The Four Operating Models of the AI Era

2025 was the year the enterprise bought AI. Nearly every large company now runs a copilot, a chatbot or an agent pilot somewhere in the business, and most can point to a team that moved faster because of it. What far fewer can point to is a single function that was redesigned around the technology rather than handed a faster tool to do the old job. That gap is the most important fact in enterprise AI right now, and it is worth being precise about why it exists.

There are two familiar ways to get this wrong. The first is to sit on the sidelines, treat AI as another overhyped technology, wait for the dust to settle and preserve the operating model that got you here. The second is to break the glass and go all in, pouring capital into tooling on the theory that more licenses, more consumption, more pilots and more seats must eventually add up to transformation. Both fail, and for the same underlying reason: neither one touches the structure of the organization.

The disciplined move sits between them, and it starts with an uncomfortable diagnostic question. Not how much AI have we bought, but which operating model are we actually running? A copilot is not an operating model. It is a faster tool bolted onto whatever operating model you already had. And the clearest lesson of the enterprise computing wave of the 1990s, the last time firms bought a general-purpose technology at this scale, is that bolting a powerful new tool onto an un-redesigned organization is precisely how you underperform it.

Why the org chart resists AI

To understand why redesign is so hard, and so valuable, it helps to ask what the org chart was ever for.

Start with the Roman legion, whose structure Block traces in its recent essay From Hierarchy to Intelligence as the origin point of the modern org chart. Eight soldiers shared a tent and a mule under a decanus; ten of those made a century under a centurion; six centuries made a cohort; ten cohorts made a legion of roughly five thousand. The structure, 8 to 80 to 480 to 5,000, was not a theory of leadership. It was an information-routing protocol built around a single human limitation: a commander can effectively coordinate somewhere between three and eight people. Two thousand years later we call that limitation span of control, and it still governs the shape of every large organization on earth.

Everything that followed was a variation on the same problem. Prussian general staff, the postwar matrix, the McKinsey 7-S framework, and more recently the experiments run by the technology sector, each was an attempt to move information through an organization faster without breaking the tradeoff at the core of hierarchy: narrowing the span of control means adding layers, and every layer you add slows the signal down.

The flat-organization experiments Block catalogues are the tell. Spotify popularized cross-functional squads with no conventional management chain, then moved back toward traditional structures as it scaled. Zappos eliminated managers and titles outright under Holacracy, and saw significant attrition. Valve ran with no formal hierarchy at all, and the model proved difficult to sustain much beyond a few hundred people. Each experiment revealed something real about the limits of hierarchy. None of them replaced it. Organizations that grow into the thousands revert to hierarchical coordination, because until now no alternative routing mechanism has been powerful enough to do what the managers were doing.

This is not a new observation dressed in AI language. It is one of the most durable results in economics. Ronald Coase asked in 1937 why firms exist at all, if markets are supposed to be efficient, and answered that a firm is what you get whenever coordinating an activity inside a hierarchy is cheaper than buying it through the market. The boundary of the firm sits exactly where those two costs cross. Later work sharpened the point. Friedrich Hayek showed that the hard problem in any economy is using knowledge that is dispersed, local and never possessed by any single mind. Luis Garicano modeled the hierarchy itself as a knowledge-routing device, in which routine problems are handled at the front line and only the rare, hard exceptions escalate to a specialist, because it is cheaper to route the exception to the person who already knows the answer than to make everyone an expert. In each case the conclusion is the same. The layers exist to move information. The manager, in the most literal sense, is a router.

Which yields the observation this entire piece turns on. If the organization is fundamentally a communication network, and Bolton and Dewatripont modeled it as exactly that, deriving the optimal number of layers from the tradeoff between communication cost and processing capacity, then a technology that collapses the cost of communication does not merely make the network faster. It changes the network’s optimal shape. Hierarchy was never sacred. It was the cheapest available answer to a coordination problem, and for the first time there is a cheaper one.

The most sophisticated wrong answer

Before naming the alternative, it is worth taking the incumbent view seriously, because the best version of it is genuinely good.

