The Agentic Firm: reinventing the company for the age of machine intelligence

A chief executive walks onto a stage and announces the company’s new AI strategy. There is an enterprise agreement with a model provider. There are several internal copilots. There is a customer-service chatbot, an AI innovation lab, a collection of pilot projects, and an ambitious productivity target. The audience applauds. Analysts describe the company as AI forward. Employees are told to experiment.

A year later, the company has thousands of AI users, hundreds of disconnected prototypes, several approved vendors, a growing token bill, unclear ownership, very little financial evidence — and no meaningful change in how it operates.

The organization adopted AI. The firm itself barely changed.

This is the mistake at the centre of enterprise AI, and it is not a mistake of effort. It is a mistake of category. Companies are treating artificial intelligence as though it were simply the next generation of information technology.

mainframe → personal computer → internet → mobile → cloud → AI

That sequence is partly right. AI does need infrastructure, software, data, security, technical talent, and integration work, exactly like the waves before it. But the sequence hides the bigger change. AI is not only a new tool the firm uses. It is a new source of thinking, production, coordination, judgment, communication — and, increasingly, of action — operating inside the firm.

A capable system today can:

• interpret a request and work out what is actually being asked
• generate documents, code, analysis, and plans
• predict, compare, and investigate
• negotiate within limits it has been given
• operate real tools and systems
• carry out a process end to end
• watch what happened afterwards
• improve when it is evaluated

That does not merely change what an employee can do with their software. It changes what an employee is, what a manager is, what a team is, what a process is, what company knowledge is, what capital is, what scale means, where authority sits, where the edge of the company should be, and what society can reasonably ask of a corporation.

Satya Nadella has put the point provocatively: AI should not be treated as a technology initiative, because it concerns the future of the firm. A chief executive who cannot explain the company’s AI capital, its operating model, and the structural advantage it is building may one day look as unprepared as a leader who cannot explain the company’s people, finances, or strategy.

That is a different question. It is bigger than automation. It is bigger than productivity. It is bigger than software. It forces us to reconsider the institution we call the firm — and this essay is an attempt to work through what that reconsideration actually involves.

What is a firm, really?

A firm is so familiar that it feels natural. People work together. Managers assign responsibilities. Departments coordinate. Budgets are allocated. Employees receive salaries. Executives decide. The company sells something.

But firms are not laws of nature. They are one way — not the only way — of organizing economic activity. A restaurant can employ chefs permanently, or it could hire a different chef for every meal. A manufacturer can own a factory, or buy every operation from independent suppliers. A software company can employ engineers, or try to contract for every line of code separately.

Ronald Coase asked the deceptively simple question at the heart of all this in The Nature of the Firm: if markets already coordinate economic activity through prices, why do firms exist at all?

His answer was about the cost of using the market. Using a market is not free. To buy something rather than organize it yourself, you have to find suppliers, explain what you need, negotiate, write the contract, check the work, resolve disputes, and renegotiate when circumstances change. Each of those steps costs time, money, and attention.

A firm appears when doing certain things internally is cheaper than repeatedly contracting for them in the market. That gives us a useful starting point: the size and shape of a company depend on the balance between the cost of coordinating inside and the cost of transacting outside.

AI reduces both. That is what makes its effect on the shape of companies genuinely ambiguous — and genuinely profound.

AI could make firms larger

AI lowers the cost of coordinating more employees, more markets, more product variants, more decisions, more information, and more operational complexity. If one management team can supervise thousands of agents, watch every market continuously, and keep processes aligned across the whole enterprise, then the amount a single firm can sensibly hold gets bigger.

A company might operate in more countries, serve more microsegments, offer more personalized products, run more experiments, and handle more customer interactions — without human coordination costs rising in proportion.

AI could make firms smaller

AI also lowers the cost of reaching outside for capability. A small team can now use agents to do research, product design, coding, financial analysis, customer support, legal preparation, marketing operations, and procurement analysis that used to require whole departments.

So a company can reach the scope of a traditional corporation with a fraction of the permanent workforce.

AI could make the firm more porous

Work may also start moving fluidly between employees, internal agents, external agents, contractors, model providers, platforms, and automated suppliers. The firm becomes less like a closed hierarchy and more like an orchestrated network of intelligence, rights, tools, capital, and relationships.

The paradox

All of this happens at the same time. AI can produce smaller human organizations with larger operational scope; more internal coordination and more external specialization; stronger central control and greater freedom at the edge.

There is no single destination. Some firms will become enormous intelligence platforms. Others will become very small, very high-leverage teams.

The competitive variable is not size. It is whether the firm can choose and govern the right boundary for each activity — deliberately, activity by activity, rather than inheriting a boundary drawn decades ago for reasons that no longer hold.

The firm is more than a cost calculation

Companies also exist because they build up capabilities you cannot simply buy in finished form. A mature company develops routines, trusted relationships, judgment, reputation, culture, operating knowledge, a shared language, problem-solving patterns, and institutional memory.

Much of that knowledge is tacit. It is not written down anywhere. It lives in sentences like these:

• “This retailer will only accept the proposal if supply is secured first.”
• “That machine becomes unstable above this speed.”
• “This kind of promotion looks profitable but usually just pulls sales forward.”
• “This executive says yes indirectly.”
• “This consumer segment notices a difference our research metric misses.”
• “This market needs the launch sequenced differently.”

A company’s advantage often depends less on having information than on knowing which information matters, how to read it, when to make an exception, and what good judgment looks like.

