Last weekend I found myself sitting in a hotel restaurant facing a surprisingly difficult decision. The breakfast buffet cost $34. Coffee was additional. Juice was extra. Alternatively, I could order à la carte. Eggs were $27. Add coffee, perhaps some fruit, maybe juice, and suddenly the economics became unclear. Was the buffet better value? How much breakfast was I actually planning to consume? At what point did the all inclusive offer become the rational choice? And perhaps most importantly, how much mental effort was I willing to spend answering these questions before I had even had my first coffee? The irony was hard to miss. In an age of AI, I was performing an optimization exercise that felt better suited to an algorithm than a human brain.

Economists have long recognized that situations like this reveal an important truth about how people make decisions. Contrary to the assumptions of classical economics, humans rarely behave as perfectly rational optimizers. Instead, we operate under constraints. One of the most influential ideas in economics comes from the Nobel Prize-winning scholar Herbert Simon, who argued that people exhibit what he called “bounded rationality.” Faced with limited time, incomplete information, and finite cognitive capacity, we do not search endlessly for the optimal solution. We settle for one that is good enough. We satisfice rather than optimize.

Standing in front of the menu, bounded rationality was not a theoretical concept. It was just breakfast. To determine the mathematically optimal choice would require estimating my appetite, comparing the marginal value of additional items, forecasting how hungry I might be later, and calculating the relative value of convenience. While possible in principle, the effort involved quickly outweighed the potential savings. At some point, the cognitive cost of solving the problem exceeded the financial benefit of getting the answer exactly right.

This is where another branch of economics enters the story. Behavioral economists have shown that decisions are heavily influenced by choice architecture, the way options are presented. The buffet was not simply a collection of food. It was a carefully designed alternative to complexity. The more confusing the individual pricing became, the more attractive the buffet appeared. Rather than calculating every possible combination, I could pay a fixed amount and stop thinking. In effect, the hotel was selling cognitive relief.

My breakfast dilemma is a trivial example of a much larger phenomenon. Insurance policies, mobile phone plans, airline tickets, mortgages, healthcare options, retirement accounts, software subscriptions, and loyalty programs all impose similar cognitive demands. Anyone who has tried to choose between dozens of health plans, compare competing mortgage offers, or maximize the value of airline reward points has encountered the same challenge. The issue is not a lack of information. It is the growing burden of deciding among too many possibilities.

Most of us cope by relying on shortcuts. We choose familiar brands. We follow recommendations. We stick with defaults. We pay extra for the problem to go away, or we postpone decisions altogether. These strategies are not signs of irrationality. They are adaptations to a world whose complexity exceeds our capacity to process every variable.

One of the most important adaptations to complexity is something known as cognitive offloading. The term describes the practice of shifting mental effort into external systems, allowing us to accomplish tasks that would otherwise exceed our individual cognitive limits. Writing allows us to store information outside our brains. Maps allow us to navigate environments we have never seen before. Calculators perform arithmetic more accurately than most people can manage unaided. Calendars remember appointments. Search engines retrieve facts we no longer need to memorize.

Civilization itself can be viewed as a series of increasingly sophisticated mechanisms for cognitive offloading. As societies became more complex, humans did not simply become smarter. They built tools, institutions, and technologies that allowed them to operate effectively within environments whose complexity far exceeded what any individual could fully comprehend.

The result was not merely convenience. Cognitive offloading expanded the frontier of what people could accomplish. A merchant could manage larger trading networks. A corporation could coordinate thousands of employees. A scientist could build on centuries of accumulated knowledge rather than starting from first principles.

AI appears to represent the next stage of this process. Previous tools helped us remember information, perform calculations, or access knowledge. AI agents may help us offload something more demanding: the continuous optimization of decisions. Much of the current discussion about AI focuses on automation. The dominant narrative is that machines will perform tasks previously carried out by humans. But an equally important shift may be the ability of AI systems to absorb the optimization burden that increasingly defines modern life.

Imagine returning to that stained, dog-eared breakfast menu, this time accompanied by a personal AI agent. Rather than calculating the economics yourself, the agent already understands your preferences, schedule, dietary goals, and spending habits. It knows you skipped dinner the night before, have a busy morning ahead, and are expected at a client lunch later that day. In seconds, it evaluates every option and recommends the one that maximizes value, not according to a generic definition of efficiency, but according to what matters most to you.

Today, individuals face hundreds of similar decisions every week. Which insurance policy offers the best coverage? Which flight provides the best balance of cost and convenience? Which healthcare provider is most appropriate? Which investment allocation best matches long-term goals? Most people simply do not have the time or expertise to optimize every choice.

As personal AI agents become more capable, these decisions may increasingly be delegated. The result is not merely greater efficiency. It potentially changes what individuals can successfully manage. As that limit expands, societies may begin to tolerate levels of complexity that would previously have been unworkable. Markets enabled societies larger than tribes. Corporations enabled enterprises larger than partnerships. Computers enabled calculations beyond human capability. AI agents may enable individuals to navigate decision environments that would otherwise overwhelm them.

What happens next is hard to predict, because adding digital labor to the mix changes the entire environment. Financial markets changed dramatically when algorithmic trading systems arrived. The algorithms did not simply make existing traders more productive. The nature of the market itself evolved because participants could process information and act at speeds beyond human capability. Trading strategies, liquidity patterns, and competitive dynamics all adapted to a world where machines had become active economic actors. Similar transformations may emerge across consumer markets, healthcare, education, and professional services as AI agents become active participants in decision-making.

This possibility raises a deeper question. Many institutions, products, and business models have been designed around the assumption that human attention is scarce and cognitive capacity is limited. Entire industries depend on the fact that consumers cannot evaluate every option, read every contract, or optimize every decision. What happens when millions of people suddenly acquire access to an always-on cognitive assistant capable of performing these tasks on their behalf?

For decades, successful products reduced complexity. They hid details, simplified choices, and abstracted away difficult tradeoffs because human attention was limited. In the future, successful products may expose more of that complexity. The winners may be those that allow AI agents to navigate a richer landscape of options, tradeoffs, and preferences than human customers could ever evaluate for themselves.

The interesting question becomes: what institutions, products, and markets were simplified primarily because humans lacked the cognitive capacity to engage with their underlying complexity? And what happens when that constraint disappears? Bounded rationality, transaction costs, and choice architecture all describe a world in which cognition is scarce. The emergence of personal AI agents suggests that we may be approaching a different reality, one in which optimization itself becomes abundant.

There is, however, an important twist. Every previous form of cognitive offloading, from writing to spreadsheets to GPS, created both gains and dependencies. We became capable of operating in more complex environments, but we also lost some of the skills that the tool replaced. Few people can navigate a city as well as they could before GPS. Few people perform arithmetic as fluently as before calculators.

The long-term question is whether AI agents become like calculators, which augmented human capability while preserving mathematical understanding, or like GPS, where many people no longer know how to navigate without assistance. The answer will determine whether abundant intelligence creates more capable humans or simply humans who are increasingly dependent on external cognition.

The challenge is not deciding what machines can think about for us. It is deciding which forms of thinking are too important to surrender.

The latest twist in the AI job replacement debate is not that machines are coming for everyone’s work. It is that, in a growing number of cases, the machines may not be cheaper. That is an awkward development for some. For the past two years, many executives have been encouraged to imagine digital labor as a form of near-frictionless substitution: fewer people, lower costs, faster output. Replace the call-center agent. Replace the analyst. Replace the junior engineer. Replace the back office. But the economics are becoming more complicated.

AI agents, coding assistants, and increasingly robotics do not simply remove labor costs. They introduce a new, highly variable, and sometimes explosive cost line: tokens, compute, orchestration, monitoring, tooling, risk controls, and infrastructure. In other words, digital labor is not free labor. In some cases, it is not even cheap labor. And that is precisely why leaders need to understand it more carefully. Their real value is not that they do the same work for less money. It is that they make entirely new kinds of work possible.

If this is news to you — and you were planning to pay for your AI strategy by firing half your workforce — you may be in for a rude shock.

Welcome to Tokenomics

The first concept leaders need to master is tokenomics. In the AI context, a token is a unit of information processed by a model. Every prompt, document, tool call, retrieved record, system instruction, chat history, intermediate reasoning step, and model response consumes tokens. In a simple chatbot interaction, that may not matter much. In an agentic workflow, where a system plans, retries, calls tools, reads files, checks work, invokes other agents, and loops through a task repeatedly, the economics change quickly.

What began as a technical metering device is now becoming a financial discipline. Tokenomics is the practice of understanding, forecasting, governing, and optimizing the cost of AI work. It asks a deceptively simple question: how many units of machine cognition are we buying, and what business value are we getting in return?

Tokenomics is likely to become a core part of corporate FinOps. The pattern is familiar. SaaS began as a way for business teams to move faster without waiting for IT. Over time, as subscription sprawl, cloud consumption, and infrastructure bills rose, finance moved in. Procurement teams scrutinized renewals. CFOs demanded usage data. FinOps emerged to create financial accountability across engineering, finance, product, and business teams. In some organizations, even technology leadership was pulled closer to finance, with CTOs and CIOs increasingly expected to justify architecture decisions not only in terms of performance and resilience, but also unit economics, utilization, and cost discipline.

AI agents will follow the same path, only faster. Token usage, model selection, inference costs, tool calls, data retrieval, latency requirements, and agentic retries will all become objects of financial control. AI spending is too technical to be left only to finance, too financially volatile to be left only to engineering, and too strategically important to be left to enthusiasm. But token governance is only the visible edge of the problem. The deeper issue is structural: companies are about to discover that digital labor is not a software feature they can simply turn on, but a new operating layer that changes infrastructure strategy, workflow design, management accountability, and the economics of work itself.

Deloitte has already warned that traditional total-cost-of-ownership models need to be refreshed for AI because tokens have become a primary unit of value and spend. Its recent work on AI token economics argues that CFOs need to connect token consumption to the P&L, model usage inflection points, and govern AI with the same rigor they apply to capital allocation. Deloitte’s survey data suggests that many companies are already generating more than 10 billion AI tokens per month, while the share expecting to exceed 100 billion tokens per month is projected to triple between 2025 and 2028. Goldman Sachs Research is even more dramatic: it expects agentic AI to drive a 24-fold increase in token consumption between 2026 and 2030, reaching 120 quadrillion tokens per month as consumer and enterprise agents scale.

When AI Works Too Well

Recent examples show how quickly AI costs can escalate. Microsoft has reportedly begun removing most internal Claude Code licenses and pushing many developers toward GitHub Copilot CLI instead. Claude Code had become popular inside Microsoft after the company opened access to thousands of employees, but the company is now winding down most usage in one major division by the end of June, with sources saying the decision is also financial.

Uber offers an even sharper warning. Uber CTO Praveen Neppalli Naga commented that the company’s surging use of Claude Code exhausted its expected 2026 AI coding budget only a few months into the year. Interestingly, Uber executives also started questioning whether higher token usage was translating into proportionally more useful consumer features, and that AI spending was creating trade-offs with hiring.

Then there is OpenClaw. Peter Steinberger, the creator of the open-source AI agent project, posted a dashboard showing more than $1.3 million in OpenAI API usage over 30 days. The bill covered 603 billion tokens across 7.6 million requests and roughly 100 coding agents. Luckily for Steinberger, OpenAI was picking up the bill. But the example still gives us a rare public glimpse into what large-scale autonomous coding can cost when budget constraints are removed.

The lesson is not that these tools do not work. Often, the problem is that they work well enough that people cannot stop using them.

That is the paradox. Cheap unit costs can still produce large bills when consumption explodes. A single token may be inexpensive. A fleet of agents running continuously across thousands of employees, tools, repositories, documents, and workflows is something else entirely.

This is why the naïve comparison between “human labor cost” and “AI labor cost” is so dangerous. It assumes that the machine is simply a cheaper substitute for the person. But digital labor behaves less like a salary and more like a utility. It scales with usage, ambition, workflow design, model choice, latency requirements, and governance discipline. A careless AI rollout can easily produce the worst of both worlds: the company still pays the salaries, adds a large new token bill, and gets only marginal productivity gains because no one redesigned the work.

From Substitution to System Design

So why would any rational executive use digital labor? Because the real selling point of a virtual workforce is not that it is cheaper. It is that it is different.

Before the rise of digital labor systems, Robotic Process Automation was typically sold as labor substitution. Take a repetitive process, map the steps, build a bot, reduce headcount or redeploy staff. The unit of value was usually cost reduction. AI agents are more interesting than that. They are not merely faster clerks. They are a different type of intelligence which is both autonmous and capable of handling ambigious situations without breaking. Importantly, their value is highest not in isolation, but when they are combined with human intelligence in novel configurations.

In healthcare, that can mean moving from episodic patient support to continuous patient engagement. A pharmaceutical company introducing a new therapy can use AI-enabled systems to check in with patients daily, reinforce adherence, surface confusion, and flag side effects or anxiety before they become reasons for discontinuation.

In insurance, digital labor can absorb the administrative burden that overwhelms organizations during moments of crisis. Allianz’s Project Nemo, launched in Australia in 2025, uses seven specialized AI agents to manage simple claims and has reported an 80% reduction in claim processing and settlement time, while still keeping humans involved in final payout decisions. Neptune Flood offers a similar lesson from Hurricane Ian, when 30% to 35% of claims were submitted through its bot, allowing customers to act at any time of day while human teams focused on higher-complexity support.

In sales, the same logic shows up as a daily intelligence layer for frontline teams. AI-driven “next best action” systems can help thousands of representatives prioritize which customers to meet, why now, what to discuss, which objections to anticipate, and what content to share. Verizon’s experience points to the power of this augmentation model: an AI assistant used by 28,000 customer service representatives helped reduce call times and contributed to a nearly 40% increase in sales through that team.

These are not one-for-one replacement stories. They are examples of system redesign. Humans remain central, but their work shifts toward judgment, empathy, persuasion, exception handling, and relationship-building. Machines take on the work that is too repetitive, too granular, too constant, too data-heavy, or too uneconomic for people to perform at scale.

The Return of Rent Versus Own

At Dell Technologies World in Las Vegas, I interviewed John Roese, Dell’s CTO and chief AI officer. He made a point that stayed with me. There is a whole class of work that has historically been uneconomic for people to do — not because it lacks value, but because the value is too small, too distributed, or too continuous to justify human effort.

Take data hygiene. Every executive knows that CRM data decays. Contacts change roles. Doctors switch specialties. Customers move. Account fields go stale. Most companies periodically pay people or third parties to clean this data retrospectively, which means the information is already wrong by the time it is fixed.

Roese has argued for autonomous “hygiene agents” — systems that continuously monitor, update, and clean databases in the background. In an earlier interview, he described these agents as a way to handle useful but neglected tasks, including CRM data cleaning, that humans often do not do because the return on manual effort is too poor.

The economics are fascinating. Suppose an up-to-date record is worth 50 cents, but it costs a dollar to fix through a cloud-based AI workflow. That makes no sense. But if you own a data center with GPUs sitting idle at night, the equation changes. Suddenly, a task that was uneconomic at 2 p.m. on rented infrastructure may be attractive at 2 a.m. on owned capacity.

This is why AI infrastructure strategy is becoming inseparable from AI operating strategy. The old cloud question — rent or own? — returns in a new form. For experimentation, renting is usually rational. For variable workloads, APIs are powerful. But for persistent, high-volume, strategically important digital labor, companies will increasingly ask whether they need their own AI factories, private inference capacity, or hybrid architectures.

The point is not that everyone should rush back on-prem. The point is that AI strategy now has an infrastructure P&L. A company that treats agentic AI as a software subscription may discover too late that it is actually running a variable-cost labor utility.

The Tokenmaxxing Trap

In my earlier piece on “tokenmaxxing,” I explored the strange new pressure on employees to demonstrate that they are being AI-powered at work. The phenomenon is understandable. Leaders have signed expensive enterprise AI agreements. Boards want proof of adoption. Managers want value. Employees want to demonstrate high performance.