A clear articulation of the human operating model in print is a 2025 book, Rethinking Operating Models, written by practice leaders at a leading consulting firm and a winner of the NYC Big Book Award. Its central diagnosis is strikingly close to the one above. It places the real source of organizational agility in the middle tiers of leadership, two to four levels beneath the chief executive, where strategy is translated into execution and global is reconciled with local. It calls that middle layer the element most often absent, and the hardest of all to build.

That is the correct diagnosis. The prescription is where it turns. Having identified the coordination layer as the most fragile and most valuable thing in the organization, the book’s answer is to invest in it, to staff it better, train it harder and design it more carefully. It even concedes the mechanism that undoes its own conclusion, noting that generative AI is sharply lowering the cost of structuring data and putting a single source of truth within reach, then using that concession to prop up the existing structure rather than to question it.

This is the pattern worth noticing across the market. The smartest incumbents have correctly identified the coordination layer as the constraint, and they are now trying to perfect the very layer that the technology has finally made it possible to remove. You do not reach the next operating model by building a better middle layer. You reach it by asking what the organization looks like when the middle layer no longer has to exist.

Four operating models

If you array enterprises by how they actually coordinate work, four distinct archetypes emerge. The first three sit on a single continuum. The fourth does not, and that difference is the whole point, so hold it in mind until we get there.

Hierarchical. The default state, and still the honest description of most large organizations. Humans are the coordination mechanism. Information travels up the chain to where decisions are made and back down to where work happens, accreting latency and distortion at every hop. AI, where it exists, is ornamental: a pilot in a corner, a chatbot on the website. This is the operating model Coase would recognize from 1937.

AI-Augmented. The copilot archetype, and the one where the overwhelming majority of “AI transformation” actually lives. Individuals get faster. The analyst drafts quicker, the developer ships more code, the marketer generates more variants. But the org chart is untouched, the decision rights are unchanged and the coordination layer is fully intact, now simply feeding faster inputs into the same slow structure. This is the copilot plateau, and it is a more dangerous place to be than it looks, because the local speed-ups feel like progress while the operating model stands exactly where it was.

We have run this experiment before. When firms computerized in the 1990s, the productivity payoff did not come from the hardware. Erik Brynjolfsson and Lorin Hitt showed that the returns to information technology depended almost entirely on the complementary organizational redesign, meaning decentralized decision rights, redesigned processes and new skills, and that companies which bought the technology without redesigning the work captured little of its value. It is the same finding, running again on new hardware. And the coordination tax it leaves untouched is not abstract. Microsoft’s own workplace telemetry finds the average employee is interrupted every two minutes by a meeting, message or notification, roughly 275 times a day, and spends 57% of the workday communicating rather than creating. A copilot makes the communicating faster. It does not ask why so much of it was necessary in the first place.

AI-Native. Here the structure itself changes. The system, not the hierarchy, holds a live model of how the business works, and uses it to coordinate. The management layers whose job was to route information thin out, because the routing now happens in software. What remains is a workforce organized around the edge, where judgment actually lives. Block describes the resulting shape well: deep specialists who build and operate the capabilities, owners accountable for specific outcomes with the authority to pull resources across the business, and player-coaches who develop people while still doing the work rather than sitting in the status meetings that used to constitute management. This is not a flatter org chart. It is Garicano’s escalation hierarchy with the router replaced: routine work handled autonomously, judgment reserved for humans, the matching function absorbed by intelligence. Microsoft, describing the same shift in its own research, calls it moving from the org chart to the “Work Chart”, where small teams assemble around a goal and use AI to close the skill gaps, rather than functions that turn a simple price change into days of meetings and handoffs.

The first three archetypes are answers to the same question: how should work be coordinated? You could stop here, call it a maturity curve and tell everyone to climb it. Most frameworks do. That would miss the most important shift of all.

Outcome-Native. The fourth archetype is not the last step on that continuum. It is a different object. The first three ask how to coordinate the work. This one asks what the firm is even for, and re-derives the answer from scratch.