This is where AI changes the theory of the firm a second time. AI can begin converting tacit organizational behaviour into things that can be stored and reused: traces, examples, tools, evaluations, context, reusable skills, and machine-executable workflows.

Nadella describes enterprise tacit knowledge as the accumulated ways people operate, exercise judgment, and demonstrate taste. The strategic challenge is to make human–agent work compound inside systems the enterprise controls, rather than letting the company’s most distinctive knowledge leak away as a side effect of using someone else’s tools.

It is worth being precise about what has actually changed, because “AI” is now used to describe four quite different things.

Software followed instructions

Traditional software performs logic a human specified: if X is true, do Y. It can be enormously complex, but its behaviour is bounded by what was programmed.

Machine learning inferred patterns

Machine-learning systems learned statistical relationships from data. They could forecast demand, classify images, score credit, predict churn, and recommend products. They changed decisions — but they were usually locked inside tightly defined applications.

Generative AI produces cognitive artifacts

Generative models can create text, images, code, summaries, plans, explanations, hypotheses, and simulations. This moved AI right up against the surface of knowledge work.

Agentic AI acts across a trajectory

An agent does not just produce one answer. It pursues an objective over several steps.

interpret objective
   ↓
form a plan
   ↓
retrieve context
   ↓
use tools
   ↓
observe results
   ↓
revise the plan
   ↓
escalate or act
   ↓
evaluate the outcome

The important change is not that the model can write. It is that the system can take part in a course of action.

This agency is still limited. Agents fail. They misunderstand. They often need a great deal of scaffolding to be useful. They have no moral or legal standing. But economically, even bounded machine agency matters, because firms can now provision not just machines that calculate, but systems that carry out parts of coordinated thinking work.

Intelligence becomes a factor of production

Classical production combines labour, capital, land, energy, materials, and technology. The AI firm adds something that resists easy classification. It looks like labour, because it performs tasks. It looks like software, because it can be reproduced. It looks like capital, because it raises output over time. It looks like infrastructure, because other work depends on it. And it looks like a service, because you often rent it by usage.

enterprise output

  =  human capital
  ×  physical and financial capital
  ×  institutional capability
  ×  machine intelligence

The multiplication sign is the whole point. Machine intelligence does not create equal value in every organization. Its return depends on complements the firm has to supply itself: clean data, clear processes, capable employees, defined decision rights, system access, evaluation, customer trust, and operational discipline.

OECD firm-level research points the same way — the productivity advantages associated with AI adoption travel with complementary assets such as digital infrastructure, technical skills, and existing digital capability. The model is increasingly available to everyone. The firm that can use it well is not.

Intelligence can be copied

A talented employee cannot be duplicated. An effective agent configuration essentially can. A useful skill can be deployed to every employee, every market, every customer, every transaction. That creates real scale effects.

But the copy carries everything, including bad reasoning, bias, incorrect policy, security holes, and poor incentives. At machine scale, good judgment compounds — and so does error.

Intelligence can operate continuously

Human work is bounded by time, energy, attention, sleep, physical presence, and fatigue. Agents can monitor and investigate without stopping.

Continuity introduces its own danger. A system that never rests can make more mistakes before anyone notices, accumulate more cost, spread an incorrect assumption further, generate circular agent-to-agent activity, and overwhelm the humans meant to supervise it.

Intelligence can improve through feedback

Nadella describes the enterprise AI system as a hill-climbing machine: give it objectives and evaluations, then improve it through data, reward, and repeated execution. In that view, the company’s evaluation criteria — its definition of what good looks like — become a serious source of proprietary value.

A company that performs a million activities and learns nothing from them is operationally busy and institutionally static. A company that turns each activity into better context, tools, policies, examples, evaluations, and decisions builds compounding intelligence.

The new production function is learning

The industrial firm converted materials into products. The knowledge firm converted expertise into decisions and documents. The agentic firm converts objectives, context, human judgment, models, tools, and evaluation into actions, outcomes, and reusable learning.

The final output is not the completed task. It is the improved ability to complete the next one.

The job was always a bundle

A job usually contains many different activities. A key account manager might analyze sales, prepare customer plans, build presentations, negotiate, resolve deductions, coordinate supply, document commitments, and manage relationships.

Those activities are bundled together partly because it would be inefficient to contract for each one separately. The employee carries context from one activity to the next. The company pays for the whole bundle.

AI makes the bundle separable. That is the quiet structural change underneath most of the noise about job losses.

Work decomposes into six units

Once you see work this way, the automation question stops being about job titles. A title is an accounting convenience. The six units are where the real design decisions live.

The job becomes a portfolio of trajectories

Instead of asking “which jobs should AI automate?”, a far more useful question is: which trajectories inside each role should be performed by humans, by agents, by deterministic systems, or by some combination?

trigger → context → investigation → decision → approval → execution → verification

This is more honest than a list of tasks, because real work unfolds through dependencies and feedback. A task list suggests things can be handed over one at a time. A trajectory shows you where the handover actually has to happen — and where it must not.

Jobs will not disappear neatly

Research on AI exposure keeps finding the same thing: occupations contain a mix of exposed and less-exposed tasks, and exposure is not the same as displacement. The ILO’s 2025 update stresses job transformation more strongly than wholesale occupational automation, and its 2026 guidance warns that exposure measures are indicators of possible change rather than forecasts of employment outcomes.