But measuring AI transformation by token consumption is like measuring innovation by electricity usage. It is not only unwise; it is unfair. Employees should not be asked to carry the burden of corporate productivity simply because leadership signed a bulk licensing deal. In most companies, that approach will produce bad economics. Salaries remain. Token costs rise. People use the tools because they are told to use them. The visible activity goes up. The actual value may not.

The most dangerous version is when companies begin cutting people to pay for rising AI costs created by poorly governed AI adoption. That is not transformation. It is managerial arbitrage — and not a very good one. The right dashboard is not “tokens consumed.” It is “value created.” Did cycle time fall? Did error rates improve? Did conversion increase? Did customer satisfaction rise? Did engineers ship more reliable code? Did salespeople spend more time with the right customers? Did claims settle faster? Did risk decline? Did the organization create a capability that competitors cannot easily copy?

Without those answers, token usage is just a vanity metric with an invoice attached.

The People Who Should Be Worried

There are people who should be worried about their jobs. But it shouldn’t be the average knowledge worker. Despite the hype, in the near term, the people most exposed are the leaders who do not understand what it takes to run a real AI transformation.

Real transformation does not come from marginal improvements to individual productivity. It comes from redesigning systems of work: where decisions happen, what data flows where, which tasks should be automated, which should be augmented, which should remain human, and how the economics of machine intelligence are governed.

This is why the frontier is already moving beyond prompt engineering. Anthropic’s engineering team describes context engineering as the natural progression of prompt engineering. The challenge is no longer simply how to phrase an instruction, but how to curate and maintain the entire context around a model: system instructions, tools, external data, message history, and the operating environment in which an agent acts.

That is a management lesson disguised as a technical one. The future of AI advantage will not belong to the companies with the most enthusiastic users. It will belong to the companies that know how to design context, workflows, incentives, governance, infrastructure, and human-machine teams around valuable outcomes.

For decades, too many managers have been rewarded for allocating tasks, counting activity, and reducing cost. Digital labor exposes the limits of that model. If intelligence becomes abundant but expensive, the scarce skill is not prompting. It is judgment. It is architecture. It is knowing where machine work creates leverage and where it merely creates a bigger bill.

The rise of digital labor should shift the debate beyond whether AI will replace jobs, or whether it will be cheaper than people. Both questions miss the point. Digital labor is a new class of enterprise capability: part workforce, part infrastructure, part capital investment. It has to be designed, governed, and combined with human judgment. The uncomfortable truth is that digital labor will not only test the adaptability of workers. It will test the imagination, competence, and relevance of the people managing them.

What is the difference between a high-performing leader and someone who is simply good at performative leadership? That question is becoming more urgent in the age of AI. The latest game in the surreal, parallel universe of big organizations is tokenmaxxing: the attempt to appear highly AI-enabled by generating, consuming, or reporting large volumes of AI usage. This is not really a story about tokens. It is a story about incentives.

Tokens, in this context, are the units of data processed by AI models. In theory, they can be a useful indicator of whether people are experimenting with new tools and embedding them into work. In practice, they can very quickly become another proxy for productivity that people learn to game.

Let’s be honest: a lot of office work is performative. Not because people are foolish, lazy, or cynical, but because organizations are extremely effective at teaching people which behaviours matter. If leaders put enough emphasis on certain visible norms, employees will adopt them, often at the expense of the underlying outcome those norms were supposed to support. If you reward being seen at the office, people will stay late. If you reward responsiveness at all hours, people will send emails at midnight. If you reward AI usage, people will find ways to use AI, whether or not their work actually improves.

For years, one of the most powerful proxies was stamina. Being first in and last out became a shorthand for dedication, especially in industries like financial services, consulting, and law. There was a time when exhaustion itself seemed to carry moral authority. The person still at their desk at 11 p.m. was assumed to be more serious than the person who had designed their work well enough to leave at a reasonable hour.

Eventually, the costs of work stamina culture became too obvious to ignore: burnout, poor judgment, brittle teams, and a generation of younger workers increasingly unwilling to treat chronic overwork as a badge of honour. Several firms, including JPMorgan Chase and Bank of America, have implemented or considered a 80-hour weekly cap on junior banker hours, particularly when not working on active live deals. In China, the infamous “996” model — 9 a.m. to 9 p.m., six days a week — helped spur the “lie flat” movement, as exhausted young people rejected the premise that ambition required the total surrender of the self.

Sadly, the new badge of commitment seems to be how visibly, frequently, and enthusiastically someone appears to be using AI. At a high level, some of this makes sense. When Jensen Huang, CEO of NVIDIA says that he would be ‘deeply alarmed’ if his $500,000 engineer did not consume at least $250,000 of tokens - he is not encouraging his employees to burn tokens to drive demand for his chips. He is talking about leverage. An expensive software programmer who is not using AI to increase output is leaving productivity on the table.

But a recent story, reported by the Financial Times, shows how quickly that logic can get out of control. Some Amazon employees have apparently been using an internal AI tool called MeshClaw to automate unnecessary or non-essential tasks in order to increase their AI token usage. The tool allows employees to create AI agents that connect to workplace software and act on their behalf. Amazon had reportedly introduced targets for more than 80 per cent of developers to use AI each week and had begun tracking token consumption on internal leaderboards. The company has said these statistics are not used in performance evaluations. But several employees told the FT they believed managers were watching the numbers.

Amazon are not alone in their zeal for token usage. The NYT reports that managers at Meta and Shopify are also factoring workers’ token consumption into their performance reviews. Leaders at Google have informed employees in non-technical roles that they are also expected to use more AI in their workflows.

A company investing heavily in AI infrastructure naturally wants to know whether its people are actually using the tools. Adoption matters. Experimentation matters. New habits do not emerge by magic. But the moment usage becomes a scoreboard, it stops being merely diagnostic. It becomes social. It becomes political. It becomes a new way to signal that you are modern, ambitious, adaptive, and aligned with the future. In other words, it becomes leadership LARPing.

There are two ways this trend could play out. The first is that we drift into a newly rebooted Taylorism, updated for the AI age. Frederick Winslow Taylor’s system of scientific management was built on the idea that work could be broken down, measured, standardized, and optimized. In its original form, it was about time-and-motion studies, stopwatches, factory floors, and the separation of thinking from doing. Managers designed the system; workers executed it. Efficiency was pursued through observation, measurement, and control.

The AI-era version will not necessarily look like a foreman with a stopwatch. It will look like a system that measures prompt counts, agent invocations, token consumption, keystroke analysis, meeting analytics, automated performance summaries, and leaderboards for behaviours that may or may not correlate with meaningful work. It will be sold as productivity intelligence. Some of it may even be useful. But without judgment, trust, and a clear connection to outcomes, it risks turning AI from a tool of augmentation into a tool of managerial authoritarianism.

That would be a tragic waste. The point of AI should not be to make humans easier to monitor. It should be to make work more intelligent, more creative, more adaptive, and, ideally, less bloody stupid.

The second path is more interesting. It requires us to recognise that trends like tokenmaxxing are not primarily technology issues. They are organizational design problems. They happen when the incentive system rewards the appearance of progress more than progress itself. People are being asked to transform work while still being evaluated through old structures. When roles, reporting lines, budgets, and compensation models remain rigid, new tools fail to make meaningful impact.

At too many organizations, people are not rewarded for improving the system. A person may see an opportunity to automate a broken process, improve a customer experience, or build an AI-enabled workflow that benefits three other teams. But if that work falls outside their job description, their P&L, or their manager’s immediate priorities, they may be discouraged from doing it. In some cases, they may even be penalized. The organization says it wants innovation, but the incentive structure says: stay in your lane.

This is why crude AI adoption metrics are so dangerous. They measure the easiest part of the problem. They can tell you whether people are touching the tools. They cannot tell you whether the organization has been redesigned to benefit from them.

Management theorists used to talk about principal-agent problems: the misalignment that occurs when one party, the agent, is supposed to act on behalf of another, the principal, but has different incentives or better information. The result is agency cost: decisions that are rational for the individual but inefficient for the organization. Now, it is not only humans acting as agents inside organizations. It is also the AI agents those humans design, deploy, and supervise. These agents may be capable of moving across systems, summarizing information, drafting code, triaging email, monitoring deployments, or coordinating workflows. But they remain trapped inside organizational structures built for a slower, more human-scaled world.

The result is a new kind of agency problem. We may have increasingly powerful digital agents operating inside organizations that have not resolved the human incentive problems around them. The tool can move faster than the structure. The agent can generate more possibilities than the organization has permission to absorb. The technology can expose inefficiencies that no one is rewarded for fixing.

This is why one of the most important emerging roles is not only the forward-deployed engineer, but the work architect.

A work architect understands how work actually happens: where decisions are made, where information gets stuck, where accountability becomes blurry, where process has accumulated because no one has had the authority or patience to remove it. They know enough about technology to see what AI can do, but enough about organizations to understand why capability does not automatically translate into impact.

In writing my new book, Abundant Intelligence, I’ve had many interesting conversations with leaders about the role of work architects. Tracey Franklin, Chief People and Digital Technology Officerat Moderna, for example, believes that the real opportunity is not merely to add AI to existing workflows. It is to rethink the workflow itself. What is the outcome we are actually trying to achieve? Which parts of the process require human judgment? Which parts require speed, memory, coordination, or pattern recognition? Where do we need an agent? Where do we need a better interface? Where do we need clearly articulated decision rights? Where do we simply need to stop doing something that no longer serves a purpose?

That is work architecture. It is part systems thinking, part organisational design, part product management, part anthropology. It is the discipline of asking not “How do we get people to use AI more?” but “How should this work now that AI exists?”

The distinction is crucial. A smart leader obsessed with AI usage will ask how many tokens were consumed. A transformative leader obsessed with cognitive leverage will ask what changed. Did cycle time improve? Did quality improve? Did customers get a better answer? Did employees spend less time on low-value coordination? Did the team make a better decision faster? Did the organization learn something it can now repeat?

Over time, aspects of the work architect role will apply to everyone. The future will not belong only to people who can use AI tools. It will belong to people who can redesign their work around them. That requires curiosity, judgment, and a willingness to challenge inherited assumptions. It also requires leaders who do not punish people for stepping outside the boundaries of a job description when the real opportunity sits between functions.

The concept of leadership is very much in play right now. I will not insult your intelligence by offering a generic list of abstract qualities leaders supposedly need in the age of AI. We have all read those lists. They are not wrong, but they are rarely useful.

What I would be looking for instead are people who are genuinely obsessed with solving problems from a particular point of view. It might be customer obsession. It might be an engineering-led desire for operational resilience. It might be a hatred of unnecessary bureaucracy. It might be the simple, underrated impulse to get rid of annoying tasks and broken ways of working. These are the people who will find the real uses for AI, not because they are trying to look like AI-native leaders, but because they have something they are trying to make better.

Those people should be rewarded. They should be given room to experiment, permission to cross boundaries, and naturally - as many AI tokens as they need. But most of all, they should be measured by the problems they solve, not the proxies they perform.

The real problem with tokenmaxxing is not that people will waste a few extra tokens. It is that organizations will once again confuse what is visible with what is valuable. Leadership in this strange new world will not be defined by who vibe codes the most reports, looks the busiest, or invokes the most agents. It will be defined by who can redesign work so that intelligence, human and machine, compounds.

And when you find those people, the best thing you can do is often the simplest: get out of their way.

One problem with the way we talk about AI and professional work is that we focus too much on what autonomous systems can achieve, and not enough on what we are willing to trust them with. Just because an AI model can generate a plausible answer, or an agent can complete a workflow, does not mean we have resolved the harder question of responsibility. Who decides when the output is good enough? Who understands the trade-offs? Who is accountable when the system fails? Engineering brings that tension into sharp relief. AI may generate the design, but a human still has to sign.

I have recently completed a major research project for the ACEC Research Institute, as part of its Firm of the Future series, exploring how AI, talent shortages, data, and digital transformation are redefining the engineering firm itself. The report (which you can read here) was grounded in detailed interviews conducted across engineering firms, technology providers, and the broader AEC ecosystem, including leaders from Autodesk, NVIDIA, WSP, Mott MacDonald, Bentley Systems, Esri, BST Global, Moffatt & Nichol, and others. We combined those conversations with prior ACEC research into AI adoption and the future engineering workforce, as well as Tomorrow’s own proprietary research into digital labor and the changing design of firms. The result was not a technology forecast in the usual sense. It was a bottom-up narrative account of how leaders inside a conservative, safety-critical, highly traditional industry are beginning to think differently about work, value, talent, and the future operating model of the firm.

The bottom-line finding was interesting. AI is already changing engineering work, but not primarily by replacing engineers. The more immediate effect is that it changes the ambition of what can be achieved. Engineers are beginning to move from manually producing a small number of design options to orchestrating systems that can generate and test thousands. They are spending less time on repetitive documentation, drawing production, and knowledge retrieval, and more time framing problems, evaluating trade-offs, validating outputs, and advising clients. In the most sophisticated firms, AI is not just making the old work faster. It is expanding the solution space. It allows engineering teams to ask a different class of question: not merely, “Can we design this?” but “Of all the technically viable ways this could be designed, which one best balances cost, resilience, sustainability, safety, speed, and long-term performance?”

The engineering industry is already under enormous pressure. Demand for infrastructure is accelerating, driven by aging systems, climate adaptation, electrification, urbanization, and the rise of digital infrastructure. The AI economy itself depends on a massive physical buildout: data centers, power generation, transmission networks, cooling systems, water infrastructure, and the industrial supply chains needed to support them. The bottleneck to AI’s future may not only be faster chips, better networking, or more efficient models. It may be the human and institutional capacity to design, permit, build, and operate the infrastructure that makes AI possible.

That is why the tightening supply of engineering talent could not be happening at a worse moment. Prior ACEC research highlighted the imbalance clearly: in the U.S., roughly 184,000 engineers left the workforce in a single year while only around 166,000 new engineers were available to replace them, creating a net shortfall of about 18,000 professionals. Globally, some interviewees in our research pointed to the possibility of a much larger gap in the workforce required to deliver critical infrastructure by 2035. Just as AI is increasing the demand for physical infrastructure, the industry responsible for designing and delivering that infrastructure is running short of the people it needs. In that context, AI is not simply an optional productivity tool. It is becoming a response to a structural capacity problem.

Automation alone will not solve the capacity constraint, because engineering is not only about producing more work. It is also a question of accountability. Even in a world where AI agents can generate designs, run simulations, identify clashes, check compliance, draft documentation, and manage workflows, the need for a human engineer to take responsibility is unlikely to disappear. Engineering is not like generating marketing copy or summarizing a meeting. A design decision can shape public safety, asset performance, environmental outcomes, and capital investment for generations. Clients are not merely buying calculations. They are buying trust. Regulators, insurers, and courts will still want to know who reviewed the output, who understood the assumptions, who accepted the risk, and who signed the drawing. The machine may generate more possibilities than any human could produce, but a person still has to decide which possibility is appropriate.

As creation becomes more abundant, validation becomes the scarce and strategic act. In a traditional workflow, design generation was the bottleneck. Engineers could only explore as many options as time, budget, and human capacity allowed. In an AI-enabled workflow, that constraint begins to move. The problem is no longer whether the system can produce enough possibilities, but whether the firm can test, verify, and stand behind them. As Julien Moutte, CTO Bentley Systems explained to us, the faster designs can be generated, the more critical it becomes to ensure they are right. “You need to test everything,” he said. “That’s what gives engineers the confidence to sign the design.”

Greater AI autonomy does not diminish the profession. It elevates it. But it also raises a difficult question for firms: how do you train engineers to develop judgment if AI automates many of the early-career tasks through which judgment used to be built? If young engineers no longer spend years doing calculations, reviewing drawings, and working through the details of delivery, firms will need new forms of apprenticeship, simulation, mentoring, and deliberate practice to build the intuition that accountability requires.

The firms that thrive in this environment will not simply be those that buy the best AI tools. They will be those that redesign themselves around a world of abundant intelligence. NVIDIA’s perspective was especially revealing here. From its vantage point at the foundation of the AI infrastructure stack, the future engineering firm begins to look less like a conventional professional services business and more like a hybrid human-AI organization.

