AI Operating Model

In March 2025 McKinsey published a finding that should give executive teams pause. Of 25 organizational attributes it tested, the one with the biggest effect on whether companies saw any bottom-line impact from generative AI was the redesign of workflows. Yet only 21 percent of respondents whose organizations used generative AI said they had fundamentally redesigned even some of their workflows¹. The same survey found that a chief executive’s oversight of AI governance was among the attributes most closely linked to bottom-line impact, and that only 28 percent of respondents said their CEO held that responsibility¹. The two figures have different bases: the 21 percent counts respondents at organizations using generative AI, the 28 percent respondents at organizations using AI of any kind. Both are self-reports from a survey, not audited results.

That is a contradiction worth sitting with. The practices that separate results from activity are not secret. They appear in survey after survey. Most organizations in this one still had not adopted them, and the likelier reason is not ignorance. It is that nobody has been given the job.

Picture the Monday morning after your board approves an AI strategy. Someone has to decide which initiatives go first, who pays for each one, who builds the shared tools, who confirms that the risks are controlled, and who answers for the business result. If those questions have no named answers, the strategy stays a presentation, however good it is. The answers, written down and used every week, are the AI operating model.

Strategy sets direction; the operating model makes it repeatable

An AI operating model is the system of people, decision rights, funding, processes, technology and governance that lets an enterprise execute its AI strategy again and again, at scale. It answers the Monday morning questions once, so that each new use case does not have to negotiate them from scratch.

What Is an Enterprise AI Strategy? used Richard Rumelt’s kernel to show that a strategy is incomplete without coherent action². The operating model is where coherent action stops being a slide and becomes a routine. Marco Iansiti and Karim Lakhani draw the same line from another angle: a firm’s business model describes how it creates and captures value, while its operating model is the system that actually delivers that value to customers, and in their account AI’s deepest effect is on the operating model³. A strategy can be written in a quarter. An operating model has to be run every week.

An organization chart is not an operating model

Ask a leadership team how it will run AI, and the answer often arrives as a new box: a Chief AI Officer, an AI center of excellence, a dotted line from each business unit. The box may be necessary. It is not the answer.

This is an old lesson. In 1980 three McKinsey consultants, Robert Waterman, Tom Peters and Julien Phillips, published an article whose title states the case: “Structure is not organization.” Structure, they argued, is only one of seven interdependent factors, alongside strategy, systems, style, staff, skills and shared values, and reorganizing the boxes rarely fixes a problem that lives in the others⁴. Their framework became the McKinsey 7-S. Four decades later, AI programs repeat the mistake it was written to prevent.

The practical consequence is a rule about order. Start from the strategy, design the capabilities and decision rights it requires, and only then decide where they should live on the chart. Reverse that order and you get restructuring for its own sake. Centralized vs Federated AI, earlier in this module, compared central, federated and hybrid structures; that choice comes after the rights are clear, not before.

A picture from computing helps here. Think of a new AI team as an application you install. The icon appears on the screen, and the application may even work well. But whether work runs smoothly depends on the operating system underneath, which you rarely see: it decides which tasks are admitted, which process may use which resources, how those resources are shared and paid for, and what happens to a task after it starts.

The test of an operating model is therefore not whether the boxes exist. It is whether a real use case can move from idea to owned outcome without a fresh negotiation at every step.

The parts only work when they fit

An AI operating model has eight components that sit around the strategy: organization, decision rights, funding, talent, technology and data, governance, processes and measurement. The point is not any one of them but that they work only together.

Michael Porter made the general argument in “What Is Strategy?”: advantage comes less from any single activity than from the fit among many, because a system of activities that reinforce one another is far harder to copy than any one of them⁵. The same logic explains why operating models fail at the joints. Strategy without funding is a poster. Funding without talent cannot execute. A platform without processes produces tools that nobody builds into daily work. Governance without measurement cannot tell whether its controls are protecting the business or only slowing it down.

Measurement deserves a word because it closes the loop. A useful operating model measures a chain, from AI activity to system performance, to workflow change, to business outcome, and feeds what it learns back into strategy and investment. AI ROI and Value Realization, in Module 03, covers how to measure that chain. For the operating model the rule is simpler: a platform that reports usage is not yet reporting value, and someone must own the difference.

Business owns the outcome; every layer owns its part

Many AI stalls trace back to one confusion. Because a technology team builds the system, people assume that team also owns the result. It cannot. The team that owns the business outcome does not hand over that accountability just because another team owns the technology.

Think of it in four layers. The executive sponsor owns the investment decision and removes obstacles across functions; John Kotter’s research on change found that no transformation gets far without a guiding coalition with enough power to lead it⁶. The business owner owns the outcome: faster port turnaround, fewer defects, better margins, whatever the use case promised. The AI platform team owns the reliability of the service it agreed to provide. Risk and assurance owns the confirmation that controls actually work. Each layer is accountable for its own part, and none can silently take on another’s.

The split is what makes a stalled initiative diagnosable. When something stops, you can ask which layer has not done its part, rather than letting every group point at another. Writing these rights down precisely, decision by decision, is the work of the next chapter, AI Decision Rights and Accountability. How the intensity of control should rise with the risk of each use case, through risk tiers and approval gates, is taught in Module 07, AI Governance.

Owners who outlast the launch

Organizations have long run technology as projects. A project is funded, builds something, delivers it and closes; the team disbands and moves on. That works for a one-off system migration. It does not work for AI, because behavior, data, users and requirements keep changing after launch. A system that was accurate in March can drift by September, and there is nobody left to notice.

