Operating models
for the AI-native
enterprise.

Alfred Sloan took over General Motors and broke it into divisions. Each division ran its own product line with its own profit and loss, while the corporate center kept hold of capital allocation, financial reporting, and strategic direction. Sloan called it the multi-divisional form, and his 1963 memoir My Years with General Motors became the operating manual for the modern corporation.

Every consulting deck on operating models since has been a variation on Sloan. Box names change, reporting lines shift between functional and product-led structures, and centers of excellence appear and dissolve, but the underlying premise stays the same: a corporation is a system for separating strategic decisions from operational ones, holding the center accountable for capital allocation, and pushing execution down toward the people closest to the work.

That premise survived the personal computer, the internet, ERP, agile, the cloud, and the entire 2010s wave of platform thinking.

Whether it survives large language models is yet to be seen.

Pick up any 2025 strategy deck on the AI-native enterprise and you’ll find roughly the same diagram, usually a pyramid or a target with three or four concentric rings. At the core sits data and infrastructure, with an AI platform layer above it, products and workflows above that, and an outer ring covering people, processes, and governance.

The recommendations follow a predictable pattern: there’s always a central AI function, a Chief AI Officer at the top, and a center of excellence underneath staffed with model engineers and prompt specialists. Use cases get federated to business units while platform standards stay centralized. Data infrastructure comes first, because models are commodities and proprietary data is the moat. Training gets rolled out, adoption gets measured, and the whole thing iterates on a quarterly cadence.

These decks are competent and largely interchangeable, all answering the same underlying question: how should a Sloan-era corporation reorganize itself to absorb a new general-purpose technology?

This is the wrong question.

Sloan’s M-form was a response to a specific constraint. In 1923, the cost of producing competent industrial output was high, and the cost of coordinating production across multiple product lines was higher. The divisional structure kept coordination costs bounded while letting each product line scale independently.

Knowledge work inherited the structure but inverted the costs. By the time Peter Drucker described the knowledge worker in The Age of Discontinuity in 1969, the cost of competent output had become the binding constraint. A piece of analysis, a report, a strategy document, a piece of working code: each one took a person or a small team weeks of effort. Coordination was expensive too, but production was where the budget went, and org charts followed the same logic. Most corporate functions are still structured around the assumption that producing competent knowledge work is the bottleneck.

Large language models broke that assumption. The cost of producing a competent first draft of almost any knowledge artifact has collapsed by something like two orders of magnitude. Code, copy, slide decks, financial models, legal review memos, market analysis, customer research summaries: a model can produce a credible first pass on any of these for cents per attempt.

The bottleneck moved to judgment about what’s worth producing, taste about how to refine it, distribution to people who can act on the result, and accountability when something breaks. Operating models haven’t caught up. Most AI-native frameworks are still optimizing the production layer.

If you take seriously that the binding constraint is now judgment rather than production, several conventional moves in the operating model deck become incoherent.

The COE format was invented for technologies that required deep specialist expertise to deploy — enterprise data warehouses, statistical modeling, and six sigma. The specialist sat at the center, business units brought problems in, the specialist applied the technique. With AI, the specialist input has become broadly available; anyone can prompt a model. The expertise that matters is judgment about the problem, which lives in the business unit.

The Chief AI Officer role assumes that AI is a discrete capability that can be governed, deployed, and measured the way IT governs cloud spend. That works for things with clear boundaries; AI as a general-purpose technology behaves more like electricity in 1910 than like a software vendor relationship in 1995. There was no Chief Electricity Officer at U.S. Steel, because electricity stopped being a capability and started being a substrate. AI is on the same trajectory, and probably faster.

The metric most decks suggest tracking is adoption; number of employees using AI tools, number of workflows automated, hours saved per role, and tools sanctioned by IT: these are cheap to measure and almost meaningless. The companies extracting actual economic value from AI have rewired specific workflows so that the model handles the production step while a human handles the judgment step, with the rewiring showing up in unit economics rather than seat counts on Copilot.

The metrics that matter sit downstream of judgment, rather than upstream of usage.

If the constraint is judgment, two genuinely different operating models emerge from the AI-native shift, and most consulting decks blur them together.

The first is the centaur model. Humans and AI collaborate on the same artifact, with the model handling production and the human handling judgment and final accountability. The org chart looks similar to today’s, but every individual contributor has roughly the output of a small team, while middle management’s coordination function shrinks. Companies that pull this off reorganize around feedback loops, which means faster review cycles, smaller artifact units, and tighter coupling between the person making the call and the person living with the consequences. Humans stay in the loop, but each one has the throughput of a small team.