When we spoke to Sean Young, Director of AECO, geospatial, and AI solutions at NVIDIA, he described a world in which each engineer may be supported by dozens of specialized AI agents. One agent might generate geometry. Another might run structural analysis. Another might check code compliance. Another might estimate cost or schedule impact. The engineer becomes the supervisor of a digital workforce, not merely the user of a software application.

That idea has profound implications for the structure of the firm. A company with 10,000 employees could eventually be managing hundreds of thousands of digital agents. Today, IT departments manage devices, networks, software licenses, and cybersecurity. In the future, they may manage digital labor: agents that need to be deployed, governed, monitored, evaluated, and improved. The organizational question shifts from “Which tools do our people use?” to “How do humans and machines collaborate as a single productive system?” In that world, competitive advantage depends on proprietary data, simulation capability, AI governance, and the ability to encode institutional knowledge into systems that amplify human expertise. The firm of the future is not defined by headcount alone. It is defined by how effectively it organizes intelligence.

Autodesk’s future roadmap points to an equally important reversal in the work process itself. For decades, the AEC industry has moved through successive waves of digitization: from paper to CAD (Computer-aided design), from CAD to BIM (Building Information Modeling), and now from model-based design toward outcome-based BIM. Each transition changed the interface between professionals and their tools. Paper made design physical. CAD made it digital. BIM made it coordinated and data-rich. But AI changes the direction of the process. Traditionally, engineers and architects constructed a model, then tested it against requirements. In Autodesk’s emerging vision, the professional starts by defining the desired outcome. For example, a hospital might need a certain number of rooms, a net-zero target, a fixed budget, a construction timeline, and particular performance requirements. The system then generates and simulates a vast number of possible designs that meet those constraints.

Outcome-based BIM is a fundamentally different way of working. The professional is no longer drawing a solution one element at a time. They are shaping the space of possible solutions. Nicolas Mangon, Vice President AEC Strategy at Autodesk described this to us as a shift in which “the desired outcomes are the input to the process.” The software becomes less a drawing tool and more an exploration engine. For standardized asset classes such as warehouses, housing, and data centers, this could push the industry toward industrialized, repeatable, highly automated design and delivery. For complex, one-off projects, it will not remove human expertise, but it will concentrate it where it matters most: judgment, trade-offs, client context, and accountability.

The same shift is also pushing firms beyond the boundaries of project delivery. A digital model created during design no longer has to disappear once construction is complete. It can become an operational twin, connected to sensors, performance data, maintenance systems, and AI analytics. That changes the role of the firm. Instead of handing over an asset and walking away, engineering firms can remain connected to the infrastructure they helped create, helping clients understand how it performs over decades. In that world, the most valuable output of a project may not be the drawing or the model, but the intelligence that accumulates around the asset over time.

Ultimately, this is a discussion that is not only relevant to people who work in the architecture, engineering and construction industry. Many other professionals in law, consulting, accounting, finance, and technology all face versions of the same question. What happens when digital systems can perform more of the work, but humans remain accountable for the consequences? What happens when the old apprenticeship pathways are disrupted? What happens when value shifts from producing outputs to selecting, validating, and standing behind outcomes? And what happens when firms built around labor scarcity suddenly have access to scalable intelligence?

The deeper lesson is that AI will change where value lives. In engineering, the greatest risk is that the intelligence embedded in projects, data, workflows, client relationships, and hard-won judgment migrates into the platforms around the firm. The same challenge now confronts every expert organization. The firms that win will be those that redesign themselves around what they know, how they decide, and why clients still trust them when machines can generate credible answers.

The most important users of enterprise software are no longer human.

For decades, the design and economics of enterprise technology have been built around a simple idea: a person logs in, navigates an interface, and performs a task. Revenue scales with the number of users, the time they spend inside the system, and the workflows they complete. That model has produced some of the most valuable companies in the world, from Salesforce to Workday to ServiceNow, and it has shaped how we think about productivity itself. That assumption is now under pressure.

I’ve been working on a project for the American Council of Engineering Companies Research Institute on the future of the engineering firm, and one conversation in particular stood out. At Autodesk, a company at the center of the engineering technology stack, executives pointed to a striking shift. Among the most advanced architecture, engineering, and construction firms, the fastest-growing users of design software are not engineers, but agents. These digital counterparts interact directly with engineering tools through APIs, generating code, running simulations, and executing complex workflows without human intervention.

As one Autodesk executive described it to me, firms are effectively hiring “digital engineers” and training them to operate inside their systems. What looks like a usage anomaly is, in fact, an early indicator of a broader shift: when the fastest-growing users of a system are non-human, the underlying economics of the platform begin to invert. When agents drive activity, pricing based on human seats starts to break down. Autodesk and its customers are already exploring usage models tied to computational work and outcomes rather than access.

This pattern is spreading quickly. Across customer service, finance, HR, and operations, a growing share of work is now handled by AI agents that never log in, never appear on a dashboard, and never occupy a seat license. They operate through APIs, event streams, and permissions systems, responding to signals and acting in real time. Pricing is shifting accordingly. Salesforce is experimenting with action-based pricing, charging per autonomous action in some contexts. Microsoft meters Copilot usage through credits tied to computational work. Zendesk and Intercom price based on resolutions. HubSpot charges for qualified leads and completed conversations.

The unit of value is moving from the user to the outcome. And once that shift happens, the entire logic of SaaS begins to unravel. Seats measure access. Outcomes measure work. Those are not interchangeable.

ServiceNow’s Q1 2026 results show this transition in action. More than half of its net-new business now comes from non-seat-based pricing, including usage tokens and connectors, alongside traditional subscription contracts. Management has also elevated its AI-specific revenue commitment, with its Now Assist suite tracking toward roughly $1.5 billion in annual contract value in 2026 as customers expand beyond seat-based deployment.

This is the beginning of what can be described as headless SaaS.

Headless SaaS is software designed for execution rather than interaction. The interface remains, but it is no longer where value is created. Work happens through execution. Systems expose capabilities as actions that can be invoked programmatically by agents, triggered by events, and governed by policy. Humans remain involved, though their role shifts toward supervision, exception handling, and boundary setting. Software becomes less of a workspace and more of an active participant in getting work done.

The largest vendors are already reorganizing around this reality. Salesforce has introduced Agentforce to enable autonomous execution. Microsoft has expanded Copilot Studio and Azure AI Foundry into platforms for orchestrating agents across systems. ServiceNow has built an AI Control Tower to monitor and govern agent activity. Workday has defined an “Agent System of Record” to track digital workers alongside human employees. Oracle and SAP are embedding agent capabilities directly into their core applications.

Taken together, these moves signal a deeper architectural shift. Enterprise software is being rebuilt around an execution layer that runs continuously. Systems of record provide state. Action layers define what can be done. Orchestration engines coordinate activity. Identity and policy frameworks constrain behavior. Observability tools track performance, risk, and cost. The system no longer waits for a user. It operates by default.

In effect, the enterprise is being reassembled not as a set of workflows, but as a system of decisions that execute continuously.

This shift is economic as much as technical. Enterprise AI spending already reflects this transition. Estimates suggest roughly $37 billion in enterprise AI spend in 2025, with nearly half flowing to application-layer systems that execute work directly. AI-native companies are reaching $100 million in revenue faster than any previous generation of SaaS.

These signals point to a deeper shift in where value is created. The real question is no longer how software is sold. It is who controls the work, and more importantly, who controls the decisions that define that work.

For the past two decades, enterprise systems have excelled at two things. They store the state of the business, and they execute transactions reliably. A CRM records the customer. An ERP records the order. A service platform records the ticket. These systems answer with precision what is true and what just happened. What they have not owned is the decision about what should happen next. That gap is now becoming the most valuable layer in the enterprise.

As agents become more capable, they do more than execute predefined steps. They interpret situations. They decide whether to escalate a case, how to price a deal, which supplier to select, or how to route work across a network. To do this effectively, they require more than data. They require context.

Context is not a dashboard or a dataset. It is the operating system of decision-making. It is the assembled understanding of a situation at the moment of action, combining signals from multiple systems, historical patterns, constraints, objectives, and risk thresholds. This contextual scaffolding determines how an organization interprets reality and what actions it considers possible. Control over context translates directly into control over decisions.

Right now, no single player fully owns this layer. Each part of the stack holds a piece of the puzzle, but none owns the complete decision loop. Application vendors control structured data. Hyperscalers control compute and orchestration. Data platforms aggregate signals. But the logic that turns context into action remains fragmented.

As a result, a new layer is emerging inside organizations. Teams are stitching together systems that pull context from multiple sources, reason across them, and act back into each platform. Agents increasingly sit above applications rather than inside them, treating enterprise systems as interchangeable components in a larger decision engine. Decisions begin to move outside traditional SaaS boundaries. What appears to be automation is, in practice, the early formation of a distributed decision architecture.

This shift has immediate consequences. It accelerates transformation by allowing organizations to target high-leverage decisions rather than redesign entire workflows. A pricing decision can be optimized independently of the broader sales process. A routing decision can be improved without rewriting the entire supply chain. These interventions compound, gradually reshaping how the system behaves.

It also redistributes control. When decisions are made outside core systems, those systems become infrastructure rather than control points. They still hold data and execute transactions, but they no longer determine how the organization thinks. That logic moves to the layer where context is assembled and decisions are made.

This is why the repricing of SaaS is only the first signal. Investors are beginning to recognize a deeper risk: ownership of systems of record does not guarantee control if decision-making migrates elsewhere. The most valuable layer in the enterprise may no longer sit inside SaaS at all. The next phase of competition will center on who controls context, who defines decision rights, and who captures value at the moment of action.

Every layer of the technology stack is now moving toward that prize. Application vendors are embedding AI into their products to retain control. Hyperscalers are building agent platforms that operate across systems. Data platforms are expanding into real-time decisioning. Enterprises, often with the help of integrators, are constructing their own decision layers that bypass any single vendor.

What is emerging is a scramble to control the decision layer of the enterprise.

There are real challenges ahead. Security becomes more complex when non-human actors initiate actions across systems. Data quality becomes more critical as errors propagate at machine speed. The economics of AI introduce new cost structures that differ from traditional SaaS. Organizations must also rethink how they manage a workforce that now includes both humans and machines.

Even so, the direction is clear. Enterprise software is moving toward continuous execution. The number of users matters less than the volume and impact of work performed. The value of a platform is measured by how many decisions it influences and how effectively those decisions translate into outcomes.

The companies that win will not be those with the most features or the largest user base. They will be the ones that control how context is assembled, how decisions are made, and how actions are executed across the enterprise. Because in the end, the future of enterprise software will not be defined by who owns the workflow.

It will be defined by who shapes the decision.

For most of modern knowledge work, the defining anxiety has been simple and persistent: am I doing enough? Enough hours, enough output, enough visible effort to justify my role and my compensation. Performance was measured in activity, and productivity was largely a function of how effectively human effort could be applied to a problem. But what changes when effort is no longer the constraint? When intelligence itself becomes elastic, abundant, and on demand, the question shifts. The rise of the token economy is often treated as a technical or financial detail, but it is something more revealing. It is emerging as a new measure of productivity, not in terms of effort, but in terms of leverage.

The early signals are striking. Some of the most sophisticated practitioners now worry less about cost discipline and more about underutilization. Andrej Karpathy has described feeling “nervous” when he does not fully exhaust his AI token allocation, treating unused capacity as lost opportunity rather than efficiency. Nvidia CEO Jensen Huang is even more explicit: “If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed.” Failing to deploy AI is no longer prudence. It is underperformance. The benchmark is shifting from how much work an individual completes to how much intelligence they can bring to bear.

This shift is best understood as a change in constraints. For decades, the bottleneck in knowledge work was human effort. Organizations were built to allocate tasks, coordinate people, and extract efficiency from limited time and attention. Generative AI introduces a different dynamic. Intelligence, once scarce and tightly coupled to individuals, becomes fluid and scalable. The limiting factor moves again. It is no longer what the system can do, but how effectively humans can direct it. On the No Priors podcast, Karpathy pointed out that the primary constraint in engineering work is no longer compute capacity. “It’s not about flops… it’s about tokens. What is your token throughput and what token throughput do you command?” Performance is no longer defined by effort, but by the ability to direct large flows of machine intelligence toward meaningful outcomes. The implication is clear. If you cannot do this effectively, you become the constraint.

In practice, this is already reshaping how work gets done. Tasks that once defined expertise are increasingly delegated to AI systems, while humans focus on structuring problems, distributing work across multiple agents, and integrating results. The role begins to resemble orchestration more than execution. Instead of writing code, drafting documents, or performing analysis step by step, individuals manage flows of machine-generated output across several tools at once, intervening at key moments to guide direction and ensure coherence. Less like a worker, and more like a system designer.

This is the emergence of a new archetype of performance: the high throughput operator. This is not the person who knows the most or works the hardest, but the one who can effectively coordinate the largest volume of intelligence. Their advantage lies in how they frame problems, how they allocate tasks between human and machine, and how they maintain quality across an expanding surface area of output. They treat AI not as a tool to be occasionally consulted, but as an always on cognitive infrastructure. Their contribution is not measured in tasks completed, but in systems directed.

In this environment, expertise does not disappear, but it changes shape. Knowledge becomes a multiplier rather than a primary source of value. The critical skill is judgment, knowing how to break problems into machine executable components, how to design workflows that produce useful results, and how to evaluate those results before errors compound. This is where cognitive leverage becomes the defining concept. Cognitive leverage is the ability to generate disproportionate value from a relatively small amount of human input. It is the difference between doing more and making more happen. A highly leveraged individual can take a complex objective, distribute the work across a network of AI systems, and recombine the outputs into something coherent and valuable. Tokens enable this process, but they do not determine its effectiveness. That depends on how well the system is designed and governed.

This introduces a familiar tension. Tokens are both a cost and a capability. The instinct to minimize usage is understandable, but it risks constraining the very resource that drives productivity. History suggests that organizations that expand into new forms of abundance outperform those that optimize too early. Electrification created advantage not because power was cheap, but because it enabled entirely new ways of organizing production. Cloud computing followed the same pattern. It won not on cost efficiency alone, but on the ability to experiment and scale. The same logic now applies to intelligence. The question is not how much is consumed, but how effectively it is deployed.

At the same time, the labor market is beginning to adjust. The routine, structured tasks that once defined entry-level roles are among the first to be automated, reducing demand for junior positions while increasing the premium on those who can operate at a higher level of abstraction. This creates a subtle but important shift. The pathway to expertise, historically built on repetition and incremental skill acquisition, is narrowing just as the need for high-quality judgment expands. Without a deliberate approach to talent development, organizations may find themselves with more intelligence than they can direct, but fewer people capable of directing it.

As models improve and costs decline, the constraint will move again. Access to tokens will matter less. The scarce resource will be judgment, the ability to ask better questions, structure problems, and intervene at the right moments. In that world, performance is no longer about what you produce, but what you can direct. Leverage becomes the defining metric.

The implication is stark. When intelligence is abundant, underutilization becomes the new form of inefficiency. Not using what is available is no longer a sign of discipline, but of misalignment. The organizations that struggle will not be those that lack access to AI, but those that fail to reorganize around it. And the individuals who fall behind will not be those who lack effort, but those who fail to expand their capacity to direct it.

The prevailing story about AI and jobs is seductively simple: break work into tasks, measure how many can be automated, and once enough of them are, the job disappears. That logic works well for routine work. But in high-stakes, human-facing roles, especially those performed by agents and advisors, it rests on a fragile assumption: that jobs are just workflows, collections of discrete steps that can be taken apart without changing where value is actually created—or whether it can be created at all.

A recent report from Anthropic tries to quantify AI’s impact on the labor market through a measure of “observed exposure,” combining theoretical LLM capability with real-world usage from millions of Claude interactions and mapping both onto occupational task data from O*NET. Its logic is straightforward: the more of a job’s tasks AI can do, and is already doing, the more exposed that job is to disruption. It is a sophisticated extension of the task-based view of work, and it produces compelling signals about where AI is already active. But it still assumes that if you can decompose a job, you can understand its value. In many roles, especially those built around human judgment and coordination, that is precisely the mistake.

