A point of view on what changes, what does not, and why most companies are getting it wrong.
AI strategy without an operating-model evolution produces activity, but not impact. Most large enterprises have an AI strategy. Very few have changed how the business runs. The two are not the same thing, and the second one is much harder. Without it, the strategy produces pilots, partnerships, and slides, but does not change the company.
The default move is to bolt AI onto the existing operating model. The incentive is understandable, and the pattern is already well-established: Hire a Chief AI Officer. Stand up an AI function. Pick relevant use cases. Establish meaningful partnerships. Approve a governance document.
Still, ticking the boxes is not what changes the operating model as long as the work flows through the same approval chains, the same budget cycles, the same governance committees, the same role boundaries that were designed for a pre-AI world. The result is friction, not transformation. After 18 months, the AI strategy has produced a portfolio of pilots that don’t scale. The conclusion drawn is usually that AI is harder than expected. The actual problem is that the operating model is not built to carry it.
This Point of View argues something simpler. An AI-enabled operating model is not an operating model with an AI function attached. It is a new type of operating model, where the unit of work, the rate of change, the location of decision rights, the build-buy boundary, and the role of the center are all different. Companies that do not redesign these dimensions will struggle to become AI-enabled.
AI is not a function
The most common pitfall is treating AI as a function and assuming that this is sufficient for AI integration and transformation. The pattern was borrowed from digital a decade ago, when digital had a coherent scope. Web, mobile, ecommerce, marketing technology. It worked because there was a defensible boundary around the function.
This does not transfer to AI. AI does not have a coherent scope. It is not a thing. When AI is set up as a function, it has to negotiate access to every other function, and this creates friction across marketing, operations, and engineering. The Chief AI Officer is challenged by the same credibility problem that affected the CIO a decade ago. In that setup, the work that a CAIO actually impacts is a fraction of the AI work happening in the company, while the work outside that scope is often where the real value is realized.
The right move is to distribute AI capability pervasively where the work is, with a thin spine at the center that owns a small number of things that have to be the same across the organization. The spine is small. The work is spread throughout the entire organization. If there is a Chief AI Officer, the role is an orchestrator and standard-setter, not the owner of the work.
This will be resisted, because the CAIO-led AI function is a comfortable structural answer to a hard organizational problem. It looks like progress on the org chart. In practice, it is mostly an organizational placeholder that inadvertently gives the rest of the company misdirection and results in no changes.
What actually changes
Five things change when AI becomes how the work gets done. None of them are about technology. All of them are about how the business runs.
The unit of work changes from project to slice
The traditional operating model is built around projects. Projects have scope, timeline, budget, and a defined endpoint. AI work does not fit this pattern. Scope reveals itself through doing. Timelines are compressed but iterative. Budgets are spent on learning before they are spent on building. The endpoint is the start of the next slice.
Operating models that still process AI work as projects result in long approval cycles, scope expansion, and governance that treats early-stage work as if it were late-stage. This results in a pipeline of pilots that never become production. The fix is not to make AI work fit projects. It is to make real slices of the unit of work.
A slice is bounded, embedded in a real workflow, sponsored by a P&L owner, with explicit decision gates. It is funded as a small bet, not as a project. It produces production capability or it stops.
The rate of change becomes a design parameter
Operating models historically assumed years of stability between redesigns. Design the model, operate it for five years, then redesign. This assumption collapses when the underlying premise changes every few months.
A model provider releases a substantially better capability. A new agent pattern makes a previously impractical workflow practical. A regulation changes what can be deployed where. None of these are exceptional events. They are routine. An operating model that needs governance that is able to balance progress and risk.
The answer is not faster governance. The answer is governance operating at two horizons. A slow horizon for things that do not change often, such as data residency, IP boundaries, ethics, or controls infrastructure. A fast horizon for things that do, like model selection, workflow design, or integration patterns. Most enterprise operating models combine this into one, which is why everything feels slow.
Decision rights shift toward the work
In a pre-AI operating model, decisions accumulate at the top. Expertise is scarce. Information takes time to reach decision-makers. Decisions take time to come back to the work.
AI changes the underlying logic. Expertise becomes augmented. Information moves faster. Decisions are made closer to the work.
This is not an org flattening argument. The org chart still matters. Accountability still matters. The change is in which decisions stay at the top – strategy, capital allocation, ethics boundaries, brand – and which move down – configuration, prioritization, day-to-day operational choices. Most enterprises have not moved any of them yet, which is one of the reasons their AI progress feels slower than it should.
The build-buy boundary moves
In a pre-AI operating model, the build-buy choice was binary. Build it (long timeline, high cost, custom fit) or buy it (short timeline, lower cost, standard fit). AI breaks the binary.
