If you run a company or a programme for senior leaders, you have probably heard both of these in the same week: “everyone is using AI” and “we cannot find the return.” Both are true, and the research we collected this month explains why.
Use is high. Measured impact is thin.
- PwC’s 2026 CEO survey (4,454 CEOs): 56% report no significant financial benefit from AI yet. Only 12% see gains in both cost and revenue. Self-reported, by an advisory firm that sells AI services.
- A study of about 6,000 executives in the US, UK, Germany and Australia, from the Atlanta Fed with the Bank of England, the Bundesbank and others: around 70% of firms use AI, yet more than 80% report no effect on employment or productivity over the last three years. They expect gains over the next three.
- UK government research on 700 AI-using businesses: 75% say workforce productivity improved, but 77% report no change in revenue. All self-reported.
- MIT NANDA’s much-quoted “95%” is looser than its headline. The report draws on 52 organisations interviewed and 153 leaders surveyed, calls itself preliminary, and says its figures are directionally accurate rather than audited. Its two headline numbers use different bases: 95% of organisations getting zero return, and 5% of integrated pilots extracting millions. Treat it as a signal, not a measurement.
Vietnam shows the same pattern. In an AWS-commissioned survey of 1,000 business leaders, 71% of adopters say AI at least broke even, yet only 13% have a consistent ROI framework and 61% are still experimenting. It is one data point among several, but it is the clearest one on the measurement gap itself.
Why the gap: we mostly measure feelings
Where time savings are measured at all, they are modest and mostly self-reported. HMRC’s randomised Copilot trial found about 2–3% of the working week saved after discounting for response bias. In a randomised trial, experienced developers were 19% slower with AI tools, though they believed they were faster.
Executives’ own language runs ahead of results. An analysis of about 490,000 earnings-call transcripts by the St. Louis Fed found that 95% of AI-related productivity sentences refer to future gains, compared with about three quarters of other productivity talk.
Time saved is also not money saved. A saved half hour only reaches the P&L if the work, the headcount or the price changes. Klarna and Commonwealth Bank both reversed AI-attributed staffing decisions, and Gartner expects half of firms that cut service headcount for AI to rehire by 2027.
What the few that do show returns have in common
- They redesign the work. McKinsey’s roughly 6% of AI high performers are nearly three times as likely to have redesigned individual workflows. Deloitte finds 84% of firms have not redesigned jobs around AI, and BCG puts about 70% of the effort in people, organisation and process.
- They train people. In a Danish study summarised by Brookings, employer encouragement, tools and training lifted chatbot use to 93% of workers, against about 41% without it. The study still found no earnings or hours effect yet, which is the same measurement point again.
- They compare against a baseline. DBS reports about S$1bn of AI value in 2025, derived by comparing customers offered AI-driven solutions with a control group. It is company-reported and secondary, but it is the right shape of claim.
What this means for leaders and programme directors
Three questions separate a pilot that will reach the P&L from one that will not:
- What is the baseline? If nobody wrote down today’s cost, cycle time or revenue for this workflow before the pilot, there is nothing to compare against.
- Who owns the number? A named budget owner, not a project team, who answers for the line that should move.
- What work changes? If the workflow, roles and incentives stay the same, the saved time evaporates.
For a programme, that points at the design: leaders should work on a real problem in their own business, with the baseline, the owner and the changed work agreed up front. That is why Civico’s programmes put senior leaders on a real problem rather than a case study.
Method note and limits
Every figure above is linked to a source we opened, and the rows are in the Use cases view. Most are self-reported surveys by firms with something to sell; where a figure is vendor-reported, secondary or from a small sample, the row says so. We have little evidence on founder-led firms specifically, and nothing yet on manufacturing, logistics or travel. The Knowledge graph shows how these findings connect to the rest of the library.