To see why, it is useful to borrow a concept from early twentieth-century psychology. The Gestalt theorists argued that we perceive patterns, not parts. A melody is not experienced as a sequence of notes, but as a whole. Rearrange the notes, and the melody disappears, even if every individual component is still present. As Kurt Koffka put it, the whole is not simply more than the sum of its parts; it is different in kind.

The same is true of many forms of work, especially those built around human interaction. What looks like a sequence of tasks on paper is, in practice, an evolving social process. Each interaction changes the next. Meaning is interpreted, not just transmitted. Decisions are shaped by timing, framing, and trust as much as by information. The outcome is not produced by completing steps, but by how people respond to them.

Consider a real estate transaction. It can be mapped as a series of steps: pricing, listing, marketing, negotiation, closing. But that is not how the deal actually happens. A buyer hesitates, not because of price, but because of uncertainty. A seller rejects an offer because it feels wrong, not because the numbers do not work. A shift in tone, timing, or phrasing can move a negotiation forward or cause it to collapse. The role of the agent is not to move the process along a checklist, but to manage a moving target of perception, emotion, and incentive. The outcome emerges from how those elements play out over time.

Financial advice operates in much the same way. Portfolio construction is increasingly straightforward. Models can optimize allocations, simulate risk, and rebalance continuously. Yet the defining moments in a client relationship rarely occur in these analytical phases. They occur when markets fall, uncertainty spikes, or life changes suddenly, and clients feel the urge to act against their own long-term interests. The critical work is not choosing the portfolio. It is navigating the client’s response to uncertainty. That is not a task. It is an unfolding interaction.

This is where task-based models begin to break down. They measure which parts of a job can be substituted, but miss how the whole situation actually works. You can automate analysis, generate documents, and manage communication flows, and still not control the outcome. Completing 80 or even 90 percent of the identifiable tasks in a role does not guarantee that a deal closes or that a decision holds under pressure. Those outcomes depend on moments of judgment, timing, trust, and emotional coordination that are not easily reduced to tasks in the first place.

A glimpse of this shift is already visible in the market. The recent convergence between Rocket, Redfin, and generative AI platforms is a case in point. Rocket’s acquisition of Redfin, followed by Redfin’s launch of an AI-powered home search experience inside ChatGPT, points toward a fully integrated, AI-native transaction stack. Discovery, pricing, brokerage, financing, and customer interaction are being pulled into a single conversational flow, collapsing what was once a fragmented process into a continuous digital experience. On one level, this is the logical endpoint of workflow automation: faster transactions, greater transparency, and radically lower friction.

But on another level, it exposes the limits of the workflow model itself. As the informational and transactional layers of an industry collapse into software, the system does not become simpler. It becomes faster, more dynamic, and often more volatile. More data does not eliminate uncertainty. It amplifies it. And as more of the customer journey compresses into software, the remaining human moments become more consequential. The role of the broker does not disappear. It becomes more valuable, precisely because it sits at the point where the process stops being computational and starts being emotional, interpretive, and irrevocable.

What AI changes, then, is not whether these jobs exist, but how they are structured. The lower layers of the work—analysis, preparation, routine communication—are increasingly handled by machines. The human role becomes more concentrated in moments that require interpretation, alignment, and commitment. The job compresses at the bottom and intensifies at the top. In fact, many of these roles, especially those requiring complex human coordination, may move toward the frontier of “high exposure” without triggering a white-collar collapse, because the work that remains is the work that matters most, and the people doing it will operate with far greater leverage than before.

This creates a less obvious but more troubling effect. Many of these professions have historically depended on apprenticeship. Junior roles provide exposure to real-world situations, allowing individuals to develop judgment over time. If AI removes much of this early-stage work, the training ground for these capabilities begins to disappear. We may become highly effective at automating large portions of the work, but less effective at developing the people who can do what remains.

The deeper point is that not all work can be broken down without changing what makes it valuable. When tasks are independent and repeatable, decomposition enables automation and scale. But when outcomes depend on how people interpret and respond to each other, breaking the work apart can strip out the very dynamics that drive results. Some jobs are workflows. Those will be automated with increasing precision. Others are social processes, where outcomes emerge through interaction over time. In those domains, AI does not eliminate the work. It raises the stakes of what remains human. AI will do most of the work. But someone still has to make it work together.

For leaders, this requires a shift in perspective. The critical question is no longer which roles have the most automatable tasks. It is where value depends on human judgment, trust, and coordination under uncertainty. Those are the roles that will not disappear, but be redefined. And they are the ones that will matter most.

Salespeople often assume they will be first in the firing line as AI reshapes work, a modern echo of Willy Loman watching the world move on from a model of selling that no longer works. There is some truth in that. The mechanics of the job are changing fast. But for those whose real work is helping other humans decide, commit, and act, the future will not be defined by how many tasks machines can perform. It will be defined by the value of what happens after those tasks are done. When value is created between people, not within steps, breaking the work apart risks breaking what makes it work.

With all the discussion about AI agents lately, you might be wondering: how exactly do you count them? If multiple agents collaborate to resolve a customer issue or approve a loan application, does that represent one digital worker or many? The question may sound trivial, but it will soon matter a great deal. Organizations will eventually track digital headcount the same way they track human employees today.

Workforces used to be easy to measure because they were made of people. Employees had identities, job descriptions, and clear places on an organizational chart. You could count accountants, engineers, or customer service representatives with simple headcount. AI agents break this model. They are not discrete in the way humans are. Agents can spawn sub-agents, operate for milliseconds, run invisibly inside software systems, or collaborate in networks that blur the line between tool and worker. What appears externally as a single agent may internally be an orchestration of models, prompts, memory systems, policy engines, and software tools. Technically the system is a constellation of components. Organizationally, however, it may still behave like a single role.

This is where the counting problem begins.

Different parts of the organization will see the same system very differently. To product marketing it may appear as one AI agent. To the software architecture team it may be a network of micro-agents. To cloud infrastructure it could represent hundreds of model calls. Finance, meanwhile, may see nothing more than a few cents of inference cost.

A simple rule of thumb helps cut through this complexity. What matters is not how many models are running, but how many decision-making roles exist inside the enterprise and how those roles interact. An agent is not defined by the number of tools behind it, but by the unit of responsibility it represents within the system. In practice, an agent is the smallest unit of autonomous responsibility in a digital workforce.

Many production agents are really bundles of models, prompts, memory systems, and tools working together behind a single interface. A customer support agent, for example, might include a reasoning model, a retrieval system, a policy engine, a summarizer, and an action executor. Technically that is a multi-agent pipeline. From the perspective of the enterprise, however, it functions as one digital worker with a defined role. If several internal components collaborate but consistently produce one coherent decision or action, it is best understood as a single agent with internal architecture, much like a human employee who relies on spreadsheets, software, and colleagues to do their job.

Research in technology governance highlights why this distinction matters. Sociologist Madeleine Clare Elish coined the term “moral crumple zone” to describe what happens when complex automated systems fail. Just as the crumple zone in a car absorbs the force of a collision, responsibility in automated systems often collapses onto the nearest human operator, even when the broader system design shaped the outcome. When organizations cannot clearly identify which digital systems act with autonomy or authority, accountability defaults to individuals rather than the architecture that produced the decision. Defining the boundaries of digital workers therefore becomes more than a technical exercise. It is a way of ensuring that responsibility is assigned where it actually belongs.

If agents are going to function as digital workers, leaders need a simple way to identify them. Here are some practical rules that might help:

The first is identity. If a system has a persistent identity inside the organization, it begins to behave like a digital worker. It can authenticate into systems, receive permissions, and perform actions that can be traced back to that identity. If a system cannot be independently identified and audited, it is probably just a component inside a larger architecture.

The second rule is lifecycle control. A system that can be provisioned, updated, paused, or retired independently has an operational lifecycle. That means it can be managed much like organizations manage applications or service accounts. By contrast, a micro-agent that appears only as part of an orchestrated chain of tasks is closer to a function than a worker.

The third rule is accountability for outcomes. A digital worker should own a measurable task or result. An IT support agent might be responsible for responding to service tickets within a defined service level. If a system contributes only a hidden sub-step within a larger workflow, it likely belongs to the system architecture rather than the workforce.

Together these rules create a surprisingly clear boundary. If a system has a distinct identity, an independent lifecycle, and responsibility for a defined outcome, it begins to resemble a digital employee. If not, it is better understood as infrastructure.

But what happens when components begin to behave like independent actors? If systems have distinct roles or objectives, if they operate asynchronously, coordinate decisions with one another, or expose separate identities and interfaces to the organization, then you are no longer looking at one agent. You are looking at a team of agents. At that point the system begins to resemble a small digital organization rather than a single worker augmented by technology.

Consider an aviation analogy. A modern aircraft cockpit contains autopilot systems, navigation computers, sensor networks, and sophisticated flight software performing thousands of calculations every second. Internally it is an extraordinarily complex digital environment. Yet operationally we still treat autopilot as part of a single role: the aircraft’s flight control system assisting the pilot.

Air traffic control, by contrast, is a distributed coordination system. Radar networks, aircraft, scheduling systems, and human controllers interact across towers and control centers. Each participant has its own responsibilities, authority, and identity within the system. What emerges is not one augmented operator but a network of interacting roles. The difference is not the number of machines involved. It is whether the system supports one role or many.

This shift from architecture to accountability is already appearing in governance frameworks. The U.S. National Institute of Standards and Technology has begun exploring how agent systems should be identified, authenticated, and authorized as they interact with digital infrastructure. The emphasis on identity and authorization reveals an important assumption: if agents are going to act autonomously inside enterprise systems, they must be treated as identifiable entities whose actions can be traced and governed.

International governance frameworks are moving in a similar direction. Emerging ISO standards like ISO/IEC 42001:2023 for AI management systems require organizations to define the scope of their AI deployments, manage them across their lifecycle, and assign accountability for their behavior. These frameworks do not attempt to catalog every model or algorithm inside a system. Instead, they focus on identifying which systems operate as actors inside organizational processes and ensuring those actors can be governed responsibly. Implicitly, they adopt the same principle: what matters is not the internal architecture of AI systems, but the role they play inside the enterprise.

For most executives, the debate about counting digital workers will likely surface first on the org chart. Should AI agents appear alongside human employees? In 2024, the HR software company Lattice briefly experimented with allowing companies to list AI employees in its platform, only to reverse course after a public backlash. At the time the idea seemed provocative, even absurd. In retrospect it may prove inevitable. If digital workers have identities, permissions, responsibilities, and measurable outcomes, they begin to resemble organizational actors rather than tools. The more interesting question may not be whether agents appear on org charts, but how their presence reshapes them. As digital workers take on operational decisions once handled by layers of management, hierarchies built around supervising people may give way to flatter structures designed to coordinate human and machine decision-making.

Yet even this debate about org charts may be missing the deeper shift underway. Org charts, after all, are still a way of counting people and managing layers of control. Agentic systems are beginning to change the underlying economics of work itself. The real transformation in organizations is not simply the number of digital workers they deploy, but the amount of decision-making capacity embedded in their operations.

Historically, firms measured productive capacity through simple metrics such as headcount or labor hours. Those measures made sense in an industrial economy where human attention was the primary constraint. Agentic systems change that equation. As organizations embed intelligence into everyday processes, the relevant question shifts from how many workers exist inside a workflow to how much cognition the system can execute. The more meaningful metrics may become things like decision throughput or the number of tasks completed autonomously. Instead of asking how many workers are involved in a process, leaders may soon ask how many decisions that process can execute per second.

So, with all that in mind, how many AI agents does it take to change a lightbulb? Arguably, none. A sensor detects the outage. A diagnostic model determines the cause. A procurement system orders a replacement. A scheduling agent allocates a technician. A workflow system verifies that the job is complete. But unless the organization has a team of highly sophisticated humanoid robots, it still takes a human to take the bulb out of the box and screw it in.

Depending on how you feel about the future of work, that may be good news for now.

A dangerous new market narrative is spreading through boardrooms and earnings calls: artificial intelligence has made companies so productive that they can slash their workforce and barely notice the difference. Analysts applaud, the stock jumps, and executives describe a future where digital labor replaces the old human-heavy operating model. Unfortunately, the economy is rarely that tidy.

Over the past year a growing list of companies has announced layoffs framed around AI-driven efficiency. Logistics software firm WiseTech Global said AI-assisted development tools were collapsing project timelines from months to days as it eliminated roughly 2,000 roles. Chemical giant Dow announced thousands of job cuts as part of an automation and AI overhaul, though weakening industrial demand clearly played a role. Autodesk and Pinterest both reduced headcount while promising to redirect resources toward AI initiatives. Insurance group Allianz has suggested that advances in AI-powered customer service and claims processing could eventually displace thousands of call-center jobs.

On the surface, AI-powered labor substitution looks like the beginning of a productivity revolution. In reality, the story is more complicated. Some companies over-hired during the post-pandemic boom and are now shrinking under the convenient cover of AI disruption. Others are desperate to signal relevance in a market obsessed with artificial intelligence. A small handful of firms are seeing real gains from digital labor. But even there executives may be drawing the wrong conclusions about what actually drives scale.

Consider the restructuring at Block. The firm announced plans to cut more than 4,000 employees, roughly half its workforce. At first glance the move resembles the early days of Twitter after Elon Musk arrived with a chainsaw and a strong view that most Silicon Valley companies were bloated.

But the Block restructuring is more deliberate than it appears. The company has spent the past several years embedding AI into its internal workflows, particularly in software development. Engineers use AI tools to generate code, test features, and accelerate product cycles that previously required large teams and layers of coordination. Leadership says the result is a dramatic jump in productivity and a surge in gross profit per employee. In other words, the cuts are not simply about removing people. They reflect a bet that software agents can remove friction from the company’s internal machinery.

Klarna tells a very different story. The Swedish payments company aggressively promoted its AI transformation, claiming that generative AI assistants were performing the work of hundreds of customer service agents. Hiring slowed, headcount fell, and executives highlighted rising revenue per employee as proof that the model was working.

Then reality intervened. Customer support interactions turned out to be more complex than a chatbot script. Financial disputes require empathy, judgment, and trust. Klarna eventually reintroduced more human service capacity and shifted toward a hybrid model where AI handles routine inquiries while people manage difficult situations.

Comparing the AI transformations of Block and Klarna reveals an important principle that many companies miss. The best target for AI restructuring is workflow friction, not headcount. Klarna initially pitched AI as a labor substitute. Block frames it as a force multiplier for smaller teams. The second framing is far more robust. When AI removes operational obstacles around skilled workers, organizations unlock real leverage. When AI tries to erase the human layer entirely, the system often deteriorates in ways that financial metrics fail to capture.

Another key difference between the two companies is where AI is deployed. Back-office augmentation is far easier than customer-facing replacement. Internal engineering, model building, summarization, quality assurance, and repetitive analysis are forgiving environments for AI agents. Mistakes can be corrected before they reach customers. Customer service is different. It involves emotion, nuance, and exceptions. Automation failures there damage trust quickly. Block’s investments sit largely in the first category. Klarna pushed too aggressively into the second.

The metrics used to justify these restructurings also deserve closer scrutiny. Revenue per employee has become the poster statistic of the AI productivity story. Klarna’s executives highlighted it repeatedly. Block has emphasized gross profit per employee. Investors love these ratios because they appear to compress efficiency into a single number.

But the math is misleading. Cut the workforce in half while revenue stays flat and the metric doubles overnight. The statistic improves even if the organization itself becomes weaker. Revenue per employee tells us what happened after the layoffs. It does not prove that the company became more scalable.

Klarna illustrates the danger perfectly. The revenue-per-employee story looked brilliant until the company realized that removing too many humans from the system degraded the customer experience and forced it to rebuild parts of the workforce. The ratio improved before the operating model was proven.

The real test of AI-driven productivity is not whether a company can survive with fewer employees. It is whether the organization can reduce the marginal cost of coordination without eroding trust. True scale in the AI era comes from redesigning how intelligence is configured throughout the firm. That means shorter decision cycles, better exception handling, lower cost to serve, stronger decision quality, and preserved customer relationships.