But “jobs will transform” can easily become a comforting cliché. Transformation can still mean fewer employees, less entry-level hiring, more intense performance expectations, wider gaps between highly-leveraged and less-leveraged workers, weaker bargaining power, and the erosion of apprenticeship paths.

Three futures, all of which will happen

All three will occur. The mix will differ by activity, industry, regulation, and — importantly — by organizational design choices that are being made right now, mostly without much deliberation.

AI can lift up inexperienced workers

In one widely studied customer-support deployment, access to a generative AI assistant increased productivity on average, with much larger improvements among novice and lower-skilled workers. The researchers found evidence consistent with the system spreading the practices of the strongest performers to everyone else.

That is a genuinely powerful possibility. AI can act as coach, translator, guide, quality checker, and distributor of knowledge. It can make expertise available to people who would previously have waited years for it.

AI can also hollow out expertise

The same technology cuts the other way. If employees accept outputs without understanding them, stop practising core skills, never encounter difficult cases, lose access to formative work, and become supervisors before they were ever practitioners, then the organization gains short-term output while losing long-term capability.

The apprenticeship problem is fundamental. Senior experts exist because they once did junior work. If agents do the junior work, where do the future experts come from?

The broken career ladder

Traditional careers follow a sequence:

AI can interrupt that sequence at the start. An employee may begin their career at “review the machine’s answer” without ever developing the knowledge required to review it. That is supervision without mastery, and it is one of the most under-managed risks in enterprise AI.

Cognitive coverage

Nadella introduces a useful idea here: cognitive coverage, by analogy with test coverage in software. Humans need a mechanism through which they understand enough of an agent’s work to learn from it and steer it responsibly.

Cognitive coverage does not require a human to repeat every step. It requires enough understanding to answer a specific set of questions:

1. 01What did the system conclude?
2. 02Which assumptions actually mattered?
3. 03What evidence was used?
4. 04Where was judgment exercised rather than calculated?
5. 05What would make this conclusion wrong?
6. 06What did I learn from this?
7. 07Am I still competent to intervene?

A team that cannot answer these questions about its own workflows is not supervising them. It is watching them.

The future worker is not just an AI user

The worker of the agentic firm needs to become an objective setter, a context provider, a workflow designer, a delegator, an evaluator, an exception handler, a risk owner, and an integrator of human meaning.

Prompt writing is a small part of this. The deeper skill is the orchestration of machine thinking toward legitimate outcomes — which is a management skill, not a technical one, and it is now needed several levels below where management skills were previously required.

From copilot to delegated work

The way AI shows up at work has been moving steadily along a spectrum. Each step hands more of the work to the machine and moves the human further toward steering.

Nadella describes exactly this progression in software development — from completion and chat toward agent modes capable of long-running delegated work — and names the emerging managerial interface as micro steering after macro delegation. You set the direction broadly, then intervene precisely.

A company may have more agents than employees

Imagine a corporation with 20,000 human employees and two million active agents: employee assistants, customer-facing agents, monitoring agents, planning agents, document agents, coding agents, testing agents, negotiation-support agents, quality agents, audit agents.

The traditional organization chart becomes incomplete. It shows reporting relationships among people. It does not show the operating population that actually performs the work.

The new organizational map

One chart is no longer enough. The agentic firm needs several overlapping maps, and most companies currently maintain none of them.

The span of control changes shape

Historically one manager could supervise only so many people. Agents appear to raise that limit — but managing agents creates its own cognitive load. A human supervising a hundred agent threads may well be less effective than a manager supervising ten people.

The bottleneck moves. It shifts from doing the work to prioritization, exception review, conflict resolution, quality control, and keeping the whole thing strategically coherent. The answer to that is not more agents. It is a better environment for managing them.

Managers become designers of control systems

Management moves closer to institutional design. That is a harder job, not an easier one, and it is worth being honest that not every current manager will want it.

Does middle management disappear?

Some middle-management work is heavily exposed: consolidating information, preparing status updates, translating priorities, monitoring completion, scheduling, reporting. AI can do much of that.

But middle management also supplies contextual judgment, conflict mediation, social cohesion, coaching, escalation, political interpretation, and responsibility. The role does not simply vanish. It splits. Weak administrative management becomes less valuable. Strong integrative management becomes considerably more valuable.

New roles the agentic firm will need

Agents are not employees

Calling agents “digital employees” is a useful analogy and a dangerous one. Agents have no consciousness, no dignity, no labour rights, no moral accountability, no human relationships, and no legitimate authority of their own.

But they are not ordinary software either

Traditional software has predictable boundaries. Agents interpret ambiguous instructions, choose among tools, generate intermediate plans, adapt after observations, and behave differently in different contexts. They need governance closer to that of operational actors than that of applications.

The safest classification is a functional one:

An agent is a nonhuman operational actor whose purpose, permissions, context, tools, evaluations, and accountability must all be explicitly designed.

Capital has always meant more than money

A firm holds financial capital, physical capital, human capital, intellectual property, relational capital, brand capital, and organizational capital. AI introduces another asset class, and most companies are currently generating it by accident or not at all.

Token capital includes:

• enterprise context — what the company means by its own terms
• agent skills that have been tested and reused
• workflow traces of how real work was actually done
• evaluation rubrics that define good in this business
• decision policies with thresholds and owners
• tool interfaces to the systems that matter
• domain-specific models trained on the firm's own problems
• institutional memory with provenance
• human feedback and expert corrections
• reusable trajectories that start the next job halfway through

Token expense versus token capital

Consider two companies that each spend €10 million on AI in the same year.

