Value shows up where it is measured

The adoption numbers and the value numbers no longer belong to the same story. 88% of organisations use AI in at least one function; 5.5% report real financial returns from it. MIT's Project NANDA reached a starker version of the same finding by a different route: in its preliminary 2025 study of enterprise generative-AI deployments, only around 5% of pilots reached measurable profit-and-loss impact.

The common reading is that AI does not work. The evidence supports a narrower and more fixable conclusion: almost nobody measures. Programmes deploy without a baseline, run without instrumentation, and then face a board asking what the spend returned, holding anecdotes.

Why the measurement is missing

Nothing in the default adoption path requires it. The tools arrive with usage dashboards, not value dashboards; they report seats active and tokens burned, never hours returned or error rates moved. Pilots are judged as demos (does it work?) rather than as operations (what did it change?). By the time value is questioned, the "before" state was never recorded, and the honest answer is that nobody can know.

That failure is structural, so the fix has to be structural too. Bolting a survey onto a finished pilot produces satisfaction scores, not returns.

Four steps, run inside the work

The method is not exotic. It is the same discipline any operational change programme uses, applied from the first day:

1. Baseline. Before building, measure how the work runs today: cycle time, cost per unit, error rate, volume. If the baseline is not measured before the build starts, the value conversation is already lost.
2. Instrument. The delivery itself records throughput and cost per task as it runs. Instrumentation added later measures a different process than the one that was changed.
3. Attribute. Credit the change conservatively, in ranges, on the organisation's own data. Discount time savings that do not convert to capacity or output; a saved hour only counts when it lands somewhere.
4. Report. A recurring value statement against the baseline, in the numbers the board already uses. Once, at handoff, is a report; recurring is a system.

My conflict of interest

I sell AI delivery, and I have a direct financial incentive to tell you your programme needs measuring and that measurement should be mine. You should know that, and it is why the method above is designed to be held against me: we baseline before we build, the instrumentation runs inside the delivery where you can see it, and the report states the return in your numbers, conservatively. If the number does not move, that is visible too.

The market signal that this discipline pays: outcome pricing is moving from the margins of professional services to its centre. McKinsey's UK managing partner reports that a quarter of the firm's fees are now driven by outcome-based terms. Pricing against results is only possible for work whose results are measured, which makes the measurement method the commercial foundation, not the reporting afterthought.

The standard to hold any provider to, me included: baselines before builds, instrumentation inside delivery, attribution in conservative ranges, a recurring report in your numbers. Providers confident in their delivery will accept it. The others will offer a dashboard.

---

Sources

• McKinsey & Company, The State of AI in 2025: Agents, innovation, and transformation, November 2025 (n=1,993): 88% of organisations use AI in at least one function; 5.5% report real financial returns.
• MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025 (preliminary industry study, not peer-reviewed): ~5% of enterprise GenAI pilots reach measurable P&L impact.
• Michael Birshan, McKinsey managing partner UK/Ireland/Israel, reported by Business Insider, November 2025: ~25% of McKinsey fees on outcome-based terms.