When you look closely at companies where AI is genuinely improving productivity, three structural shifts appear:

1. Coordination compression. Artificial intelligence reduces the friction between analysis, decision making, and execution. Code generation, automated testing, rapid experimentation, and internal agents executing workflows shrink the distance between an idea and its implementation.

2. Decision leverage. Humans move up the stack. Instead of performing every task themselves, they supervise systems that generate and evaluate options at scale.

3. Cost-to-serve decoupling. AI systems handle routine work so efficiently that the marginal cost of serving another customer or processing another transaction begins to fall.

That is the real signal of scale. Not fewer employees but lower coordination cost per decision. From this perspective, the market’s fascination with AI layoffs misses the bigger story. Artificial intelligence is not simply a tool for replacing workers. It is a technology for redesigning the architecture of work. Companies that treat AI primarily as a headcount reduction strategy may discover that they have optimized a ratio while weakening the system that created the value.

The winners in the AI era will not be the companies that eliminate the most employees. They will be the ones that redesign work so that every human decision is amplified by machines. Headcount may fall, but that will be a consequence of scale, not its cause.

There is, perhaps, a small silver lining in the current wave of AI anxiety. Not long ago, the dominant fears revolved around killer robots, runaway superintelligence, and apocalyptic scenarios that ended with data centers being nuked from space. Today the panic is more grounded, and in many ways more sophisticated. We are no longer imagining machines conquering humanity; we are worrying about white-collar unemployment ticking above 10%, mortgage books wobbling in San Francisco, and private credit portfolios unraveling because software agents can write code faster than junior analysts. The monsters have moved from science fiction to the balance sheet.

The 2028 Global Intelligence Crisis from captures this shift perfectly. Subtitled “The Consequences of Abundant Intelligence,” it presents a fictional macro memo from the near future in which cheaper, more capable AI triggers a white-collar job apocalypse, hollows out discretionary spending, and destabilizes housing and credit markets. It is cleverly constructed and economically literate, and its viral spread reflects a genuine unease among investors and executives. Yet for all its sophistication, the argument ultimately rests on a mispricing of what abundant intelligence actually means.

What the authors frame as an Intelligence Collapse Scenario is more accurately understood as an Intelligence Reconfiguration Scenario. The difference is not semantic. It is structural. The real question, one I have been exploring extensively in my own work on abundant intelligence, is not whether digital labor transforms the economy, but how that transformation is architected: who retains authority, who captures the surplus, and how judgment is redistributed when execution becomes abundant.

Written as a retrospective memo from June 2028, the essay sketches a world in which “abundant intelligence” delivers surging productivity alongside double-digit unemployment, as white-collar professionals, once the engine of discretionary consumption, are structurally displaced. The authors’ central mechanism is what they call “ghost GDP”: output and corporate profits rise on paper, but income no longer circulates through households because machines do not earn salaries or spend money. As wages contract and consumption weakens, asset prices and credit structures built on stable high-income employment begin to crack. Each firm’s rational decision to substitute software for labor aggregates into a systemic feedback loop, where declining demand justifies further automation, reinforcing what they portray as an intelligence displacement spiral with no obvious stabilizer.

It is a compelling story. But it rests on a critical modeling assumption that deserves scrutiny: that machine intelligence primarily substitutes for human work, and that wages are the only meaningful transmission mechanism of economic value. The memo treats intelligence as if it were a fixed pool of salaried labor. When machines perform that labor, value supposedly disappears from the system unless it flows through paychecks. That is a 20th-century production model applied to a 21st-century technology.

The deeper question is not whether machines can perform more tasks. It is how organizations reallocate judgment, authority, and ownership when intelligence becomes abundant. Modern enterprises are not simply collections of jobs. They are architectures of decision rights. Someone allocates capital. Someone signs off on compliance. Someone bears legal liability. Someone determines acceptable risk. AI systems can draft, optimize, simulate, and execute. They cannot absorb responsibility in the way institutions require.

When intelligence becomes abundant, value does not evaporate. It migrates. The constraint shifts from execution to orchestration. As digital labor absorbs routine analysis, drafting, coding, optimization, and coordination, the residual human contribution does not simply shrink in importance. In many cases, it becomes more leveraged. When a single executive, engineer, or strategist can direct systems that generate ten times the output of a traditional team, the marginal impact of their judgment increases, not decreases. The value of being correct when machines execute at scale rises sharply.

Consider how capital markets reward decision-making authority today. Portfolio managers do not earn fees because they personally process every data point. They earn fees because their judgment governs large pools of capital. The more leverage embedded in the system, the more valuable the individual exercising oversight becomes. Digital labor functions in a similar way. When output scales non-linearly but decision rights remain concentrated, the marginal productivity of judgment rises. Digital labor does not erase authority. It amplifies the consequences of those who hold it.

The crisis scenario assumes a simple substitution dynamic: one AI agent replaces one $180,000 employee. Multiply that across the economy and aggregate demand collapses. Yet real organizations rarely operate through one-to-one replacement. They operate through reconfiguration. Some roles disappear. Others expand. A smaller number of individuals may control far more productive systems. Income distribution may widen. But that is not the same as permanent economic contraction. If AI substitutes 50% of white-collar labor and multiplies the productivity of the remaining 50% by five, the income dynamics look radically different from pure elimination.

The memo also models only one economic effect of cheaper intelligence: substitution. It largely ignores two others that accompany every dramatic fall in input cost: scale expansion and new use cases. When a core production factor becomes cheaper, usage tends to explode. Lower-cost intelligence reduces the price of experimentation. It lowers barriers to entry. It enables new products and services that were previously uneconomic. Legal advice, design support, financial modeling, research assistance, and personalized education have historically been constrained by scarce human hours. As digital labor lowers those constraints, the total addressable market for intelligence-intensive work expands.

Abundant intelligence increases the number of problems worth solving. When launching a company, prototyping a product, or analyzing a market requires fewer human hours and less capital, more individuals can participate. Each new venture generates demand for coordination, oversight, trust-building, governance, and strategic direction. In that sense, digital labor expands the surface area of the economy itself. Execution becomes cheaper, but the need for judgment does not contract. It often intensifies.

This does not imply a frictionless transition. Routine cognitive labor will be commoditized. Middle layers may compress. Inequality may widen before it stabilizes. But the equilibrium outcome is unlikely to be mass professional obsolescence. It is more plausibly a bifurcation: execution becomes abundant, while high-leverage judgment, accountability, and system design become more valuable.

Many AI doomer scenarios share a hidden assumption: that artificial intelligence evolves rapidly while humans, organizations, and markets remain fixed in place. Capabilities improve. Tasks disappear. Wages fall. Systems fracture. Yet history suggests the opposite dynamic. Every general-purpose technology has triggered dislocation followed by reinvention, with new skills repriced, institutions redesigned, and entirely new industries emerging around the technology itself. The industrial revolution reorganized labor and capital. Electrification reshaped production. The internet created markets that were previously unimaginable. Betting that AI will transform cognition while leaving human adaptability unchanged is to ignore the most consistent pattern in economic history.

Abundant intelligence will commoditize certain forms of work. It will also elevate what remains scarce: judgment under uncertainty, ethical accountability, cross-domain synthesis, and the willingness to assume responsibility when automated systems fail. The real risk is not that machines change everything. It is that we misinterpret what is changing. Intelligence is becoming abundant. Judgment is not.

The future will belong not to those who resist digital labor, nor to those who deploy it blindly, but to those who understand how to redesign authority, ownership, and value creation around it.

The AI crisis arrived without fanfare. There were no alarms, no cascading red dashboards, no breathless messages from the security operations center. At 3:17 a.m., somewhere between Singapore and Rotterdam, an AI routing agent inside a global logistics company glitched slightly. Just milliseconds. But that agent sat at the center of thousands of shipments, negotiating contracts, rerouting containers, balancing fuel costs and port congestion in real time.

That week, the company cloud provider had quietly shifted workloads to a different region after an energy price spike. A frontier model vendor had rolled out an update that subtly changed how the system reasoned. An external technology partner, granted limited access months earlier, had folded usage patterns into broader product improvements now available to competitors. Nothing was hacked. Nothing was stolen. Yet by quarter’s end, delivery times slipped, margins thinned, and the firm’s once sharp operational instincts felt strangely generic.

This was not a cyberattack. It was a sovereignty failure. The company did not control the intelligence it depended on. Small external shifts in infrastructure, models, and learning loops compounded into strategic drift, and the firm had no easy way to recalibrate. In a world where AI systems mediate core decisions, enterprise sovereignty is not about keeping intruders out. It is about ensuring that when the intelligence layer beneath your business moves, you are the one steering it.

Over the past year, sovereignty has been framed largely as a geographic issue. Should data sit on premise or in the cloud? Should models be hosted domestically or offshore? These are not irrelevant considerations, but they increasingly resemble economic trade offs rather than existential strategic choices. Compute can be shifted. Data centers can be mirrored. Regulatory constraints can often be engineered around. The deeper question is more uncomfortable: how much of your end to end supply chain of intelligence do you actually control, and how much of it rests on layers you neither see nor govern?

At Davos, NVIDIA CEO Jensen Huang described AI not as a monolithic breakthrough but as a five layer cake. At its base sits energy. Above that, chips and computing infrastructure. Then cloud data centers. Then AI models. And finally, the application layer where intelligence expresses itself in products, services, and workflows. Each layer must be financed, constructed, and operated. Each embeds its own capital intensity, geopolitical exposure, and technical constraints. Huang’s point was that this platform shift will generate economic activity across sectors, from power generation and advanced manufacturing to cloud operations and software development. Yet implicit in his metaphor is something else: each layer is also a potential point of sovereign vulnerability.

The volatility of AI economics rarely surfaces in the user interface. It is buried in the architecture. A token is not simply a unit of text—it is a compressed signal of infrastructure. Each one carries the fingerprint of a GPU generation, the power draw of its rack, the bandwidth of its interconnects, the latency across regions, and the complexity of the model architecture behind it. When electricity prices spike, inference costs don’t just rise—they ripple across the stack. When a new chip improves performance per watt, cost curves bend. When storage or network throughput lags, user experience suffers. Tokeneconomics is infrastructure economics rendered in milliseconds. And the companies that ignore this hidden volatility risk finding that their margins are tethered to physical and geopolitical forces they neither see nor control.

Many executives assume that if their AI applications are functioning smoothly today, their strategy is secure. But surface stability can mask structural fragility. A change in model licensing terms can flow upward into customer facing experiences. A regulatory restriction on cross border data flows can constrain training pipelines. A reliance on a single orchestration framework can make it prohibitively expensive to migrate to an alternative provider. In this context, sovereignty is not about physical location. It is about strategic leverage and the capacity to reconfigure your intelligence stack when conditions change.

Microsoft CEO Satya Nadella made a similar argument in Davos when he suggested that the physical location of a data center is “the least important thing” for AI sovereignty. What matters, he argued, is whether a firm can embed its tacit knowledge into model weights that it controls. If you cannot distill your proprietary customer data, operational history, and institutional expertise into models under your governance, then you are effectively leaking enterprise value into external systems. Nadella predicted that corporate sovereignty in the AI era would become one of the most discussed topics in boardrooms this year. His insight reframes sovereignty away from geography and toward cognition.

Nadella is correct that weights matter, but wrong to imply that they are sufficient. Fine tuned models represent compressed organizational memory. They encode patterns from years of transactions, customer interactions, supply chain disruptions, and strategic decisions. In that sense, they resemble a vault, a dense numerical artifact containing the essence of how a company operates. But focusing exclusively on weights risks missing a more profound shift that is now underway. We have moved from retrieval computing, where competitive advantage stemmed from accessing information efficiently, to generative computing, where advantage emerged from synthesizing novel outputs from large scale learned patterns. We are now entering the era of agentic computing, in which systems do not merely answer or generate but plan, coordinate, execute, and adapt across complex workflows.

In an agentic world, sovereignty extends beyond a single model. It resides in how intelligence is orchestrated. It lives in the design of workflows that determine which tasks are automated and which require human judgment. It is expressed in the guardrails that constrain autonomous action, the verification loops that ensure reliability, and the feedback mechanisms that continuously refine performance. Two companies may license the same foundation models, run on the same cloud infrastructure, and even possess similar volumes of data. Yet their outcomes can diverge dramatically. The difference lies not only in what they know, but in how they configure what they know.

Intelligence configuration is the emerging frontier of competitive advantage. How is work decomposed between humans and machines? Which decisions are delegated to agents and which are escalated to managers? How are agents granted access to internal tools and external APIs? How are exceptions surfaced and resolved? How is institutional knowledge encoded in prompts, policies, reinforcement learning loops, and monitoring dashboards? These design choices shape how value is created and captured. They determine whether intelligence accumulates within the enterprise boundary or dissipates into shared platforms.

Enterprise sovereignty, then, is less about isolation and more about optionality. It is the ability to switch model providers without dismantling your workflow architecture. It is the capacity to retrain systems on proprietary data without renegotiating fundamental platform dependencies. It is the discipline of mapping your exposure across energy, chips, infrastructure, models, and applications, and understanding where concentration risk resides. As intelligence becomes the essential ingredient in every transaction and interaction, the boundaries of the firm become cognitive as much as physical.

There is an enduring story about senior Coca Cola executives who know the secret formula and are not permitted to fly on the same plane. Whether apocryphal or not, the symbolism captures a core truth about value creation. Certain assets are so central to a company’s future that their concentration represents a strategic risk. In the AI era, your secret formula may not be a chemical recipe locked in a vault. It may be a constellation of fine tuned weights, proprietary reinforcement loops, curated data pipelines, and uniquely configured networks of agents working in concert with human teams.

Defending enterprise sovereignty is ultimately about defending that constellation. It requires recognizing that the real attack surface is not only cybersecurity but dependency. It demands that boards and executives look beneath the interface layer to the stack, and beneath the stack to the configuration of intelligence itself.

The next disruption may not arrive as a breach notification. It may appear as a subtle shift in energy pricing, a model update that alters performance characteristics, or a vendor policy change that constrains how your data can be used. Enterprise sovereignty is your capacity to absorb intelligence shocks, reconfigure your architecture, and ensure that the secret formula of your organization remains firmly within your control.

In the mid nineties, designers did not know what online shopping was supposed to look like. So they borrowed from the physical world. Early retail websites featured isometric shopping carts gliding down digital aisles. Shelves were rendered in crude 3D. You clicked arrows to “walk” through a store. After a while, search bars replaced aisles, and recommendation engines became the new merchandising layer. Eventually, mobile screens collapsed the store into a feed. We are now at another such inflection point. Retailers are redesigning the storefront again. But this time, the shopper may not even be human.

Over the past eighteen months, the largest U.S. retailers have begun quantifying the impact of AI-powered shopping assistants. Walmart’s Sparky, embedded directly into its mobile app, is one of the clearest early case studies. On its most recent earnings call, Walmart disclosed that customers who engage with Sparky generate average order values approximately 35 percent higher than non-users, and that roughly half of U.S. app users have tried the assistant.

Lowe’s has reported similar traction. Its Mylow assistant answers nearly one million customer questions per month, and the company has stated that customer engagement with Mylow more than doubles conversion rates. In-store, the associate-facing Mylow Companion tool has been linked to a 200 basis point increase in customer satisfaction scores. The lesson is straightforward. When AI is embedded directly in high-intent surfaces and connected to fulfillment and inventory systems, it can drive measurable commercial outcomes.

Amazon, meanwhile, has positioned its AI assistant Rufus as a generative shopping guide capable of answering product questions, comparing items, and supporting research. But the more interesting move may be Amazon’s “Buy for Me” feature, which allows customers to purchase select items from third-party brand sites without leaving the Amazon app. That blurs the line between retailer and intermediary. Amazon becomes not just a marketplace, but a purchasing agent. According to Amazon CEO, Andy Jassy, Customers who used Rufus were about 60% more likely to complete their purchase.

The stakes are high. Retail media, the practice of selling sponsored placements within retailer ecosystems, has grown into a global market projected to approach $200 billion by 2026–2027. For companies such as Amazon, advertising has become one of the highest-margin segments of the business. If human eyeballs are replaced by machine queries, sponsored placements stop influencing people and must start influencing algorithms. Retail media becomes less about persuasion and more about protocol.

