This paper has spent a month saying nobody measures what AI saves. In the NHS, somebody is

Three times this month we have printed a version of the same finding. On 1 September, that every figure in circulation about AI saving time in British legal work came from a company selling the software. On 14 September, that eleven organisations asked who checks those numbers had not replied. This morning, that the professional body for solicitors holds no independent measure of its own.

Read back to back, that is a paper telling you nobody measures anything. So here is the other half, and Friday is the right day for it.

In the NHS, somebody is measuring, and rather well.

The National Institute for Health and Care Research is funding an independent evaluation of the AI note-taking tools now spreading through hospitals and GP surgeries. It is run from University College London and Cambridge. It has a published protocol. It ran a first phase that mapped what evidence already existed, and it says in writing that most of what it found was short pilots, some of it produced or paid for by the companies selling the tools.

The second phase collects data across four NHS trusts until January and reports in February 2027.

Now the part I want to point at, because it is the question this paper has been asking since June about every job we have covered.

The study does not only ask whether the machine saves time. It asks where the saved time goes. Those are its words, near enough: whether documentation time falls, and how any time released is used in practice, whether that is more time with patients or absorbed into existing workload.

Nobody else we have written about is asking that. Not the suppliers, who measure the saving and stop. Not the regulators we wrote to. Not the trade bodies.

I am not going to dress this up further than it deserves. The evaluation has not reported. The study that produced this summer's striking numbers was a different thing: NHS-led, but run with the supplier as technology partner, with the supplier's chief executive among its authors and no control group. Leah sets out both, in that order.

And the early signal points somewhere uncomfortable. Where time released has been measured so far, a good share of it went into seeing more patients per shift. That is a real gain for someone waiting on a trolley at two in the morning. It is not a clinician getting home earlier, and the two keep getting reported as though they are the same thing. Tomas has the case against all of this, which is what Fridays are for.

What I take from it is narrower than good news and more useful than a headline.

The evidence we keep asking for can be built. Somebody scoped it, funded it, published the method, named the outcomes in advance and put a date on the answer. It needed a research council, two universities and about two years.

Every sector this paper writes about could have that. Most of them have a supplier's press release and a trade body that has not replied to us.

— M.

This note is mine: the view, and the call to run it. It begins as a draft, drawn from work the AI and I have researched and argued out together, the same way every desk in this paper is made, and I answer for every line because I read every line. Those desks run on models built by Anthropic, one of the labs sitting on the very scoreboard we report, so we tell you plainly: we cover this from inside it.