The defining capital question

A chief executive should be able to answer one question:

What did our firm learn from yesterday’s work that it can now do better, more safely, or more cheaply tomorrow?

If the answer is “nothing”, the company is consuming intelligence without accumulating capability. That can go on for years while looking, from the outside, like a successful AI programme.

Evaluations are intellectual property

An evaluation defines what good performance means. That definition is specific to your business, and it is much harder to copy than any model.

The model can be replaced next quarter. A high-quality, hard-won definition of success cannot. Nadella’s hill-climbing framing makes this explicit: objectives, rubrics, reward dimensions, and the data used to improve the system can end up more differentiating than the generic model underneath.

Strategy becomes learning-loop design

A durable advantage probably does not come from having today’s best model. Model rankings change. Costs fall. Open models improve. Providers compete relentlessly.

The advantage comes from having the best loop:

What belongs on the balance sheet?

Traditional accounting will not recognize token capital as a separate asset any time soon. But management still needs internal capital discipline, because what a company measures determines what it builds.

A firm that measures seats, prompts, active users, and tokens consumed will optimize activity. A firm that measures capability will build it.

No firm produces intelligence alone

A single enterprise AI outcome may depend on electricity, data centres, semiconductors, cloud infrastructure, foundation models, open-weight models, data providers, enterprise context, agent platforms, tool integrations, and human expertise.

That is a supply chain, and it deserves the same rigour as any other. Nadella expects firms to increasingly need to understand how their AI supply chains compound value while preserving meaningful organizational boundaries and ownership.

A company can look diversified at the application layer while depending entirely on a single foundational layer underneath. The diversification is cosmetic.

Frontier dependence

The most capable model is appropriate for novel scientific reasoning, ambiguous strategic analysis, difficult coding, and low-frequency high-value judgment. It is excessive for extracting a standard debit note, applying a rebate rule, checking a known policy, matching structured records, or routing an approval.

The model portfolio

A mature firm does not have “an AI vendor”. It has a portfolio, and it routes work to the cheapest thing that reliably does the job.

The objective is not maximum intelligence per request. It is the required outcome at the appropriate cost, latency, privacy, reliability, energy use, and governance level.

Token efficiency is not using fewer tokens

A cheap model that produces repeated errors is expensive. A frontier model doing trivial deterministic work is also expensive. Both mistakes are common, and they are usually made by different parts of the same company.

Energy becomes strategic again

AI feels weightless because its outputs arrive digitally. Its infrastructure is not weightless at all.

The International Energy Agency estimates data-centre electricity consumption is growing rapidly, with AI-focused facilities a major driver. Its 2026 update projects data-centre electricity use rising from roughly 485 TWh in 2025 to around 950 TWh by 2030 in its central case, while stressing significant uncertainty and local grid bottlenecks.

So the agentic firm depends on energy availability, grid capacity, cooling, water, hardware supply, geopolitics, and capital markets. AI strategy quietly reconnects digital strategy to industrial policy.

Sovereignty is operational, not rhetorical

Enterprise sovereignty does not require owning every model or data centre. It requires the ability to choose among providers, move workloads, protect critical context, preserve learning, maintain continuity, avoid irreversible lock-in, and operate within the jurisdictions you are subject to.

Sovereignty is the preservation of strategic choice.

Cheap thinking does not make strategy automatic

Strategy is not the production of more analyses. It is the commitment to a course of action under uncertainty. AI can generate more scenarios, more research, more competitor analysis, more forecasts, and more options. That may improve strategy. It may also create analysis abundance and commitment scarcity.

Attention becomes the scarce resource

When analysis is expensive, organizations ration analysis. When analysis is abundant, they have to ration attention instead. The scarce inputs become executive focus, organizational commitment, the willingness to stop things, and clarity of purpose.

The value of a good question rises

When machines can produce many answers, the quality of the question matters more. The company has to define which outcome matters, which trade-offs are acceptable, what evidence would change the decision, what must not be optimized away, and whose value is being counted.

Strategy becomes an objective-function problem

Strategy is what determines the complete objective. In an agentic firm, writing the objective function is a strategic act, not a technical detail delegated to whoever configures the system.

Taste becomes more valuable

Taste is the ability to tell adequate from exceptional, locally correct from generally plausible, coherent from merely impressive, and meaningful from statistically optimized. As average-quality production becomes abundant, high-quality selection becomes the constraint.

The agentic firm needs leaders who can define excellence, not only request output.

The strategy process becomes continuous

This does not mean changing strategy every day. It means maintaining a living model of the conditions under which the strategy should change — and noticing when those conditions arrive rather than when the calendar says so.

Strategy and operations converge

When a firm can continuously observe customers, update forecasts, simulate supply, measure promotions, track commitments, and reallocate resources, the boundary between strategy and operations blurs. Strategy becomes embedded in decision rules. Operations become a continuous test of strategy.

Hierarchy was partly a bandwidth solution

Organizations built layers because information and authority could not flow directly among everyone. A hierarchy compresses complexity: many signals go to a manager, who produces a summary, which goes to a senior manager, who makes a decision.

Agents change the compression mechanism. They can summarize, detect exceptions, preserve the underlying detail, simulate options, and route decisions. Once compression is no longer the reason for the layer, the layer needs a different justification.

The risk of hypercentralization

If executives can watch every transaction and every employee through AI, authority can become extremely centralized. The centre may conclude it no longer needs local judgment. That produces something like an algorithmic command economy inside the company.

