How much is AI being used in your organisation, and what is its impact? It sounds like a simple question. Yet when a senior leader in a financial services company asked me recently how I’d address it, I realised I couldn't give a good answer. Neither, it turns out, can most organisations, most regulators, or most governments.

As I investigated this further, I saw that judging whether AI is delivering value requires us first to be able to see it. Yet much of what passes for AI measurement today counts activity and assumes impact. Licences bought, pilots launched and staff trained are important. But they’re not evidence of delivered results.

So, I set out to answer a narrower question. If we wanted to build an “AI monitor” for an organisation, a community, or a country, what would it measure, and what would we learn by doing it?

The most useful step I found was to stop asking “how much AI?” and break the question into a chain. Each link asks something different, and each is harder to measure than the one before.

The pattern this chain of questions exposes is striking. Most measurement stops at the first two links in the chain and implies the last three. For example, a dashboard showing that 80% of staff have a Copilot licence tells you about access. It says nothing about whether anything has improved.

When I ask organisations how they measure their AI adoption, they most often point to three sets of data: tokens consumed, user feedback, and vendor tool usage. These obvious measures are helpful, but each comes with a catch.

Tokens measure compute, not value. Counting the text an AI system processes is a reasonable signal of intensity inside one organisation. However, newer models and agents burn far more tokens for the same task, units differ between vendors, and once token use becomes a target, people will find ways to consume more of them. It is rather like judging a factory's output from looking at its electricity bill.

Surveys are good for who, weak for how much. Asking people what they use and what it does for them reveals adoption, attitudes and the unofficial use that no system logs. Yet people misremember, and their sense of benefit is unreliable. In a 2025 study by METR, experienced developers were measured as taking 19% longer to complete real tasks with AI tools, while still believing afterwards that AI had made them about 20% faster.

Embedded AI is visible only to vendors. Increasingly, AI arrives inside software people already use, from office tools to CRM systems. Its use shows up in vendors’ admin dashboards and licence counts, and almost nowhere else. As AI becomes infrastructure, “AI use” stops being a useful category. The better question becomes how tasks and outcomes are changing.

To test the framework, I mapped what the UK publishes against it. The good news is that adoption is now well tracked. The Office for National Statistics business survey shows 30% of UK businesses using at least one AI technology in September 2026, up from 9% three years earlier. It also asks what firms use AI for and whether it has changed their headcount: 6% of AI-using firms with ten or more staff report cutting jobs because of AI.

The ONS household survey adds the human picture. In June 2026, 63% of 16 to 29 year olds in Great Britain used AI for work or study, compared with 3% of over-70s. Confidence is also slipping: the share who think AI brings more risks than benefits rose from 25% in August 2024 to 38%.

Even so, a basic question exposes the limits. Ask how many people use AI and you get three credible answers. The ONS figure implies around 67% of adults knowingly used it in the past year. Ofcom puts it at 54%, and Microsoft’s telemetry at 44% of 15 to 64 year olds. None is wrong; each asks a different question.

The real gap is further down the chain. Of 26 indicators I tracked, 17 have some UK figure. Yet only two of the four outcome indicators have any figure at all, and neither is robust. The nearest thing to an outcome measure is the cross-government Copilot experiment, in which civil servants given 20,000 licences reported saving 26 minutes a day. That figure is self-reported, for one tool over three months.

It would be easy to read this as a British failing. It is not. I compared the UK with the United States, the European Union, Ireland, Canada, Singapore and Australia, and every one of them now measures business adoption. Their headline figures range from 12% of Australian businesses (ABS) to around 20% across the EU (CSO Ireland) and the 30% reported by ONS, although each asks a different question of a different set of firms.

Some countries see things the UK does not. Statistics Canada tracks whether firms train their staff in AI, something I couldn’t find in UK official statistics. The US Census business survey records whether AI now does tasks that employees used to do, a rare glimpse of automation rather than adoption (Minneapolis Fed summary). In turn, the UK is unusually strong on public attitudes and trust.

The most important finding is what is missing everywhere. None of the seven publishes an official measure of what AI changes. The UK and Singapore (IMDA) come closest, and both rely on what people say about themselves. Hence the question every government is asking, “Is AI improving productivity?”, is one that no national statistics system can yet answer.

To see what this looks like in practice, I built a prototype UK AI Monitor, which you can see at AIMonitor.uk. It maps every indicator onto the five areas, shades each one by whether the UK publishes a figure, and compares the UK with other countries. Its central message fits on one line: the UK can see who uses AI, but it cannot yet see what AI changes.

Building it taught me that a monitor should not pretend to count all AI use. That is impossible. Instead, it should track a consistent set of signals, be open about where they run out, and push for the few investments that would fill the biggest gaps.

For the UK those are clear. Link the businesses that say they use AI to their productivity and employment data. Require public bodies to report the results of AI they deploy, not just describe the tools. Build a trusted panel of organisations willing to share anonymised data on what their staff actually do with AI.

The same logic applies inside any organisation. Most leadership teams I speak to can tell me how many AI tool licences they hold. Far fewer can tell me what has changed as a result. If you want to know where your own monitor would go dark, try these four questions.

If the honest answer to the third question is no and you’ve never asked the fourth one, you’re in good company: so is every national government I looked at. However, that is not a reason to stop asking. Seeing AI clearly is the necessary first step to managing it well. Perhaps what I put together at AIMonitor.uk will help you see where to start.