At the same time, AI is reshaping retail far beyond the front end. Amazon has reported that AI-driven forecasting has delivered a 10 percent improvement in long-term national forecasts for deal events and a 20 percent improvement in regional forecasts for millions of popular items. Walgreens has disclosed that its micro-fulfillment centers now fill approximately 16 million prescriptions per month, with shipped volumes up 24 percent year over year and roughly 40 percent of total prescription volume at serviced stores handled through these facilities. Best Buy has reported that AI-driven call summarization has reduced average engagement time in customer service interactions by nearly 5 percent.

These are early but tangible indicators of what might be called digital labor: AI systems that do not merely answer questions but execute tasks, compress cycle times, and reallocate human effort. The front-end assistant and the back-end automation are converging. Yet the most destabilizing force may not be retailer-owned assistants at all. It may be the rise of personal autonomous agents and AI-native browsers.

In 2025, OpenAI introduced Operator, a browser-based agent designed to handle repetitive tasks such as filling out forms or ordering groceries by interacting directly with web interfaces. Perplexity launched Comet, an AI-native browser explicitly marketed around delegating tasks such as shopping and applying promo codes. OpenAI later introduced Instant Checkout and an Agentic Commerce Protocol, allowing merchants to integrate directly so that users can complete purchases within ChatGPT with tokenized payments and explicit confirmation flows.

OpenAI has stated that ChatGPT now serves more than 800 million weekly users. Even if only a fraction of those users begin experimenting with agentic shopping, the distribution leverage is extraordinary.

Unlike Sparky or Rufus, these agents do not belong to a retailer. They operate across the open web. They can log into accounts, compare products across sites, and execute multi-step workflows. Some rely on formal protocols and APIs. Others use UI automation to mimic human browsing behavior. That distinction is not merely technical. It is strategic. Retailer agents optimize inside a walled garden. Personal agents optimize across the entire digital landscape.

The tension has already surfaced publicly. In late 2025, Amazon threatened legal action against Perplexity over its agentic shopping tool, alleging covert access to customer accounts and disguised automation. This is the first visible skirmish in what may become a broader contest over who controls the decision interface in commerce.

From the consumer’s perspective, the promise is appealing. Instead of browsing, filtering, and comparing manually, the user delegates intent: restock the pantry under a certain budget, plan a themed event, optimize purchases for sustainability. The agent executes. From the retailer’s perspective, however, the shift is existential. If the customer’s AI shops on their behalf, the traditional surfaces for merchandising, branding, and advertising are reduced or eliminated.

Adoption signals suggest consumers are becoming more comfortable acting on AI guidance, even if full delegation remains nascent. Adobe’s holiday retail reporting indicates that nearly half of consumers expressed trust in AI-driven shopping experiences in 2025, and that a majority of users who encounter AI-generated links click through on them. Traffic from AI chatbots to e-commerce sites has surged year over year, albeit from a relatively small base. The shift from answering to doing is underway.

Still, the security and governance challenges are nontrivial. There is growing research to demonstrate that large language model–integrated systems are vulnerable to indirect prompt injection, in which malicious instructions embedded in web content are treated as commands by an agent. In early 2026, a prompt injection vulnerability in an open-source agent toolchain was exploited to distribute OpenClaw, highlighting how quickly such systems can become vectors for abuse. When agents can execute transactions, prompt injection is no longer a hallucination problem. It is a financial risk.

These dynamics point to three plausible five-year scenarios.

In the first, retailer-centric assistants dominate. Sparky, Rufus, Mylow, and similar tools remain embedded in retailer-owned apps and sites, driving higher basket sizes and deeper loyalty. Retail media adapts to conversational interfaces but remains within the retailer’s walled garden. The retailer controls the intelligence layer and the economic capture.

In the second, platform-centric assistants become the primary gateway. Consumers initiate shopping journeys within ChatGPT, Gemini, or AI-native browsers. Retailers integrate via standardized commerce protocols, supplying product data and fulfillment capacity while ceding control of the initial interaction. Retail media migrates into new forms of sponsored recommendations within AI environments. The intelligence layer shifts upward.

In the third, personal agents gain traction. Consumers rely on persistent AI systems that maintain context across retailers and sessions. The website becomes less a destination and more an endpoint, a structured data feed and fulfillment engine optimized for machine legibility. Retailers compete on API quality, delivery speed, transparent pricing, and reliability rather than on visual merchandising.

Physical infrastructure remains decisive across all three scenarios. AI can mediate choice, but it cannot yet deliver a package. Forecasting improvements, warehouse automation investments and the fine-tuning of proprietary retail AI models are likely to remain the backbone of any agentic promise. If agents make purchasing instantaneous, fulfillment performance becomes even more visible.

The deeper shift is not from websites to chat interfaces. It is from human-driven browsing to delegated decision-making. The shopping journey is collapsing. What once required browsing, comparison, and deliberation is being compressed into a single delegated instruction. As AI assistants evolve from answering to acting, the central strategic question for retailers is no longer how to build a better interface. It is who sits between the customer and the transaction.

The next storefront may not be a store at all. It may be a negotiation between your AI and someone else’s.

By 10:47 a.m. on Wednesday morning, billions of dollars had evaporated from wealth management stocks. There had been no earnings miss. No regulatory shock. No fraud. Just a press release from a startup announcing an AI-powered financial planning tool that could analyze tax returns, generate scenarios, and personalize investment strategies in minutes. Within hours, asset managers in London were sliding in sympathy. Brokerage firms in New York were down sharply. Days earlier, legal publishers and data providers had suffered similar fates after the launch of new AI research tools. A wealth manager and a legal publisher have very little in common. Yet, as reported by the FT, their stocks fell for the same reason in the same week.

Investors were no longer analyzing industries. They were scanning for automation exposure.

This is the early shape of a new valuation regime. Markets are beginning to price artificial intelligence risk everywhere, but they are doing so bluntly. Entire segments are being discounted not because revenues have collapsed, but because someone somewhere might automate part of what they do. Software companies are punished because AI agents could reduce seat licenses. Insurance brokers are sold because an app can compare policies. Professional services firms are repriced because generative models can draft, summarize, and create advice. In each case, the reaction is category-level. The assumption is that if AI can do a task, then firms built around that task must be structurally impaired.

In the short to medium term, this pattern is understandable. Equity valuation depends on assumptions about margins, growth rates, and the durability of competitive advantage. AI destabilizes all three at once. It threatens fee structures by lowering the cost of delivering knowledge work. It compresses barriers to entry by making sophisticated capabilities widely accessible. It makes long-term forecasts harder because productivity gains are nonlinear and unevenly distributed. Faced with this uncertainty, analysts default to caution. They lower multiples, raise discount rates, and trim guidance. When in doubt, they sell first and revisit later.

But this transitional phase of broad devaluation should not be confused with long-term structural decline. We are witnessing the first-order reaction to a general-purpose technology. History suggests that when a foundational technology emerges, markets initially punish exposure to perceived risk before they learn to differentiate between those who will be disrupted and those who will be transformed. The current wave of repricing reflects anxiety about automation, not yet insight into configuration.

This distinction matters. The dominant valuation gap of the last two decades separated technology companies from traditional firms. Software commanded premium multiples because it was asset-light, scalable, and defensible. Industrial, financial, and service firms were valued more conservatively because they were labor-intensive and capital-bound. AI is beginning to dissolve that divide.

As I argue with my co-author, Nitin Mittal in our new book, Abundant Intelligence: How Digital Labor Will Rewrite the Rules of Business - the next valuation frontier will not be tech versus non-tech. It will be cognitively leveraged versus cognitively constrained. Cognitive leverage is the ratio of useful intelligence applied to a problem relative to its cost. Digital labor, in the form of AI agents, adaptive robots, and machine reasoning systems, dramatically reduces the unit cost of cognition and getting things done.

When intelligence becomes abundant, the economic question shifts from access to configuration. The firms that will command premium valuations are not simply those that deploy AI tools, but those that redesign their operating models around scalable intelligence. They will reallocate work between humans and machines deliberately, increase the speed of decision cycles, and expand the scope of what each employee can accomplish.

In the interim, however, we should expect volatility and value destruction. As AI tools improve, business models built on selling standardized cognitive outputs will come under pressure. Subscription software priced per user may face headwinds if autonomous agents can perform tasks across platforms.

Professional services firms that bill by the hour may struggle if clients expect AI-augmented productivity gains to translate into lower fees. Education platforms that monetize answers will compete against free generative tutors. These shifts can compress margins and reduce growth rates before companies adapt.

Yet the longer-term effect is more nuanced. Digital labor does not simply eliminate work; it redistributes and reconfigures it. When routine analysis, drafting, or coordination becomes machine-augmented, human effort can be redirected toward higher-order judgment, creative synthesis, and system design.

Organizations that treat AI as a bolt-on efficiency tool may realize incremental savings. Those that redesign workflows, governance structures, and incentive systems around blended human-machine intelligence can unlock operating leverage that traditional metrics struggle to capture.

This is where equity markets will eventually refine their lens. Rather than discounting entire sectors based on automation exposure, investors will begin to assess how firms configure intelligence. Do they own proprietary data that improves their models? Have they redesigned processes to eliminate redundant human friction? Are they able to scale output without proportional headcount growth? Can they increase revenue per unit of cognition, not just revenue per employee? These questions cut across industry boundaries.

Consider two wealth management firms. Both face AI-driven planning tools. One treats them as back-office assistants to reduce paperwork. The other integrates digital agents into client onboarding, portfolio construction, risk modeling, and continuous engagement, enabling each advisor to serve twice as many clients with higher personalization. The first experiences margin compression. The second expands capacity and improves outcomes. From a sector perspective, both are “wealth managers.” From a cognitive architecture perspective, they are fundamentally different enterprises.

The same divergence will emerge in insurance, law, engineering, healthcare, and manufacturing. Some firms will cling to labor-based models and see multiples compress. Others will build operating systems around digital labor, improving speed, scale, and scope while reducing error and latency. The equity market will eventually recognize that intelligence configuration, not industry label, determines sustainable advantage.

In the near term, markets are pricing fear. The volatility in software, financial services, and professional content reflects a rational awareness that the old economics of knowledge work are unstable. But indiscriminate selloffs obscure a more strategic reality. AI is not simply an automation wave; it is a redefinition of how organizations create and capture value. The abundance of intelligence shifts scarcity toward design, orchestration, and governance.

For CEOs and boards, this is not merely a technology strategy question. It is a capital markets question. If valuation increasingly reflects cognitive leverage, then leadership must measure and manage it explicitly. They must understand where human judgment is essential, where machine autonomy adds speed and precision, and how the two interact. They must move beyond pilot projects toward systemic redesign. Because the market will not wait for perfect clarity. It will continue to scan for vulnerability and reward adaptability.

The recent selloffs across unrelated sectors are an early signal of this shift. Investors are searching for a new organizing principle in an AI-shaped economy. They have not yet found the right metric. When they do, the valuation gap that defined the digital era will be replaced by a new one. It will not separate technology firms from traditional industries. It will separate those who scale intelligence from those who merely consume it.

For the last year, the debate around AI at work has split into two unhelpful extremes. On one side, we have breathless talk of “AI coworkers,” complete with onboarding rituals, performance reviews, and soft-focus imagery of humans and machines collaborating happily at their desks. On the other, we have an anxious counter-reaction that insists this language is dangerous, misleading, and fundamentally wrong, because AI systems are not people and should never be spoken of as if they were. Both camps miss the point. The real question is not whether machines deserve human metaphors, but whether leaders understand what kind of economic force they are unleashing, and what kind of organization that force demands.

Calling something “labor” has never meant it is human. It means it performs work, incurs cost, produces output, and sits inside a system of incentives, controls, and trade-offs. We already accept this logic in countless places without controversy. We speak of mechanical labor, industrial labor, and even “work” performed by capital assets, logistics networks, or energy infrastructure. The word labor is not a sentimental label. It is an accounting term, a political term, and a strategic one. It tells us where effort is applied, how value is created, and who captures the surplus.

The discomfort with the phrase “digital labor” often comes from confusing metaphor with mechanism. The fear is that if we talk about AI as labor, executives will start treating software like employees, importing human management practices into systems that do not need motivation, morale, or meaning. That fear is not unfounded. We have already seen organizations fall into the trap of grafting new technology onto old organizational charts, preserving familiar roles and routines while claiming transformation. But that failure is not caused by the term. It is caused by shallow thinking. Bad metaphors do not invalidate good economics.

What digital labor actually names is a shift in how work gets done, measured, and priced. AI systems do not simply assist humans. They execute tasks end to end, at variable cost, with increasing autonomy, and with performance characteristics that are fundamentally different from human workers. They scale instantly, improve unevenly, fail in strange ways, and demand oversight that looks nothing like traditional management. Pretending this is just “capability” without acknowledging its labor-like effects does not make organizations wiser. It makes them blind.

This blindness shows up most clearly in how firms talk about productivity. When AI is framed purely as a tool, leaders focus on local efficiency gains. Faster reports. Cheaper analysis. Fewer errors in routine tasks. These improvements matter, but they are not the transformation. The transformation comes when the cost of performing cognitive work collapses and organizations are forced to rethink which activities are scarce and which are abundant.

At first, this shows up as redistribution. Tasks move. Responsibilities shift. What once required teams now requires supervision. What once consumed days collapses into minutes. Work does not disappear so much as it migrates, flowing toward the edges where judgment, context, and accountability still matter. But redistribution is only the visible surface of change. If leaders stop there, they mistake motion for progress.

The deeper shift occurs when organizations recognize that collapsing cognitive costs undermine the logic of existing processes. When work becomes cheap and fast, many structures no longer make sense. Approval layers exist because information was scarce. Handoffs exist because humans were slow. Entire organizational designs evolved to manage limitation, not to maximize value creation. Digital labor exposes this reality relentlessly, forcing a question most firms avoid: if this process were designed today, knowing what machines can now do, would it exist at all?

This is why digital labor cannot be reduced to a workforce debate. Labor is not just something you manage. It is something you allocate. It competes with capital. It reshapes bargaining power. It determines how value flows through the firm. When AI performs meaningful portions of knowledge work, the organization is not merely adopting a technology. It is redefining its production function. Ignoring this reality because the word “labor” feels anthropomorphic does not make organizations more precise. It makes them strategically incoherent.

The opposite reaction, fear of mass job displacement, suffers from a similar lack of depth. It assumes a zero-sum replacement model, where machines simply take human jobs and the story ends. History suggests something more complex. Technological shifts rarely eliminate work in aggregate. They reprice it. They change where value is created, which skills command a premium, and which roles lose their economic justification. The political consequences are real, but they are not caused by machines acting independently. They are shaped by decisions leaders make about deployment, governance, and distribution.

Digital labor does not automatically destroy jobs. It destroys certain task bundles. It exposes inefficiencies that were previously hidden inside roles. It forces organizations to confront how much of their structure exists to coordinate human limitation rather than to create value. In doing so, it often increases demand for judgment, system design, oversight, and creative problem solving, even as it reduces demand for routine execution. The danger is not that machines work. The danger is that institutions fail to adapt.

One reason this debate remains stuck is that we lack a language for hybrid systems. Tools are subordinate. Workers are autonomous. AI agents are neither. They act independently within boundaries, they learn from feedback, and they require governance rather than supervision. Calling them tools understates their agency. Calling them coworkers overstates their humanity. Digital labor sits in between. It is productive capacity that must be designed into workflows, not bolted on.

The deeper challenge is not reallocating tasks, but reimagining the system itself. When work can be executed at near-zero marginal cognitive cost, many processes cease to make sense in their current form. Entire functions exist today to compensate for latency, error, and coordination overhead that no longer need to exist. Digital labor does not simply improve the organization. It questions whether the organization, as currently designed, is still fit for purpose.

This is where transformation either accelerates or stalls. Redistribution optimizes the existing machine. Reinvention questions whether the machine should exist at all. Firms that stop at redistribution preserve their hierarchies and routines while quietly automating away the substance inside them. Firms that pursue reinvention redesign decision rights, collapse layers, and rebuild workflows around intelligence rather than headcount.