But local reality stays complex, and no central model has perfect context.

The agentic firm must avoid confusing visibility with understanding.

The opportunity for intelligent decentralization

The same technology can equip frontline employees with enterprise knowledge, scenario analysis, policy guidance, and decision support — which lets more decisions move toward the edge while staying governed.

Functions become capability platforms

Finance, HR, legal, and IT can shift from processing every request to providing policies, tools, agent skills, evaluation, exception handling, and assurance.

Finance stops preparing every analysis by hand and instead maintains the governed financial intelligence used across every agent. Legal maintains contract tools, policy rules, review triggers, and approved language. The function becomes a platform other people build on.

Teams may form around decisions

Instead of permanent structures based mainly on expertise, organizations can assemble temporary human–agent teams around a launch, a customer negotiation, a supply crisis, a strategic question, or a quality incident. Agents can pull together the relevant information, models, experts, prior decisions, and tools within minutes.

The unit of organization shifts from the department to the outcome.

Stable underneath, fluid on top

Management was built for information scarcity

Managers historically spent much of their time collecting reports, distributing instructions, monitoring completion, and filling information gaps. AI reduces all four. What remains is more demanding, not less.

Management becomes the allocation of agency

A manager now allocates work among themselves, employees, agents, automated workflows, contractors, and specialist systems. The central question becomes: which actor should exercise which form of agency, under which constraints?

The manager as objective setter

An agent cannot work out the legitimate objective from operational data alone. Someone has to decide whether the priority right now is growth, margin, safety, resilience, learning, fairness, or speed. The manager becomes responsible for the quality of the objective itself.

The manager as evaluator

Managers must define acceptance criteria, failure thresholds, escalation conditions, and evidence standards. Evaluation moves from an afterthought to a core managerial skill — arguably the core one.

The manager as coach

Employees still need feedback, recognition, psychological safety, development, meaning, and belonging. An automated work allocator cannot replace the human relationship through which people become capable and committed.

The manager as moral agent

An AI system can optimize a workforce reduction. It cannot legitimately decide what the company owes to affected employees, communities, customers, or society. Managers remain responsible for the human consequences of organizational choices, and that responsibility is not transferable to a system.

Productivity is not activity

AI reliably increases documents written, code generated, analyses produced, meetings summarized, and messages sent. Those are outputs. They are not outcomes.

The productivity illusion

An organization can generate twice as many presentations while making no better decisions. It can produce more code while accumulating technical debt. It can send more marketing content while becoming less distinctive. It can resolve more customer contacts while never fixing the reason people are contacting it.

Most AI programmes report the top two layers because they are easy to instrument. The bottom two are the only ones that compound.

The evidence is mixed, and that is informative

Field research has found meaningful productivity gains in structured tasks such as customer support. Recent executive surveys report positive but very uneven firm-level effects. At the same time, deep integration remains far less common than broad experimentation, and macroeconomic gains appear to depend on complementary organizational redesign rather than on tool access alone.

The AI J-curve

General-purpose technologies usually require process redesign, new skills, complementary infrastructure, and organizational learning before big returns appear. During the transition, companies pay for duplicated work, training, integration, governance, temporary productivity loss, and vendor experimentation.

The wrong conclusion is “AI created no value”. The equally wrong conclusion is “any AI investment will eventually create value”. The dip is normal; the recovery is earned. Transformation has to be managed and measured, not simply believed in.

Metrics worth building

The document was the unit of knowledge work

Knowledge workers produced reports, presentations, emails, spreadsheets, and documents, and those artifacts became the organization’s memory. But a document records a conclusion without the path that produced it. You inherit the answer and lose the reasoning.

Trajectories become first-class artifacts

A human–agent trajectory can preserve the objective, the context retrieved, the tools used, the assumptions made, the alternatives considered, the feedback given, the decision taken, and the outcome that followed. That is a far richer record than a static report.

Imagine a customer-negotiation system that knows which package was proposed, why it was chosen, which assumptions mattered, what the retailer accepted, what was actually executed, and what economic value followed. The next negotiation starts with institutional learning instead of one person’s recollection.

Memory requires governance

Not everything should become organizational memory. A firm must avoid preserving unsupported rumours, discriminatory assumptions, personal speculation, obsolete policy, confidential cross-customer information, and unverified conclusions.

Forgetting is an organizational capability

A firm that never forgets will retain bias, preserve errors, violate privacy, apply outdated context, and make experimentation dangerous. The agentic firm needs deliberate forgetting, designed as carefully as remembering.

Who owns the learning?

One of the hardest questions in this whole area: who owns the learning produced when a person works with AI?

The resulting agent skill emerges from all three. Current intellectual property and employment frameworks do not cleanly capture that hybrid production, and the issue will become central to employment contracts, model training terms, worker mobility, trade secrets, compensation, and collective bargaining.

Governance cannot be a policy document

A company can publish principles — human oversight, transparency, fairness, security — and change nothing. Operational governance requires architecture, not a statement.

The governed agent estate

Every enterprise agent should have a complete record. If you cannot produce this for every agent in the company, you do not have an estate; you have a population.

Nadella emphasizes the same requirements from the platform side: agent inventory, identity, inspectability, auditability, sandboxing, security, data protection, and execution-time assertions.

Agent identity

The enterprise must always be able to say which agent acted, for which user, under whose delegated authority, using which model and tools, at what time, on which data. Anonymous agency is incompatible with accountability.

