AI in UK HealthTech

Why this briefing

In February 2026 an NHS England-sponsored evaluation at Great Ormond Street Hospital reported that an AI ambient-voice scribe raised direct care time by 23.5 per cent and cut documentation time by 51.7 per cent, across nine sites and 192 clinicians. It is a serious piece of work, and its findings are strong. It also states plainly what it did not measure: transcription errors, hallucinations, omissions, clinical inaccuracy — none of it, anywhere in the report.

Four months later the MHRA published the results of testing the same product in its own regulatory sandbox. Without active guardrails, out-of-scope behaviour — the system moving from documentation into clinical decision support — was detected in approximately 39 per cent of real-world notes, falling to 20 per cent with guardrails active. The same report warns that "users may reduce scrutiny of outputs over time in response to consistent and reliable product performance." NHS England's £10bn technology commitment, announced weeks later, is built on the first measurement.

Nothing currently connects the two. When the MHRA confirmed in July 2026 that most of these products are not medical devices, it placed the most widely deployed clinical AI in Britain outside the Yellow Card scheme — the national route by which a device fault is reported and counted. A patient safety charity states it plainly: there is currently no reliable mechanism to detect, report or count an AI ambient-voice incident at all.

The human moment

A clinician reads an AI-generated note before signing it. Early on they read closely, because the system is new. Months in, the system is right almost every time — the MHRA's own sandbox recorded a real-world precision score of 0.989 — and the regulator's own warning is that scrutiny falls exactly because performance is reliable. One trust that measured this directly found 44.4 per cent of its clinicians had, at some point, encountered the system inventing information. The safety case for every one of these products is that a human reads the output. Nobody is currently measuring whether that reading is still happening at month six.

What's inside

Seventeen pages, written for an executive who knows health well and AI not at all:

• A note on the numbers — what is included, what was excluded, and why
• The sector this lands on, and the largest technology commitment the NHS has made
• The pressure point — two arms of the state, the same product, two different pictures
• What the sector is facing, from four separate evaluations to the national imaging rollout
• Five risks, each named by a regulator, a trade body or the statute book
• The threat side — the reporting gap, and the clinical AI patients reach for that isn't the NHS's
• The UK legal position, and why section 80 did not liberalise health data
• Beyond the UK, and what the EMA and FDA's joint principles do and do not cover
• Seven things that can be done now, none requiring the National Commission to report first
• Three levels, defined by where accountability sits
• Benefit realisation, and the four measures a business case is currently missing
• An eight-workstream way in

The discipline behind it

Every figure carries a source. There is no market-size figure for AI in UK health technology anywhere in this briefing, because every one located traced back to a paywalled commercial research house with no published method. A regulator's own commentary on AI adoption was excluded despite ranking highly in search, because it dates from June 2022 and reads as current. A trade-press figure of 20,000 clinicians in a London rollout was superseded by NHS England's own published numbers and dropped.

One disagreement is reported rather than resolved. Two national surveys of GP AI use disagree — the Nuffield Trust and Royal College of General Practitioners found 28 per cent of GPs using AI clinically, a separate commercial-panel survey found 40 per cent. They used different frames and different recruitment, and the Nuffield Trust study has the stronger sample. The briefing leads on 28 per cent and names the disagreement, rather than choosing the larger number because it makes a better headline.

Who it's for

An NHS trust digital or clinical safety lead, a board member being asked to approve a wider ambient-voice rollout, or anyone who wants a plain, sourced account of what an NHS trial and a regulatory sandbox each actually found about the same product. It is written to work for all three.