The NHS has what legal services does not: an independent evaluation of an AI tool, with the method published before the answer

Ambient voice technology, often called an AI scribe, listens to a consultation and drafts the clinical note for the clinician to check. It is spreading quickly through the NHS. NHS England announced a national rollout on 4 July.

This desk set out to establish whether anyone is measuring what it does, independently of the companies selling it.

Somebody is.

Who is doing the measuring?

The National Institute for Health and Care Research funds a Rapid Service Evaluation Team, run from University College London and Cambridge with the Nuffield Trust. Its ambient voice study carries the project reference NIHR156380.

Phase 1 reported on 4 February 2026. Phase 2 collects data from August 2026 to January 2027 across four NHS trusts: two acute trusts using the tools in outpatients and accident and emergency, and two mental health trusts in outpatients. Results are expected in February 2027.

Phase 2 has three parts. A quantitative analysis of routine NHS records, including electronic health records and the Emergency Care Data Set, comparing before and after the tools arrive and, where possible, consultations that used them against consultations that did not. A health economic analysis, which will produce a return-on-investment tool any NHS organisation can run on its own numbers. And up to 36 interviews with staff.

What does it set out to measure?

The productivity strand is written as two questions rather than one. Does the technology produce measurable savings in documentation time, and how are those savings used in practice, for example more time with patients or absorbed into existing workload.

That second question is the one this paper has failed to find anyone else asking.

What did phase one find?

The team screened more than 500 records and found 21 studies that met its criteria. Most were short-term pilots or improvement projects, often under six months. Evidence was most consistent on reductions in documentation time, including work done after hours. Evidence on patient experience, safety and cost was limited and used inconsistent measures. Findings on the accuracy and quality of the notes were mixed.

The team also states plainly that much of the existing evidence comes from small, short-term pilots and is sometimes generated or commissioned by the technology vendors.

It then mapped the market. It identified 32 products, 14 with some claimed use in the NHS or social care. Of those 14, 11 were registered with the medicines regulator and 11 appeared on NHS England's supplier registry. The map records that claimed product use was often difficult to verify independently.

That is the same wall this desk hit in legal services, reached by a different route and written down by people with no product to sell.

The study that produced the headline numbers is a different thing

Much of the coverage of AI scribes in Britain rests on an evaluation led by Great Ormond Street Hospital across nine London sites and more than 17,000 patient encounters. It reported a 23.5 per cent increase in direct patient interaction time, appointments 8.2 per cent shorter, a 35 per cent fall in clinicians feeling overwhelmed by notetaking, and at St George's accident and emergency, 13.4 per cent more patients seen.

Those are NHS findings, and they are not independent of the supplier. The company whose product was tested was the technology partner, and its chief executive appears among the authors. The design compared the same clinicians before and after, with no control group. This desk has asked the company what its role was and whether the results have been peer reviewed.

Why is the prize worth having?

The evaluation records that clinical documentation can take up to a quarter of a clinician's working day, and that clinicians spend an average of two hours outside clinical hours on it.

Something that gives a fraction of that back matters. What nobody can tell you yet, in this country, is who ends up with it.