Consider a function your company outsourced twenty years ago: claims processing, accounts payable, tier-one support. It was externalized not because it was unimportant, but because coordinating it inside the firm cost more than buying it from a specialist who had scale. That was a Coasean calculation, and at the time it was correct. Now collapse the coordination cost and run the calculation again. The work comes back inside, running against the same live model of the business as everything else. And the vendor contract, the service-level agreement, the quarterly business review and the account team that existed to manage the handoff all dissolve with it. Nothing about the work changed. The boundary moved.

The same logic runs outward. An industrial supplier that sold components can begin selling the outcome those components produce, guaranteed uptime rather than parts, because it can finally observe and coordinate that outcome at acceptable cost. A handful of firms have done this for decades, but only those able to fund bespoke telemetry and a dedicated coordination apparatus to support it. That cost is now collapsing, and with it the barrier that kept outcome-based models the preserve of a few. The unit of value stops being the task or the role and becomes the delivered outcome itself. Structure, pricing and the boundary of the firm all re-derive from that outcome, rather than the outcome being assembled from a structure that already exists.

The distinction matters because it changes what transformation is for. The first three archetypes are, ultimately, about doing the existing business more efficiently. That is real value, but it is value measured in cost and speed. Outcome-Native is about what you sell, how you sell it and how you create it in the first place: the higher-order move from efficiency to growth, from a cheaper version of today’s business to a different one. Every prior reorganization in the two-thousand-year history above redistributed coordination among humans. This is the first that removes the need for the layer, and so the first that can change the shape of the firm rather than merely its efficiency.

How to move forward

None of this argues for racing to the fourth archetype. It argues for knowing which one you are running, and moving deliberately. In practice that comes down to three disciplines.

Diagnose honestly. Locate your organization among the four archetypes without flattering it. The common error is to mistake AI-Augmented for AI-Native, to see individuals moving faster and conclude the operating model has changed, when the coordination layer is fully intact and simply better fed. If a routine cross-functional decision still takes days and a chain of meetings, you are on the copilot plateau, whatever your tooling budget says.

Decide where the economics actually point. Not every business should sprint to Outcome-Native. The right target depends on how much of your cost and your competitive position is bound up in coordination. A business whose value is mostly coordination will be reshaped by this. One whose value is mostly physical or relational, far less so. The disciplined move is to pick the archetype your economics justify, not the one the market is loudest about.

Do not buy your way ahead. The failure mode of the sidelines is inaction. The failure mode of breaking the glass is spending your way into an archetype you have not built the foundations for. Redesign is the work, meaning decision rights, process and the shape of the teams, and the technology is what makes the redesign pay. That was true of computers in the 1990s. It is true of AI now.

The performance gap between the organizations that redesign and the ones that merely equip is going to widen, the way it did in every prior technology wave, and it will widen quietly, because faster individuals feel like progress right up until a competitor who rebuilt the operating model is doing something you structurally cannot match. The question is no longer whether AI will change the operating model. It is whether you will redesign yours deliberately, or leave it to be defined by whoever moves first.

Sources

R. H. Coase, “The Nature of the Firm,” Economica (1937).
F. A. Hayek, “The Use of Knowledge in Society,” American Economic Review (1945).
P. Bolton and M. Dewatripont, “The Firm as a Communication Network,” Quarterly Journal of Economics (1994).
E. Brynjolfsson and L. Hitt, “Beyond Computation: Information Technology, Organizational Transformation and Business Performance,” Journal of Economic Perspectives (1998).
L. Garicano, “Hierarchies and the Organization of Knowledge in Production,” Journal of Political Economy (2000).
K. McMillan, G. Kesler and A. Kates (eds.), Rethinking Operating Models (Kogan Page, 2025).
Microsoft & LinkedIn, “Will AI Fix Work?” 2023 Work Trend Index Annual Report (May 2023).
Microsoft, “Breaking Down the Infinite Workday,” Work Trend Index Special Report (June 2025). [275 interruptions/day; 57% of the workday spent communicating; “Work Chart” language]
Organizational history of span of control and the flat-organization experiments draws on Block/Sequoia, “From Hierarchy to Intelligence” (March 2026).