From boardroom hype to the cost model
In late 2023, boardrooms filled with words like exponential, transformational and AI everywhere. Strategy decks glowed with generative demos. Pilot programmes multiplied. By 2026 the tone has changed. The slide that matters most is the cost model, not the product roadmap. The question has moved from “What could AI do?” to “What did it actually deliver, and what did it cost?”
The shift is not a retreat. It is a reckoning, and finance is leading it.
The evidence that ended the belief phase
Every emerging technology passes through a phase of belief before it reaches a phase of proof. AI's belief phase peaked between 2023 and 2024. Adoption surged: Stanford's 2025 AI Index reported that 78% of organisations were using AI in 2024, up from 55% a year earlier. Spending followed, but most initiatives were still experiments.
Then came the data.
In mid-2025, MIT's NANDA initiative published The GenAI Divide, based on a review of more than 300 public deployments, interviews with 52 organisations and a survey of 153 senior leaders. It reported that 95% of the organisations studied had seen no measurable profit-and-loss impact from their generative AI pilots. The sample was small and the findings preliminary, but the headline captured the mood. In the same year, S&P Global Market Intelligence found that 42% of companies had abandoned most of their AI initiatives, up from 17% in 2024, and that the average organisation scrapped 46% of its proofs of concept before they reached production.
These numbers are sobering, but they are not evidence that AI failed. They are evidence that optimism outpaced operating discipline.
In the early phase, AI was treated as innovation theatre. Pilots launched across functions without ownership. Tools were bought before workflows were redesigned. Adoption was reported without reference to profit, cost displacement or cycle time. Success was inferred from usage rather than outcomes.
By 2026 that posture is untenable, and boards have noticed. EY's review of 2025 Fortune 100 filings found that 48% now cite AI risk as part of the board's risk oversight, three times the 16% of a year earlier.
Why the CFO now owns the AI conversation
The maturation of enterprise AI is less about technology than about economics. And unlike most enterprise software, AI's economics are variable.
A traditional licence is a fixed cost. AI inference is a marginal one: every query, summary and agent step consumes compute billed by the token or the call. Usage that grows faster than forecast becomes a budget overrun, and a model that is cheap in a pilot can be expensive at enterprise volume. That is why finance teams, rather than innovation teams, are setting the terms in 2026.
Three disciplines define the shift.
- Unit economics before scale. Every use case is costed per transaction: the inference cost, the human review that remains, and the cost it displaces. If the cost per resolved case is not lower than the process it replaces, scaling it only scales the loss.
- Cost ceilings and driver-based forecasts. AI budgets are tied to volume drivers, with explicit ceilings, alerts and variance reviews, in the same way mature organisations manage cloud spend.
- Fewer use cases, embedded deeper. Instead of scattering pilots across marketing, HR, finance and service, organisations concentrate on a small number of high-value processes and redesign them end to end. The difference is integration, not technical sophistication.
A worked example
Consider a hypothetical customer service assistant that drafts replies for human agents. In the pilot, a few hundred queries a week cost almost nothing to run, and agents report saving time. The business case looks obvious.
At full volume the picture changes. The assistant now handles tens of thousands of queries a week, each with a longer context as the knowledge base grows. Every reply still needs a human check, so the saving per case is a few minutes rather than the whole case. Integration, monitoring and model updates add a fixed cost that the pilot never carried. Unless someone has modelled the cost per resolved case at that volume, and compared it with the cost of the process it replaces, nobody can say whether the assistant is creating value or simply moving cost from payroll to the cloud bill.
The numbers in each organisation will differ. The discipline is the same: cost it per transaction, at real volume, before scaling it.
Governance belongs in the same frame. You cannot control the cost of a system nobody owns, and you cannot forecast spend on agents nobody has registered. Ownership and cost control are two sides of the same discipline.
The winners in 2026 treat AI as infrastructure. They assign named owners. They tie every model to KPIs that move revenue, cost, risk or resilience, an approach set out in more detail in From Acceleration to Accountability. They accept trade-offs between accuracy, speed and expense. They understand that scale without cost visibility is exposure, not advantage.
The rest optimise for breadth. They track adoption rates instead of value, and describe AI strategy in abstract terms (innovation, transformation, alignment) without anchoring it in unit economics.
Five questions finance should ask of every AI proposal
Realism is easiest to apply as a short set of questions that every proposal has to answer before it is funded.
- What manual step disappears if this works, and what does that step cost today?
- What is the fully loaded cost per transaction at expected volume, including inference, integration, monitoring and the human review that remains?
- How does the cost behave if usage doubles, and what ceiling triggers a review?
- Who owns the outcome, and who can switch the system off?
- What evidence will show, within two quarters, that it is working, and what happens if it is not?
None of these questions slows a good proposal. They stop weak ones from consuming the budget that strong ones need.
Realism feels less exciting, and that is the point
For leaders, the move from optimism to realism is also psychological.
Optimism is energising. It creates permission to experiment and justifies bold hiring and ambitious roadmaps. It can also delay accountability: when belief is high, execution flaws are forgiven.
Realism is less glamorous. It asks harder questions, forces trade-offs and requires shutting down pilots that people are emotionally invested in. Often it requires the CFO to say no.
Yet realism restores credibility. When finance leaders insist that AI proposals include full cost models, covering data preparation, change management, infrastructure and ongoing operations, they are not slowing innovation. They are protecting it from collapse. When boards demand clear ROI timelines and risk oversight, they are not resisting AI. They are making sure enthusiasm does not become liability.
The question inside teams has changed too. In the optimism phase it was “Can we use AI here?” In 2026 it is sharper: “What manual step disappears if this works?” If nothing disappears, nothing changes. And if nothing changes, nothing is gained.
You can hear the difference in the room. The language is plainer. The ambition remains, but it is channelled.
Ambition, now with a budget
The end of AI optimism does not mean the end of AI ambition. It marks the beginning of institutional competence.
The organisations entering 2026 with a durable advantage share common traits. They have narrowed their focus, embedded AI in core workflows, assigned ownership, put cost controls in place, and tied their metrics to real outcomes: revenue growth, risk reduction, cycle time and resilience.
They have replaced belief with discipline.
The organisations reporting clear returns from AI are not luckier. They are more rigorous. They measure, reinvest the capacity they free up, and close the feedback loop. The rest are still moving from narrative value to measured value.
In 2026, what separates the winners from the write-offs is not whether they adopted AI early. It is whether they matured quickly.
Optimism was necessary to begin.
Realism is what makes it worth continuing.