Some argue these systems should be treated purely as capital investments, governed through portfolio logic rather than workforce thinking. There is truth here, but it is incomplete. Unlike traditional capital assets, digital labor is not static. Its performance can drift. Its costs can spike. Its failures can be rare but catastrophic. It requires continuous monitoring, retraining, and oversight. These are operating realities, not one-time investments. Treating AI purely as capital underestimates the work required to keep it reliable and aligned.

More importantly, capital language alone cannot capture the social and political dimensions of this shift. Labor has always been about power as much as productivity. Who controls work. Who benefits from efficiency gains. Who absorbs risk when systems fail. When AI performs work at scale, these questions do not disappear. They intensify. Avoiding the language of labor may feel safer, but it often obscures the very consequences leaders must confront.

The real issue, then, is not whether digital labor is the right phrase, but whether leaders are willing to engage with its implications. Digital labor does not mean machines are people. It means work has become programmable. It means cognition has a marginal cost. It means organizations must design systems in which humans and machines jointly create value, each doing what they do best, without pretending they are interchangeable.

This reframing also explains why superficial adoption fails. Simply inserting AI into existing roles preserves the old logic of the firm. It optimizes locally while leaving global structure untouched. True transformation requires rethinking workflows from first principles, asking which decisions should be automated, which should remain human, and which should be shared. It demands new metrics, new governance models, and new leadership capabilities.

Language matters because it directs attention. When we talk about digital labor, we force a confrontation with cost, substitution, and value capture. We ask how cheap cognition reshapes strategy. We ask who benefits when intelligence becomes abundant. Avoiding the term does not eliminate these questions. It merely delays them.

In the end, the future of work will not be decided by metaphors, but by design. Organizations that succeed will be those that treat AI neither as an employee nor as a gadget, but as a new productive force that reshapes everything it touches. They will measure it rigorously, govern it deliberately, and integrate it thoughtfully. Digital labor is not about making machines more human. It is about making organizations more intelligent.

At first glance, Disney’s recent moves look contradictory. The company announces a sweeping partnership with OpenAI that allows its characters to appear inside generative tools like Sora, while almost simultaneously firing off cease-and-desist letters to Google and pressing forward with aggressive litigation against Midjourney. To some observers, this looks like confusion. To anyone who has watched Disney for long enough, it looks like something else entirely. This is a company that has spent a century mastering the art of adapting control to new forms of participation.

The OpenAI agreement is not a casual collaboration or a vague pilot. Disney is putting real money behind it, reportedly a $1 billion equity investment in OpenAI as part of a three-year partnership that brings more than 200 characters into Sora and ChatGPT Images, including animated, masked, and creature characters across Disney, Pixar, Marvel, and Star Wars. The deal reportedly includes warrants that give Disney the right to acquire additional equity, and it is structured with careful exclusions, such as no talent likenesses or voices. In other words, Disney is not handing over the keys to its entire creative universe. It is licensing a very specific kind of use, under a very specific set of constraints, inside a distribution channel that it can influence

To understand what is really happening, it helps to revisit an older idea. In my first book, Futuretainment (2009), I argued that the defining shift in media was not digital distribution but participatory consumption. The audience was no longer content to sit back and watch. They wanted to interact, remix, customize, and inhabit the worlds they loved. This was not a fringe behavior. It was a structural change driven by new tools and social norms. From fan fiction to mashups, from mods to machinima, people were already treating entertainment as something to play with rather than something to receive.

At the time, many media companies framed this behavior as piracy or infringement. But my argument was that remix culture was not a rejection of creativity or authorship. It was an expression of deeper engagement. Fans were not stealing stories because they did not value them. They were reworking them precisely because they cared. Participation was becoming the new signal of loyalty. The mistake was assuming that control meant suppressing these behaviors rather than shaping them.

Generative AI takes that argument and detonates it at scale. What once required niche skills now takes a prompt. Anyone can visualize a character, extend a narrative, or invent an alternative version of a familiar world in seconds. This is not a new desire. It is a new interface. And it forces a hard choice on rights holders. You can either fight participation everywhere, or you can decide where and how it happens.

Disney has been through this before. Mickey Mouse is not just a character. He is a legal and cultural artifact that has repeatedly sat at the fault line between new technology and shifting consumer behavior. When Mickey debuted in 1928, synchronized sound was the disruptive force, and Disney quickly realized that technology could be an amplifier rather than a threat. As new reproduction technologies emerged, from television to home video to digital distribution, Mickey’s image became both ubiquitous and fiercely protected. Each extension of copyright around Mickey was not simply about money. It was about maintaining authorship and brand coherence in the face of wider access.

The famous copyright extensions that critics dubbed the “Mickey Mouse Protection Act” mirrored moments when copying became easier and distribution more diffuse. Xerox machines, VHS tapes, the internet. Every time the tools changed, the legal perimeter shifted. Disney’s genius was never in freezing culture in place. It was in ensuring that participation happened on terms it could manage.

Seen through that lens, the OpenAI deal makes sense. Disney is not endorsing a free-for-all remix culture. It is licensing participation into a controlled environment. OpenAI becomes a sanctioned playground where Disney characters can be used, explored, and recombined within guardrails that preserve identity and value. The company is not abandoning copyright. It is operationalizing it for a world where imagination is interactive by default.

The contrast with Midjourney and Google is telling. The lawsuits and cease-and-desist letters are not about fans experimenting with characters for fun. They are about systems that can produce near-identical replicas, that blur the line between inspiration and duplication, and that operate outside any negotiated framework. From Disney’s perspective, this is not participatory culture. It is unlicensed industrialization of its intellectual property.

What is really being negotiated here is the future of consumption itself. Entertainment is shifting from products to platforms, from finished artifacts to living systems. Characters become interfaces. Worlds become sandboxes. The value no longer sits only in the story that is told, but in the space that is created for others to tell stories of their own.

This is exactly the trajectory I described in Futuretainment. The future belongs to companies that can design for agency without surrendering authorship. The winners will not be those who lock everything down, nor those who let everything go, but those who curate participation in ways that feel empowering rather than restrictive.

Disney’s move with OpenAI is a bet that the next generation of fans will expect to play with stories, not just watch them. The company is choosing to license the remix rather than endlessly litigate it. At the same time, it is drawing a bright line around who gets to build the tools that make that remix possible.

This is not hypocrisy. It is strategy. Mickey Mouse has survived radio, television, home video, cable, the internet, and streaming by adapting the boundaries of control to each new medium. Generative AI is simply the latest chapter in that long history. The lesson for the rest of the media industry is clear. Participation is no longer optional. The only real question is whether you design for it, or spend the next decade trying to sue it out of existence.

For years, Waymo’s autonomous vehicles were known as the politest drivers in San Francisco. They came to full stops, waited patiently at four-way intersections, yielded generously, and behaved with a kind of algorithmic courtesy that seemed almost naive. Then, as the Wall Street Journal recently reported, they began to behave very differently: darting around double-parked trucks, merging aggressively, accelerating the instant a light turned green, even performing the occasional illegal U-turn. Basically, like a NYC taxi driver.

But here’s the plot twist. This wasn’t a malfunction. It was an upgrade. Waymo’s leaders explained that the cars needed to become “confidently assertive” because extreme politeness was creating safety issues. A vehicle that followed the rules perfectly was behaving in ways that other drivers did not expect. And in traffic, unpredictability is a form of risk.

This is a revealing lesson for the broader challenge of AI safety. We tend to assume that the safest systems are the ones with the most rigid guardrails. But in many real-world settings, risk arises not from breaking rules, but from breaking expectations.

I learned this years ago when I lived in Istanbul. I love driving - especially loud, classic, American muscle cars. But that aside, frankly - I’m also a cautious, rule-following driver. I quickly realized I could not safely drive in Turkey. Istanbul’s traffic isn’t chaotic — it’s coordinated on a different logic. Drivers move in fast, assertive rhythms, signaling intention less through indicators than through momentum and micro-negotiations. My careful Western driving style was out of sync with the system. I was the anomaly, and therefore the most dangerous person on the road.

This is exactly the dynamic that game theory helps illuminate. Human environments operate less like rulebooks and more like coordination games, where stability comes from shared expectations about how others will behave. These equilibria vary by culture, city, even neighborhood. They evolve through practice, not regulation. In such systems, a player who obeys the written rules but violates the unwritten equilibrium becomes unpredictable — and unpredictability destabilizes the game for everyone else.

Early autonomous vehicles made this mistake. They entered the traffic system playing a different game. Their rule-bound behavior created what game theorists call strategy incoherence. Humans behaved according to local norms; the AI behaved according to formal laws. Safety problems emerged not because the AI was reckless, but because it was misaligned with the equilibrium.

This brings us to an idea I explored in a Harvard Business Review article a little while back. When determining how much autonomy to grant AI agents, leaders usually focus on how big a potential risk is. But a more useful question is what kind of risk it is. Some problems are complicated — governed by fixed rules and stable relationships. Others are ambiguous — shaped by context, norms, and feedback loops. And some are uncertain, where neither rules nor data can reliably guide decisions.

Driving, despite its legal structure, is not complicated. Traffic laws are complicated. Driving is ambiguous.

Ambiguous problems are ones where AI systems become safer by engaging more deeply with the real world. More context improves their judgments. More interaction refines their models. As I wrote, “AI agents won’t always get it right, but they learn fast.” That is why Waymo doesn’t rely on teleoperation when a car is confused. Human operators offer guidance, but the AI must continue making decisions. The system strengthens not through constraint, but through exposure. Seen through this lens, Waymo’s shift from hyper-cautious to assertive makes sense.

The safest behavior is not always the most conservative. It is the behavior that best matches the expectations of the people around the system. A decisive merge, a quick acceleration, or a tactical lane change may appear aggressive, but if it aligns with local driving norms, it is actually the more predictable and therefore safer choice.

Economists have a parallel concept. Milton Friedman argued that inflation is driven not just by prices but by expectations. Once people expect inflation, they act in ways that make it real. Human systems are governed by beliefs about how others will behave. Traffic is no different. An autonomous vehicle that violates these beliefs, even while obeying the law, injects uncertainty into a coordination game that depends on shared assumptions.

This insight applies far beyond robotaxis. As AI agents proliferate inside organizations, they will increasingly operate in domains filled with tacit norms: customer service, workflow orchestration, decision support, compliance, prioritization. These environments resemble ambiguous games, not rigid rule systems. The most dangerous agent will not be the bold one, but the one that acts out of sync with human expectations.

Current AI guardrail strategies don’t fully account for this. They treat AI systems as if they operate in complicated domains where the primary goal is to restrict behavior. But ambiguity requires a different approach. The goal is not to eliminate variation, but to shape behavior so that systems remain consistent with the equilibrium of the environment they inhabit.

This is the real challenge of autonomous systems. We are not simply programming machines to follow rules. We are teaching them how to participate in human coordination. The next frontier of AI safety is not technical constraint, but behavioral coherence — designing agents that understand the social, cultural, and contextual signals that make their actions legible and predictable to the humans around them.

Waymo’s assertive driving is an early glimpse of this future. The first generation of AI systems obeyed the rules. The next generation must understand the game.

Tokyo moves with the cool precision of an algorithmic machine. Step out of Shinjuku Station at dusk and the city unfurls around you in layers of neon haze—LED kanji flickering like loose packets on a network, billboards pulsing with the heartbeat of a vast digital organism. Crowds slide past in silent, perfect non-collision, as if everyone is running the same invisible protocol. In a metropolis of 37 million, you brace for chaos. Instead you get choreography—an improbable calm humming beneath the circuitry of the streets.

Greater Tokyo is the largest metropolitan region in the world. By any rational measure, this density should produce disorder: competing demands, clashing priorities, frayed nerves, and amplified frictions. Yet the opposite happens. Trains arrive to the second. Streets stay clean without armies of inspectors. Public spaces remain safe without overt policing. People follow norms no one states but everyone understands. The overall effect is so natural you barely notice it—until you step outside the city and realize how rare it is.

After spending the last week immersed in this city, I’ve come to believe that its stability is not the product of engineering alone, though its infrastructure is superb. The deeper mechanism is a cultural operating system: a shared layer of simple behavioral norms. Respect for shared space. Consideration for strangers. The quiet choreography of queues, thresholds, and ritualized subway behavior. Tokyo’s order is not enforced; it emerges. It is the aggregated result of millions of small, internalized decisions that compound into large-scale stability.

This stands in stark contrast to other global cities—places I won’t name, because I visit them too often—where disorder persists despite a thicket of regulations, penalties, and punitive enforcement. More rules do not necessarily produce more order. In fact, they often signal its absence. Tokyo proves that in complex, densely interactive environments, norms beat rules every time.

And this, surprisingly, may be the most important lesson for the future of AI governance.

As organizations shift toward architectures built on swarms of autonomous agents, we are entering a world that will behave far more like a dense, dynamic city than a traditional computing system. Agentic technologies are inherently nondeterministic; they don’t simply execute instructions, they interpret, infer, and negotiate context. They learn. They misread. They interact in ways no designer can fully anticipate. And as these agents begin to operate at massive scale—potentially millions or billions coordinating workflows, optimizing supply chains, or engaging consumers—the system will not obey top-down control. It will exhibit emergence.

Today, most attempts to govern AI rely on constraints: filters, rules, compliance layers, and external guardrails. It’s an understandable instinct—if the system is unpredictable, tighten control. But this mirrors the cities that depend on punishment because shared norms are weak. It produces brittle governance: systems that behave well in the lab but fracture in the wild, the moment agents encounter novel situations or interact in unexpected combinations.

Tokyo offers a different blueprint. In complex adaptive systems, order doesn’t scale through restriction; it scales through coherence. Systems researchers describe coherence as the stable patterns that arise when independent components align their behavior without central direction. In complex environments, coherence—not control—is what prevents chaos.

Tokyo works because its behavioral substrate is aligned, not because its laws are draconian.

Instead of trying to script every action, the city embeds simple, universal behaviors at the lowest layer—norms that guide and constrain without prohibiting, shaping tendencies rather than dictating outcomes. These norms don’t eliminate uncertainty; they channel it. They create a predictable distribution of behavior even when individuals vary widely. They generate order by default, not by decree.

If we want safe, stable multi-agent AI ecosystems, we need to take a similar approach: embed normative priors into the foundation of our systems. Normative priors are behavioral defaults—embedded assumptions about how an agent should act, resolve uncertainty, coordinate with others, and interpret human intent. They bias agents toward pro-social behavior before they ever encounter real-world data. They’re not hard rules but foundational dispositions, guiding the agent toward predictable, human-aligned behavior as it learns, adapts, and interacts at scale.

Instead of specifying every rule, we define the behaviors we want to emerge:

• Respect human intent and oversight.

• Be transparent about actions, goals, and reasoning.

• Minimize unintended impact and avoid unnecessary escalation.

• Default toward cooperation when interacting with other agents and humans.

• Defer to humans and seek clarification when uncertain.

• Communicate uncertainty and limitations explicitly.

These are not constraints. They are behavioral tendencies—simple norms that, when shared across vast numbers of agents, create emergent governance: stability arising from alignment rather than force. And just as in Tokyo, perfect compliance isn’t required; predictable tendencies are enough to produce large-scale order.

This matters because the next generation of organizations will not resemble pyramids of reporting lines. They will be polycentric networks of humans and machines making real-time decisions. Trying to centrally police every action in that environment would be as futile as trying to direct traffic at Shibuya Crossing with a whistle. The only scalable strategy is to get the substrate right—to design agents that behave predictably even when acting autonomously.

Tokyo demonstrates that the simplest norms—embedded deeply and enacted consistently—can produce astonishing forms of order in environments that should, by all conventional logic, be chaotic. The city’s quiet choreography is not the result of constant oversight. It is what happens when a system’s dynamics are shaped by deep behavioral protocols—predictable patterns emerging without central control.

If we want our future societies of human and machine intelligence to function with similar coherence, we shouldn’t begin with constraints. We should begin with norms. The lesson Tokyo offers is deceptively simple: in complex systems, stability is not imposed. It is cultivated. And the most powerful form of governance is not enforcement, but alignment at the foundation.