Permissions must be narrow

The authority ladder

Rather than arguing about whether agents should be “autonomous”, define levels and move workflows up the ladder only when the evidence supports it.

Runtime assertions

Governance has to exist during execution, not only before deployment. A pre-deployment review is not enough if the system can deviate at runtime.

never     release quarantined inventory
never     approve a claim above available authority
never     send customer pricing without approval
stop      if customer identity is ambiguous
escalate  if confidence falls below threshold

Observability

The firm must be able to reconstruct what the agent observed, what it inferred, which tool it called, what result it received, why it escalated, and what action followed. How much reasoning to record is a real trade-off between audit, security, privacy, intellectual property, and technical feasibility — but “none” is not an option.

Human-in-the-loop is not a magic phrase

Human review fails when:

• reviewers are overloaded
• approvals become habitual
• the context behind the recommendation is hidden
• responsibility is ambiguous
• the review happens too late to change anything
• the human does not have the expertise to judge

The correct design is not adding an approval button. It is creating a meaningful human decision.

Approval fatigue

If employees receive hundreds of agent approvals a week, they will approve without thinking. Good governance automates low-risk activity, aggregates similar decisions, prioritizes material exceptions, shows the consequence clearly, and periodically tests whether reviewers are actually paying attention.

Independent evaluation

The agent that proposes an action should not always be the only system assessing it. High-impact workflows deserve deterministic validation, a specialist evaluator, a policy engine, or independent human review — preferably more than one of them.

What the board should ask

1. 01Which decisions are delegated to AI?
2. 02What is the largest possible uncontrolled action?
3. 03Where does agent activity affect employees or customers?
4. 04Which capabilities depend on a single provider?
5. 05How is our proprietary learning protected?
6. 06How do we know the systems are working?
7. 07What happens during failure or attack?
8. 08Who, by name, is accountable?

AI governance is becoming part of corporate governance.

Regulation is moving from principle to obligation

The EU AI Act entered into force in 2024 and is applying on a phased timeline, covering prohibited practices, AI literacy, general-purpose models, and high-risk systems. Exact requirements depend on your role and use case, but the direction is unambiguous: organizations increasingly need lifecycle documentation, risk controls, human oversight, and accountability rather than voluntary experimentation.

The NIST AI Risk Management Framework points the same way, treating AI risk as a lifecycle responsibility — govern, map, measure, manage — with trustworthiness considerations running across design, deployment, use, and evaluation.

AI can democratize capability

A small company can now reach sophisticated analysis, multilingual communication, software production, legal preparation, creative capability, and operational planning that used to require serious institutional scale. That genuinely lowers barriers to entrepreneurship.

AI can also concentrate power

Frontier AI depends on large-scale compute, expensive training, semiconductor supply, cloud platforms, vast capital, and proprietary data. The infrastructure layer has significant scale economies. OECD analysis has begun examining competition across AI infrastructure and the strategic risks created by concentration in compute, cloud, chips, and adjacent layers.

Decentralized capability, built on centralized infrastructure. That is the paradox of democratized intelligence.

Platform power

A model or agent platform can influence what firms are able to build, what behaviour is permitted, how data is processed, how costs evolve, which businesses stay viable, and how learning is captured. The provider becomes part utility, part labour market, part infrastructure company, part knowledge intermediary — a combination we do not have good institutional precedent for.

Corporate sovereignty

The agentic firm must avoid becoming a distribution layer for another company’s intelligence. It should preserve ownership or control of its proprietary context, its evaluations, its customer relationships, its operating knowledge, its learning trajectories, and its strategic choices.

Antitrust questions

Policy makers will need to consider model-provider concentration, cloud–model integration, chip supply, data advantages, platform self-preferencing, agent marketplaces, switching costs, and interoperability.

The future competitive unit may not be the model at all. It may be the entire AI supply chain.

Efficiency is not neutral

An agentic workflow decides who is monitored, who is trusted, who receives discretion, who is evaluated, who carries risk, and who benefits from productivity. The architecture contains a labour philosophy whether leadership acknowledges it or not.

Algorithmic management already exists

Organizations already use data-driven systems to allocate work, monitor employees, evaluate performance, schedule shifts, and direct workflows. ILO research documents both efficiency benefits and real risks around surveillance, work intensity, autonomy, privacy, and psychosocial well-being.

Agentic systems can intensify those tensions, because they can interpret employee behaviour, generate managerial recommendations, personalize pressure, monitor continuously, and initiate actions on their own.

The intelligent panopticon

An employer could, in principle, know:

• which employee performed each action
• how long it took them
• which information they accessed
• whether their language showed uncertainty
• how often they disagreed with the AI
• whether they followed the recommended steps

This may improve quality. It may also destroy autonomy and trust, and those two outcomes are not mutually exclusive.

The right not to be reduced to a metric

Human performance contains dimensions that resist quantification: care, creativity, moral courage, mentoring, trust, community, long-term judgment. A firm that values only machine-visible output risks becoming efficiently inhuman.

Worker participation

Employees should not always meet AI as a system imposed on them. Social dialogue can help determine which systems are introduced, which data is collected, how performance is evaluated, what appeals exist, how gains are shared, and how roles are redesigned. ILO case studies suggest worker voice supports approaches that complement skills, protect autonomy, and keep change inside labour and social protections.

Who receives the productivity dividend?

AI increases output. Where the gain goes is a separate question.

The distribution is not determined by the technology. It is determined by institutions, market power, ownership, and politics.