The second is the agentic model. Functions get decomposed into pipelines where AI agents handle production and routine judgment, with humans intervening only at exception points and final approval. This is closer to what Marc Andreessen described in his 2011 Wall Street Journal piece on software eating the world, except the software now writes itself. It’s how Stripe handles fraud detection at scale, and how Klarna ran the customer service replacement that its CEO Sebastian Siemiatkowski claimed in 2024 was doing the work of 700 agents. The org chart shrinks dramatically, and the people who remain are doing exception handling and orchestration of the agent pipelines themselves.

These are different operating models with different economics, talent profiles, governance requirements, and failure modes. A consulting deck that treats them as points on a spectrum, rather than as architectural choices the company has to make explicitly, is selling confusion as nuance.

Sloan’s structure depended on middle managers. They translated strategic intent into operational execution, allocated work across teams, reviewed output, escalated exceptions, and absorbed political risk when things went wrong. They were the connective tissue.

Most middle management work is judgment work, which means most of it is exposed to AI in ways that production work isn’t. A senior manager’s role is hard to fully automate, since part of it exists to absorb accountability when something breaks, and accountability is hard to delegate to a model. Compression is the more achievable move. The team that used to need a director, two senior managers, and four ICs can probably function with a director, one senior manager, and three ICs who each have AI doing what the fourth IC and one of the senior managers used to do.

The companies actually pulling this off skip the AI-native reorg announcement and freeze hiring in the middle layer instead, letting attrition do the structural work, and reinvesting the savings in IC compensation for people who can do judgment work at higher leverage. The org chart that emerges is flatter and more barbell-shaped, with more senior individual contributors and fewer mid-level managers per unit of output.

This is the operating model change actually happening at many large companies in 2025 and 2026. It doesn’t fit on a McKinsey ring diagram. It looks like a hiring freeze.

The deck format itself is part of the problem. A consulting engagement has to produce something legible to a steering committee, with clear workstreams, named accountabilities, and a timeline measured in quarters. The actual operating model change in an AI-native shift, the one driven by judgment and feedback loops and middle management compression, doesn’t produce that kind of artifact. It shows up as small accumulating changes in how individual teams do their work, most of which are invisible at the level of an org chart.

Clayton Christensen made a related point in The Innovator’s Dilemma in 1997 about why incumbents keep losing to disruptors. The incumbents have measurement systems calibrated to their existing business. The disruption shows up first in metrics they don’t track, in segments they don’t serve, and in workflows they don’t measure. By the time the disruption becomes legible to the existing dashboards, it’s already too late to respond.

The AI-native operating model has the same property. It shows up first as small productivity gains in specific workflows, as one IC suddenly producing what used to take a team, as a manager noticing they can run a tighter unit with AI in the loop, and as a CFO realizing that headcount stayed flat while output climbed two quarters in a row. None of this is captured by adoption metrics or COE charters.

There’s a related diagnosis from Ronald Coase, whose 1937 paper The Nature of the Firm argued that firms exist because internal coordination is sometimes cheaper than market transactions. AI lowers the cost of internal coordination and the cost of market transactions at the same time, but unevenly. Some functions, like routine analysis and content production, become cheaper to do externally through models and contractors than internally through full-time staff. Other functions, like judgment under genuine uncertainty and accountability for outcomes, become more concentrated inside the firm because they can’t be easily contracted out. The boundary of the firm gets redrawn, but not in a single direction. Some companies will get smaller and more concentrated. Others will get larger and more diffuse. The deck framework can’t model this because the deck framework assumes the boundary is fixed.

Instead of asking how to reorganize for AI, the question worth asking is which feedback loops in the company are now too slow: where artifacts take a week to produce when they should take a day, and where roles exist primarily to coordinate work that the model can now coordinate without help.

Paul David made a related point about electrification in his 1990 paper The Dynamo and the Computer. Factories had electric power by 1900, but the productivity statistics didn’t move for another thirty years. The intermediate decades got spent retrofitting electric motors into floor plans designed around a single central steam engine, running new technology through the architecture of the old one. The factory layout was the operating model. The deck format didn’t exist yet.

The current decks are doing the same thing in reverse, drawing org charts around a production cost that already collapsed and calling the reorganization a future-state initiative. If David’s timeline is anything to go by, we’re somewhere around 1905.