AI is in the budget. It’s not yet in the results.

The tools are running. Everyone says it is going well. The productivity hasn’t moved.

You are not alone in noticing this. In early 2026, researchers from Stanford, the Bank of England, and the Federal Reserve published a survey of nearly 6,000 senior executives across the US, UK, Germany, and Australia. Nine in ten reported no measurable impact from AI on either employment or productivity over the previous three years. This was not a survey of sceptics or laggards. It was a broad cross-section of business leadership — most of them actively using AI tools, most of them expecting results.

The macro numbers tell the same story. US total factor productivity grew 0.8% in 2025, down from 1.5% the year before. Goldman Sachs’ chief economist, reviewing the 2025 data in January, concluded that AI contributed negligibly to US economic growth. Hundreds of billions in capital expenditure. No aggregate output to show for it yet.

This does not mean the technology is wrong. What it means is that the technology is not the lever.

The organisations seeing real returns — around 6% in McKinsey’s 2025 global survey — share a specific pattern: they have redesigned workflows end to end, not bolted AI onto existing processes; they measure against business outcomes rather than usage metrics; and they have assigned explicit human ownership to the AI system — someone whose job includes feeding it, auditing it, and deciding when the output is not good enough to act on. The first two factors show up clearly in the McKinsey data. The third is the one I would add from what I have seen in the programmes that do not.

The organisations not seeing returns have done the opposite. They purchased licences, encouraged adoption, counted usage hours, and declared the transformation underway. The tool is present. The operating model around it has not changed. That gap — between having AI and using it in a way that produces measurable output — is where the investment is being lost.

There is a separate finding worth sitting with. Anthropic published research in early 2026 that ran a controlled experiment with engineers learning new technical material — one group with AI assistance, one without. The AI-assisted group finished marginally faster, but not by a statistically significant margin. They scored 17 percentage points lower on comprehension tests. What this describes is not AI failing to produce output. It describes AI producing output while reducing the understanding that would otherwise develop from doing the work. The organisation gets the artefact. It forgoes some of the understanding that would have developed from building it.

This is distinct from the productivity question, but not unrelated. If the people doing your most important work are getting faster at producing things and slower at developing the judgement underneath them, the return on AI looks different across a three-year horizon than it does in this quarter’s usage report.

The question worth asking before the next renewal conversation is not whether your people are using AI. It is whether any of that activity is showing up in the numbers your business actually runs on.