The four-day week or the twelve-person department?

If AI doubles productivity, a company can produce twice as much, employ fewer people, reduce working hours, improve quality, lower prices, or expand into new markets. Different organizations will choose differently, and the choice reveals values and incentives rather than technical constraints.

Pro-worker AI

Recent research distinguishes AI that substitutes for labour from AI that expands human capability, creates new tasks, and makes expertise more valuable. Acemoglu, Autor, and Johnson argue that market incentives may under-produce these pro-worker forms unless companies and policy makers deliberately support them.

The future of work is not technologically predetermined. It is designed.

A corporation operates with social permission

The firm depends on society for educated workers, infrastructure, legal enforcement, property rights, energy, communities, customers, and political stability. In return, society expects more than legal compliance. It expects the company to create legitimate value.

AI may weaken that permission

Public resistance will grow if AI is experienced mainly as layoffs, surveillance, deskilling, environmental burden, concentration of wealth, deception, or exclusion. A promise of eventual productivity will not be enough.

Nadella argues that the AI industry has to produce practical, widely understood outcomes rather than abstractly promising future jobs or prosperity. Companies, communities, and countries need proof that AI creates positive-sum value and does not merely extract their data or displace their agency.

What a positive-sum transformation looks like

The environmental contract

AI’s physical infrastructure affects electricity prices, water, land, grid investment, emissions, and local communities. A company should not describe AI as immaterial. It should account for the infrastructure that makes machine intelligence possible.

The community question

When a company builds or uses AI infrastructure, communities will reasonably ask who gets the jobs, who pays for grid expansion, who bears the environmental impact, who gains the tax revenue, what happens to electricity affordability, and what public value results. Social permission has to be earned locally as well as globally.

A corporate issue and a state issue

Companies determine how AI is deployed. Governments determine rights, liability, competition, taxation, education, infrastructure, labour protection, public procurement, and national security. The agentic firm will emerge through the interaction of private strategy and public institutions, not from either alone.

National comparative advantage

AI tends to amplify existing strengths. A country strong in manufacturing, logistics, agriculture, finance, science, design, or public administration can use AI to deepen that capability. But access to AI infrastructure and skills is uneven, which means amplification is also divergence.

The Global South

AI could let countries leapfrog limited access to expertise, education, health knowledge, business services, and software capability. It could also create new dependency on foreign models, foreign cloud providers, foreign data centres, and foreign standards.

Digital sovereignty cannot be reduced to storing data locally. It requires the capacity to create, adapt, govern, and benefit from AI.

Education and the new firm

If firms increasingly value objective setting, evaluation, synthesis, judgment, and human collaboration, then education systems built around reproducing answers become less aligned with work. The new divide runs between people who can direct abundant intelligence, interrogate it, evaluate it, and combine it with real domain experience — and people who cannot.

Taxation

If firms produce more with fewer employees, governments face lower labour-tax revenue, higher capital income, regional disruption, and greater retraining needs. Policy debates will include corporate taxation, capital taxation, social insurance, wage subsidies, training accounts, shorter working weeks, and universal basic services or income.

No formula will fit every society. But ignoring the distributional question will not prevent it from arriving.

Corporate political power

A company that controls models, infrastructure, data, identity, work platforms, and information flows exercises power beyond ordinary market participation. Democratic institutions will need to ensure that private intelligence infrastructure does not quietly become unaccountable public governance.

What happens when ten people can do the work of ten thousand?

The possibilities include extraordinary entrepreneurship, lower barriers to entry, winner-take-most competition, reduced employment, extreme founder leverage, and much weaker organizational checks. A tiny firm can create enormous societal consequences without any of the governance infrastructure a traditional corporation carries.

What happens when agents transact with agents?

Agents may negotiate procurement, buy advertising, manage inventory, contract services, and allocate capital. Machine-speed markets could improve efficiency. They could also generate rapid collusion, cascading errors, flash crises, opaque pricing, and commitments nobody intended. The law will have to determine when an agent’s action binds the firm.

Can an agent be a fiduciary?

An agent may advise a board or allocate investments. But fiduciary duty requires loyalty, care, accountability, and legal responsibility. An AI system possesses none of these in the human or legal sense. The responsible humans and the institution cannot delegate the obligation away.

Can a company become immortal?

An agentic firm can preserve founder decisions, employee expertise, operating patterns, and institutional memory far beyond any individual’s tenure. That could strengthen continuity. It could also trap the company inside the judgment of its own past.

Who has the right to challenge the corporate memory?

If an agent says “historically, we do not promote people with this profile”, the company must ask whether that history represents valid experience, outdated context, or structural discrimination.

What happens to trust?

We have historically trusted organizations partly because identifiable people stood behind decisions. An agentic organization may communicate through machines at every interface. Customers and employees will ask: is this a real commitment? Can a human be reached? Who is responsible? Is this being recorded? Am I negotiating with a system optimized against me?

Trust has to become an explicit design requirement.

What happens to corporate culture?

Culture is learned through stories, observation, rituals, incentives, promotion, and social interaction. Agents will encode culture into daily recommendations. They can strengthen it. They can also freeze slogans into rigid rules, or quietly amplify the unofficial biases nobody wrote down.

What remains uniquely human?

No permanent boundary can be guaranteed. But several responsibilities remain central: choosing legitimate ends, accepting moral responsibility, forming genuine relationships, caring, creating shared meaning, exercising judgment where values conflict, granting democratic legitimacy, and experiencing consequences.