The new pattern is assemble-and-adapt. A foundation model is sourced from a provider and combined with an integration layer, workflow logic, evaluation capability, and deployment infrastructure. Adapted to the use case. Crucially, none of these are permanent. They are reassembled as components improve.
Most enterprise procurement functions are not configured for this. They process AI deals like they processed ERP deals a decade ago. Multi-year, single vendor, large committed spend, exhaustive specification. The right pattern is short cycles, multiple providers, outcome-aligned commercial structure, and optionality preserved at the model layer. Procurement has to be redesigned, not just trained.
The role of the center changes from doing to enabling
A pre-AI center owned things. IT. Data. Digital. Centers were doers. They produced capability and pushed it out to the business.
An AI-enabled center is different. It does not own most AI work. The work is distributed where the value is. The center owns a core set of elements that ensure consistency across the organization. Identity, standards, platforms, controls, model relationships, and cross-cutting capability libraries.
The center is a spine, not a body. It is small, senior, and mostly invisible from the outside, but able to get involved as needed. When it becomes the most visible part of the AI work, something has gone wrong.
Where the difficult decisions actually sit
Three structural decisions determine whether the operating model can carry AI work or breaks under it. None are technology decisions. Each is being made right now inside most large enterprises, and most of them are being made without the right inputs.
Who is accountable for outcomes
Most enterprises put AI under the CIO, the CTO, or a newly hired Chief AI Officer. This sounds like accountability, which it is only if those individuals also have P&L accountability and benefit from the AI work.
AI work that does not have a P&L owner does not have a real sponsor. It has a budget owner. The two are not the same, and the difference is what determines whether the work scales.
The CIO or CAIO are accountable for platform readiness, talent readiness and supply, and shared services. The business is accountable for outcomes. Where this assignment is unclear, the work stays in pilot mode, because no one is paying a real price for it failing to scale. The visible symptom is a pipeline that looks busy and a P&L that is not moving.
How AI work is funded
AI work is not an IT project, nor should it be funded like one.
The better funding instrument is portfolio capital with explicit decision gates. A pool of funding is allocated to AI work in a business domain, with the capability owner authorized to deploy it across multiple slices. Each slice has gates. Continue, refine, or stop. Stops are normal. The portfolio metric is value created across the pool, not the success of a single slice. This is closer to how venture capital allocates than to how IT funds projects, and that comparison is the right one.
The CFO has to be part of this redesign. Without the CFO, the existing budgeting machinery will reassert itself within a year, and the operating-model change will quietly reverse.
How governance is designed
Governance designed in advance of production experience governs imagined risks. Real production risks are different. They surface through delivery. Operating models that try to design comprehensive governance upfront fail to reflect the real-life risks, slow the work, and become a constraint.
Governance should evolve through delivery. The slow-horizon items, such as data, IP, ethics boundaries, and controls, are set early and rarely revisited. The fast-horizon items, workflow patterns, validation methods, monitoring approaches, and escalation paths, emerge as the work produces the real signal. Production experience teaches the organization what to govern.
The workforce shape that follows
A few observations on talent that follow from the operating-model decisions above. They are not separate from the operating model. They are a consequence of it.
The most valuable people in an AI-enabled organization are not the technical specialists. They are the operators who understand a domain deeply and can work fluently with AI capability inside it. Senior underwriters who can configure AI agents to triage their portfolio. Senior engineers who drive code generation as part of their workflow. Senior clinicians who can adapt AI suggestions inside their practice. The technical specialists matter, but they are not the binding constraint. The binding constraint is the supply of AI-fluent senior operators inside the business.
A second observation. The middle of the workforce is where the most disruption will land, and where most companies are least prepared. Senior people will be augmented and will retain their roles. Junior people will be hired with AI fluency built in. The middle, where the work has historically been about pattern execution and coordination, will compress. Operating models have to make this transition deliberate.
What to measure
A useful test, twelve to eighteen months into operating-model redesign, is whether the company can point to three things.
Production slices in real workflows, sponsored by P&L owners, with measurable economic impact.
Time-to-production for a new use case has dropped from quarters to weeks. The operating model is now producing AI work fast enough to matter, with quality high enough to scale.
The operating model itself has visibly changed. Funding instruments look different, and so do decision rights, governance, and the workforce shape. If none of these have moved, AI work has not yet touched the operating model. It has been absorbed by the existing one.
A closing thought
The companies that grow to become genuinely AI-enabled will not be the ones with the best AI strategy. They will be the ones who redesigned their operating model around AI work, instead of running AI work inside an operating model that was not built for it.
This redesign is happening now, and the gap between the companies doing it and the ones that are not is widening every quarter. Step one is understanding why the operating model is the constraint. Then the work can begin, and AI work can scale.