Mik Kersten’s Project to Product makes the argument for software generally: funding work as temporary projects rather than long-lived products with stable owners cuts the link between the work and the business outcome it is meant to serve⁷. For AI the case is stronger still.

So durable AI work takes three other forms. A product, such as a demand-forecasting assistant, has an owner who keeps measuring and improving it for as long as it runs. A platform, the shared foundation described in AI Platform Strategy, has an owner who operates it for every team. A shared capability, such as evaluation, serves everyone and needs its own owner too. The question to ask of any AI initiative is who owns it the day after launch, and the year after. If the honest answer is “the project team, until it disbands”, you have a temporary program, not an operating model. The governance of that life cycle, from reassessment after material change to controlled retirement, belongs to Module 07.

Funding rules shape behavior more than budgets do

An operating model needs a funding mechanism, not only a budget number, because the way money flows teaches people how to behave.

At one extreme the central team pays for everything. Business units treat AI as free, ask for more than can be delivered, and a queue forms in front of the central team. At the other extreme each business unit pays for everything. Units build their own tools, duplication grows, and nobody funds the shared platform or the evaluation service that everyone depends on.

The balanced mechanism mirrors ownership. Central funding pays for the enterprise platform, common capabilities and governance. Business funding pays for domain products, workflow redesign and the outcomes those units own. A small shared pool covers the few initiatives that genuinely cut across the enterprise. Because each unit pays for what it owns, it has reason to adopt; because shared assets have their own budget, they survive the next cost review. The same logic places people and data. AI Talent and Capability Strategy and Data Strategy for AI explain which capabilities sit centrally and which stay close to the domains. How to compare the cost of each option belongs to Module 08, AI Economics.

Right-size the model, then let it evolve

Operating-model complexity should match the size and ambition of the organization. Copying a global blueprint is a common and expensive mistake.

A global group may need an executive AI council, a central platform, central governance and specialists, plus product teams in the businesses that own workflows and outcomes; the center provides leverage, and the business owns the value. A mid-sized company may need none of those as separate units. A small central team, existing engineering, named domain owners and a few partners can be enough. A heavily regulated enterprise may keep risk, security and platform standards central while the business still owns each use case and its outcome.

Operating models also change over time. Many organizations move from ad hoc experiments, to coordinated standards, to a shared platform, to persistent products, and finally to AI built into how the enterprise runs. This is not a mandatory sequence; organizations enter at different points, and a higher ambition, as Defining AI Ambition showed, demands more of the model. What matters is that the model is designed for the ambition actually chosen, and redesigned when that ambition changes.

Story: a strategy with no implementation arrangements

In February 2025 Australia’s Auditor-General tabled a performance audit of how the Australian Taxation Office governed its use of artificial intelligence. It reads like a post-mortem of a missing operating model.

The ATO was not short of strategy or capability. It had adopted an automation and AI strategy in October 2022, and by May 2024 it had 43 AI models it had built itself running in production⁸. What it had not built was the system around them. The auditors found that the ATO had “not established fit-for-purpose implementation arrangements” for its strategy “or defined enterprise-wide roles and responsibilities.” For the 14 models built and deployed between July 2023 and May 2024, they found no evidence of ongoing performance monitoring and reporting. Nearly three quarters of the models in production lacked a completed data ethics assessment, and the office did not have sufficient central visibility of where AI was being used. The overall verdict: “partly effective” arrangements⁸.

Read the findings through the lens of this chapter and each one is an operating-model component that was missing. Roles and responsibilities had not been defined across the enterprise. Models went live without anyone visibly owning their performance afterwards, which is the project-to-product gap. The center could not see the whole portfolio, which is a measurement gap. Assurance had not kept pace with deployment.

The repair also followed the operating model rather than the org chart. During the audit the ATO set up a data and analytics governance committee in September 2024 and, in November 2024, named its Chief Data Officer as the accountable official for AI. It agreed to all seven recommendations, which included aligning its automation and AI strategy with enterprise requirements, defining organizational structures, governance and accountabilities, and setting up performance measurement for the strategy⁸. None of this was a new technology, and none of it questioned the talent of the teams that had built 43 models. The audit’s lesson is the one that applies to any enterprise: the strategy said where to go, and the people could build, but no one had been given the job of making execution repeatable.

What this means for leaders

The operating model is the executive team’s own work product, not something it can delegate to an AI team. The team can design a platform; it cannot assign the business outcome to a business leader, rewrite the funding rules or decide which layer owns assurance. Those are leadership decisions, and they are the ones most often left blank. The practical test is the one the auditors applied: pick real initiatives and check whether every part of their execution has a named owner.

Check yourself

1. An AI operating model is mainly a question of where the AI team sits on the organization chart.MythFact
2. When a central AI team builds a system, it should also own the business result.MythFact
3. Funding every AI initiative centrally tends to create a queue in front of the central team.MythFact
4. An AI system that is still running needs an owner after the project team disbands.MythFact
5. The ATO audit found that the tax office lacked an AI strategy.MythFact
6. A mid-sized company should copy the operating model of a global AI leader.MythFact

Reflection: the operating system underneath

What comes next

An operating model has eight components, and one of them stalls AI work especially often. When AI decisions cross business, technology, data, security and risk, uncertainty about who decides often becomes the bottleneck. The next chapter, AI Decision Rights and Accountability, sets out who should decide, who should answer for the result, and how to avoid both uncontrolled AI and decision paralysis.