Human dignity is not a temporary technical advantage. The mistake is to define human worth by whatever machines cannot yet do.

Most of these are not technology failures. They are design failures, and every one of them is visible early to anyone willing to look.

Pulling all of this together produces a stack. It is not a technology diagram — the top two layers contain no technology at all — but it is the shape a firm needs if machine agency is going to operate inside it safely.

REFOUND is our framework for rebuilding a firm around human–agent intelligence. Seven moves, in order, because each one depends on the one before it.

The phases overlap on purpose. Nothing here waits for the previous stream to finish, because the learning from each one changes the next.

An agentic firm should not be defined by the number of agents it deploys. It should be defined by the quality of the institution it becomes.

A strong agentic firm:

• treats AI as a source of capability, not spectacle
• redesigns work rather than automating appearances
• owns and compounds its distinctive knowledge
• uses frontier intelligence only where frontier intelligence is needed
• gives every agent an identity, an owner, a boundary and an evaluation
• preserves meaningful human authority
• develops employees rather than making them cognitively dependent
• decentralizes judgment where local knowledge matters
• measures outcomes rather than activity
• shares enough of the productivity value to retain legitimacy
• accounts for the physical and social infrastructure it consumes
• accepts responsibility for every delegated action

The goal is not a company without people. The goal is also not a company where people perform ceremonial approvals for machines. The goal is a firm in which human and machine capabilities are composed deliberately.

The strongest institution combines them without pretending they are the same.

The industrial firm was built around physical capital. The managerial firm was built around hierarchy and administration. The knowledge firm was built around educated labour and digital information. The agentic firm will be built around the continuous interaction of human capital, token capital, machine agency, enterprise context, institutional authority, and social legitimacy.

This is not a software upgrade. It is a refounding moment, and it is worth being explicit about which assumptions are being retired.

But several things remain genuinely scarce:

• legitimate purpose
• trusted relationships
• responsibility
• attention
• physical resources
• high-quality judgment
• social permission

The future firm will succeed by understanding both sides of that ledger. It has to exploit the abundance without destroying the scarcity that gives the institution meaning.

The fundamental question is not how many employees AI can replace. Nor is it how many agents we can deploy. It is:

What combination of humans, machines, capital, authority, and learning will let this institution create more legitimate value than any other way of organizing the same activity?

Which takes us back to where we started. Why does the firm exist? In the AI age, the answer may be this:

A firm that does not learn will be outperformed by one that does. A firm that does not govern may be damaged by its own agency. A firm that does not develop its people will hollow itself out. A firm that does not create shared value may lose permission to operate.

So the winners of the AI era will not simply be the companies with the best models. They will be the companies that build the best institutions around them. They will know what they are trying to become, what knowledge they uniquely own, where machines should act, where humans must decide, how learning compounds, how power is constrained, how value is shared, and how trust is earned.

AI is not a technology. It is not even merely the future of work. It is the force that compels us to reconsider the future of the firm.

And that future is still ours to design.

The theory of the firm. Ronald Coase’s The Nature of the Firm established the question of why economic activity is organized inside firms rather than entirely through markets, linking the answer to the costs of using the price and contracting system.

The firm-level impact of AI.Recent NBER research emphasizes distinguishing invention, adoption, internal capability building, outsourcing, and actual usage when assessing AI’s effects on firms. Other studies associate AI investment with changes in workforce composition and organizational hierarchy, while 2026 executive evidence points to heterogeneous productivity effects across firms and sectors.

Adoption versus deep integration.Stanford’s 2026 AI Index reports widespread organizational AI use but much earlier-stage deployment of agents across business functions. A separate 2026 study of large US companies similarly distinguishes general AI usage from deep integration into production and business processes.

AI, tasks, and jobs. ILO and IMF research emphasizes that occupational exposure varies by task, country, worker group, and potential complementarity. Exposure does not mechanically determine job loss, but labour-market effects can be uneven and require serious attention to reskilling, job design, and worker protection.

Pro-worker versus substitutive AI. Recent economic research distinguishes labour-augmenting, automating, expertise-levelling, capital-augmenting, and new-task-creating technologies, arguing that market incentives may not automatically generate the forms of AI most supportive of worker capability and demand.

Algorithmic management and worker autonomy. ILO research documents the growing use of algorithmic systems to allocate, monitor, and evaluate work, highlighting both efficiency gains and risks concerning surveillance, privacy, work intensity, autonomy, and psychosocial well-being.

AI governance.NIST’s AI Risk Management Framework and Generative AI Profile provide lifecycle-oriented approaches to governing, mapping, measuring, and managing AI risk. The EU AI Act establishes phased legal obligations for providers and deployers, with requirements varying by role and risk category.

AI infrastructure and energy.The IEA’s 2025 and 2026 analyses document rapid growth in data-centre investment and electricity use, the increasing contribution of AI-focused data centres, and the significance of grid, energy-security, affordability, and sustainability constraints.

Competition and infrastructure concentration. OECD analysis examines competition issues across AI infrastructure, including scale, capital intensity, cloud, compute, and related bottlenecks that may shape market entry, dependence, and innovation.

The source conversation. This essay is also grounded in the discussion between Satya Nadella and Reid Hoffman, particularly its treatment of AI as the future of the firm, the interplay of human and token capital, enterprise tacit knowledge, the AI supply chain, human–agent work, cognitive coverage, agent governance, model efficiency, and the need for tangible positive-sum outcomes.