Elon Musk recently predicted the end of the smartphone. “In five or six years,” he said, “we won’t have phones in the traditional sense. What we call a phone will really be an AI edge node — no apps, no OS, just AI.” It’s easy to dismiss such statements as provocation, but he may be right for reasons that have nothing to do with hardware or his views on the mass adoption of AI-generated content.

The smartphone model has simply become too slow for what comes next. The traditional loop of unlocking, tapping, and waiting for an app to respond belongs to an era when people tolerated friction. In a world of predictive, context-aware AI, that’s already starting to feel clumsy. The real replacement for the smartphone isn’t a new device, it’s a new rhythm — where intelligence anticipates rather than waits, and latency, not interface design, determines how fast we get things done. Latency — the tiny delay between intention and action — will be what separates systems that feel instant and real from those that seem clumsy or obsolete.

The British company Nothing has already begun to edge toward Musk’s vision. Its new AI platform, Essential, lets users build mini-apps through simple natural-language prompts. There’s no coding, no app store, no interface to navigate — just a conversation. You describe what you want, and the phone generates it on demand. It’s a small glimpse of a world where computation is ambient and anticipatory, not something we command but something that happens around us.

In that world, latency becomes experience. When a system hesitates — when a chatbot pauses mid-sentence, a car’s sensor reacts a moment too late, or a warehouse robot stutters before adjusting course — the illusion of intelligence collapses. The difference between seamless and frustrating, safe and catastrophic, often comes down to the time it takes for data to travel and a model to respond.

And that’s why latency is fast becoming one of the biggest strategic challenges for the AI industry. Real-time applications like autonomous vehicles, live fraud detection, and industrial robotics all depend on split-second inference. In conversational systems and virtual assistants, every extra second of delay erodes trust. In large-scale AI training, “tail latency” — the drag caused by the slowest servers or packets — can extend job completion by hours and waste millions of dollars in idle GPUs.

The parallels with high-frequency trading are instructive. A decade ago, hedge funds spent fortunes co-locating their servers beside exchanges to shave microseconds off their trades. Today, the same logic applies to cognition itself. The firms building the fastest loops between users, data, and models will deliver experiences that feel almost precognitive — systems that answer before you’ve finished asking.

That insight is reshaping the entire technology stack. Verizon and Amazon Web Services have announced AI Connect, a new long-haul fiber network designed specifically for generative workloads, where every millisecond of delay compounds across billions of inferences. Cisco’s Unified Edge initiative takes the opposite approach, moving computation closer to where people and machines actually work — retail stores, factory floors, clinics — so that decisions can happen locally rather than waiting for a distant cloud.

But no company has grasped the implications of latency more clearly than NVIDIA. In partnership with Nokia, it’s building what it calls AI-native radio networks, embedding GPU compute directly into the next generation of 6G towers. The network itself will run inference, reducing the time between sensing and decision to almost nothing.

In Germany, NVIDIA and Deutsche Telekom have committed over a billion euros to build the country’s first Industrial AI Cloud, powered by 10,000 of NVIDIA’s new Blackwell GPUs. The goal isn’t just AI sovereignty — it’s cognitive proximity. By turning telecom geography into AI factories, they’re collapsing the physical distance between data and intelligence.

Each of these moves reflects the same realization: that the future of AI won’t be decided by the size of your models but by how close they are to the moment of action. Intelligence that’s distant is expensive, slow, and brittle. Intelligence that’s near — that operates at the edge, embedded in the environment — feels instant, intuitive, and alive.

This shift is already visible at the hardware level. Edge processors now capable of running 13-billion-parameter models on-device are cutting inference delays by up to 70 percent. Tasks like translation, image recognition, and predictive text can happen locally, while heavier reasoning is handled in the cloud. It’s a new kind of cognitive choreography — the mind distributed between body and brain. The closer intelligence sits to reality, the faster it learns from it.

Low latency doesn’t just make systems faster; it makes them possible. A self-driving car navigating traffic, a surgeon consulting an AI model in real time, or a security network detecting threats before they unfold — all depend on microsecond responsiveness. In these environments, latency is no longer a nuisance; it’s a liability.

Every technological revolution has its limiting factor. For the industrial era, it was energy; for the digital era, it was bandwidth. In the era of artificial intelligence, it may be latency. The companies and countries that master it will own the rhythm of the future — setting not just the pace of communication, but the tempo of thought itself. For businesses, this means competitive advantage will hinge not just on algorithms or data, but on architecture. The winners of the next decade will be those who can build the fastest feedback loops — compressing the distance between sensing and understanding, decision and action.

Latency is no longer just a concern for network engineers. It’s now a measure of how fast a business can think and respond. In an AI-driven world, a few milliseconds can decide whether a service feels intuitive or frustrating, whether a company anticipates its customers or lags behind them. The same logic that’s may ultimately render the smartphone model obsolete applies to business itself — those still waiting for input will be overtaken by those that predict and act. Reducing that gap isn’t just technical work — it’s strategy.

Every technology revolution offers a brief window of unreasonable advantage—an era when the bold can seize opportunities that later become ordinary. This is that moment for AI. The cost of cognition is collapsing. The rules haven’t caught up. The field is open to anyone ambitious enough to rewire their work, their organization, or their industry around machine intelligence. But none of these conditions will hold for long.

Like the early internet before paywalls and advertising, or the dawn of ride-sharing before prices rose to match reality, we are living through a period of unsustainable abundance. And history is clear: free rides don’t last.

This is the best of all possible times to embrace AI, and the worst time to hesitate.

1. AI Is Still an Unfair Advantage

AI has already entered the mainstream, but mastery remains uneven. Most people are still using it for trivial tasks—writing professional sounding emails, summarizing meeting notes, or using their chatbot as a digital therapist.

Those who know how to truly leverage AI are operating in another universe entirely. When you train ChatGPT Pulse on your own writing, reports, and meeting transcripts, it’s like waking up each morning to a team of researchers who have curated your private edition of the Harvard Business Review.

Across industries, the same pattern is emerging. Marketing teams are deploying AI agents that can generate, test, and localize thousands of campaign variations overnight—then feed the winning ideas directly into production systems. Supply-chain leaders are using autonomous forecasters that continuously adjust logistics based on weather, pricing, and geopolitical data, eliminating weeks of human analysis. Product teams are training domain-specific copilots that surface customer insights, simulate competitive dynamics, and recommend strategic trade-offs before a meeting even begins.

What once required committees, consultants, and coordination now happens as a real-time conversation between leaders and their digital counterparts—an entire cognitive layer of the enterprise that never sleeps, learns continuously, and compounds advantage for those who know how to direct it.

Right now, there’s still enormous alpha to be captured simply by being better at using AI than your peers. A mid-level consultant, analyst, or creative can plug into these tools and achieve leverage that once required an entire team. They can produce better insights, faster outputs, and more refined products—and sell them at a premium—because the knowledge and implementation gap remains wide.

But that window will close. As AI literacy spreads and models become embedded directly into platforms and workflows, the arbitrage disappears. To add value, you’ll need either deep engineering skill or very specific domain expertise. Everyone else will simply go straight to their own AI agents to get the job done.

2. Someone Else Is Paying for Your Intelligence

AI is absurdly, unsustainably cheap.

For $20 a month—or even for free—you can access sophisticated systems capable of reasoning, summarizing, coding, and designing at levels that would have required teams of specialists and millions in infrastructure only a few years ago.

But remember Uber’s early days, when rides were subsidized to be cheaper than public transport? Or Spotify, when streaming felt like a miracle before licensing costs caught up? Eventually, reality asserted itself. The same will happen here.

AI’s current economics are a gift. OpenAI, Anthropic, and others are still losing money on every query, propped up by investor subsidies and growth ambitions. Someone is paying the bill - just not you.

The energy, compute, and data costs of cognition will ultimately find equilibrium. Prices will rise, or access will be constrained. We are in the VC-subsidized golden age of cheap intelligence—a temporary arbitrage between technological possibility and economic gravity.

If you are a leader, this is your chance to build capability at minimal cost. Soon, you’ll be paying full price for the same cognitive horsepower.

3. Rulemakers Haven’t Caught Up

Regulators are still asleep at the wheel. AI currently sits in a grey zone: too new to be tightly controlled, too powerful to remain that way for long. The early web followed the same pattern. Before copyright enforcement, ad-tracking, and compliance bureaucracy took hold, there was a brief, chaotic explosion of creativity. That was when Google, Amazon, and PayPal were born.

Today’s AI landscape feels similar. The EU has passed its AI Act, and the U.S. is drafting its own frameworks—but enforcement is years behind. China is experimenting with guardrails, yet innovation continues unabated. For now, companies can build, deploy, and adapt with remarkable freedom.

Morgan Stanley’s AI wealth management assistant, trained on hundreds of thousands of internal research reports, was rolled out before clear regulatory guidance existed. The firm gained a first-mover advantage that will be harder to replicate once compliance frameworks tighten.

This window won’t stay open. Once the “regulatory-industrial complex” fully awakens—lawyers, auditors, ethics boards, and oversight committees—AI projects will slow, costs will rise, and freedom to experiment will shrink.

The pioneers will already have built the muscle memory of how to move fast and learn safely. Everyone else will be stuck writing policies. And by then, the biggest and most powerful companies will have captured the regulators—negotiating their own private sandpits with complex, compliance-heavy rules that only they can afford to follow. What begins as public safety will harden into private privilege, locking in incumbents and closing the door on new entrants who arrived just a little too late.

4. Outsize Returns Are Still Possible

Big things start small. In the 1990s, a graduate student project became Google. A scrappy online bookstore became Amazon. A simple payment tool became PayPal.

AI is at the same inflection point. Early movers are already turning modest bets into exponential value.

Harvey, the legal AI startup, began as a simple interface to an open API. Two years later, it’s embedded across global law firms, transforming how attorneys draft, review, and reason about complex documents. Cursor, which started as an AI coding assistant, has quietly become the command center for entire engineering teams—an integrated environment where agents not only write code but understand context, maintain memory, and orchestrate multi-step builds autonomously. And Runway, once an experimental video editor, now powers creative pipelines across media and marketing, collapsing production cycles from weeks to hours.

What these examples share is timing: each turned a narrow use case into a new cognitive infrastructure, capturing the kind of leverage that only exists in the early, chaotic phase of adoption.

Outsize returns always accrue in the early, unruly phase of adoption, before efficiency replaces imagination. The same pattern will play out across every industry—from healthcare and finance to manufacturing and education.

5. Ground Truth Still Exists

The final advantage of the current moment is epistemic: the data that feeds AI models is still relatively pure.

We haven’t yet reached the “hall of mirrors” stage where AI systems are swallowing their own synthetic output. But it’s coming. As AI-generated text, images, and code flood the internet, the risk of “model collapse”, where outputs become increasingly derivative, becomes increasingly likely.

For now, models are still largely trained on a foundation of human-authored books, research, and journalism. They reflect eons of human expertise and craftsmanship. That’s what makes today’s results so surprisingly coherent and useful.

But as the signal-to-noise ratio declines, verifying truth will require more time, tokens, and oversight. Leaders will need to invest in data provenance, curation, and trust architectures.

In other words, cognition is cheap now partly because the world’s knowledge base is still intact. Once that begins to erode, clarity will become a premium commodity.

6. The Window Is Closing

Every advantage we’ve just discussed—unfair knowledge, cheap economics, regulatory freedom, outsize returns, and reliable ground truth—is temporary.

Knowledge will spread. Costs will rise. Regulation will harden. Returns will normalize. Data will decay. The AI era won’t end, of course, it will just become ordinary, bureaucratic, and expensive, like every other mature technology before it.

For now, though, we inhabit a rare moment when a mid-sized company can act with the power of a global enterprise, and a single individual can operate with the leverage of an entire team. The smart leaders are not waiting for “maturity.” They are capturing cognitive territory while it’s still unclaimed.

Those who hesitate in periods like this tend to rationalize their inaction. They say they’re waiting for “best practices,” or for “the hype to settle,” or for “clarity on regulation.” But in doing so, they miss the only window when experimentation is both cheap and high-return.

Clarity will come but only after the opportunity has passed.

As I often say in my talks: the future favors the bold.

If you drive past a new data-center build in Texas or Virginia today, you might catch a strange sight: a row of trailer-mounted jet engines idling beside concrete shells and cooling towers. They’re not there to fly—they’re there to think.

These machines once powered Boeing 747s. Now, refitted by companies like ProEnergy, their GE CF6-80C2 cores have been reborn as compact, 48-megawatt gas turbines—each one capable of powering a hyperscale AI facility. It’s an image that could have come straight out of a William Gibson novel: icons of a prior industrial age salvaged from the scrapheap, retrofitted with sensors and software mods, and repurposed as power plants for frontier AIs.

For decades, gas turbines were a symbol of decline. Renewables were ascendant, climate pledges were tightening, and the great manufacturers—GE Vernova, Mitsubishi Heavy Industries, Siemens Energy—were shrinking their divisions. In 2017 Siemens announced nearly 7,000 job cuts, warning of “disruption of unprecedented scope and speed.” Demand for large turbines had collapsed from 400 units a year to barely a hundred.

Yet suddenly, they’re back. The reason isn’t geopolitics or industrial policy—it’s artificial intelligence.

Power Is the New Bottleneck

The explosion of generative AI has transformed the economics of power. Every new data-center campus is a micro-city of computation, each building drawing hundreds of megawatts to train and run models that consume orders of magnitude more electricity than traditional cloud workloads. Utilities from Georgia to Dublin are warning of shortages. The bottleneck in the AI boom is no longer GPUs; it’s gigawatts.

Sam Altman put it succinctly in a recent blog post: “Our vision is simple: we want to create a factory that can produce a gigawatt of new AI infrastructure every week… it will require innovation at every level of the stack, from chips to power to building to robotics.”

That phrase—a gigawatt a week—reframes how we think about compute. For the first time in history, intelligence is scaling like heavy industry. The product isn’t just code; it’s capacity. And energy is now the fundamental input to cognition.

What’s happening beneath the surface is a brutal equivalence between computation and energy. To generate intelligence at scale, we will need to convert ever more power into cognition. Jensen Huang of NVIDIA calls this the “token generation rate per unit of energy.” The efficiency of an AI factory, in other words, can be measured not just in flops or model size, but in how many tokens of useful thought it can produce per megawatt.

That equation changes everything. It ties the fate of the digital economy to the physical grid. It means the next generation of leaders in technology, policy, and finance will need to think like energy strategists.

Thought Thermodynamics

The convergence between compute and energy isn’t just a matter of supply and demand. Something deeper is at play. Every leap in intelligence, human or artificial, is powered by a transformation of energy. The steam engine amplified muscle; electricity amplified industry; computation now amplifies thought. Each wave of progress turns energy into a new form of leverage.

The AI revolution is simply the latest expression of that principle. Training models like GPT-5 or Gemini 2 isn’t an abstract digital process; it’s a physical one, requiring power, cooling, materials, and space. Behind every query sits a chain of turbines, transformers, and transmission lines converting natural resources into cognition.

This is what Altman’s vision captures so well. The AI factory is not a metaphor at all—it’s literal. It’s an industrial stack running from chip fabs to power plants, from robotic assembly to energy markets. And like the steel mills and assembly lines of a century ago, it will define a new geography of productivity and a new class of strategic assets.

As nations race to build sovereign AI capacity, the question of where intelligence lives is becoming inseparable from where energy is available. Data-center clusters are springing up near hydro dams, nuclear plants, and gas hubs. Energy policy is becoming industrial policy for the cognitive age.

There is a certain irony in fossil-fuel machinery powering the most advanced software humanity has ever built. I’d love to see one in person. I can just imagine it: turbines spining beneath sodium lights on the edge of the desert, exhaling heat and noise into the night—half relic, half prophecy.

Jet engines are not the end of the story, they are just a cyberpunk patch - a temporary bridge between the industrial and cognitive eras. The real challenge for leaders is not just developing smarter algorithms, but aligning intelligence with the infrastructure that powers it. In the years ahead, energy strategy will become AI strategy. The organizations that understand that connection first will shape the next economy.