104 NHS theatres, $1bn in AI billing overcharges

Two NHS organisations opened themselves up to an AI evaluation this month, one testing devices in a live regulatory sandbox, the other testing whether AI can shave dead time off a surgical list. Around the same time, a US insurer worked out that AI scribes and coding tools had added roughly $1 billion to hospital bills in a single year, not because patients got sicker, but because the tools got better at finding billable detail. Read together, the two stories tell you something useful about where an independent practice actually stands right now: the tools are moving from pilot to production faster than most owners have time to track, and the same tool that saves you an hour a day can just as easily inflate a bill nobody meant to inflate. Five developments to 27 September 2026, and what each one means for the decision in front of you.

1. MHRA opens a live NHS sandbox for AI devices, in Manchester

The MHRA and Manchester University NHS Foundation Trust announced a joint sandbox on 10 September 2026, giving AI-enabled medical devices, wearables and Ambient Voice Technology products a real NHS setting to be tested in before wider approval, rather than a lab bench or a vendor demo. The initial focus is tools that flag which patients with long-term conditions are at greater risk of complications, so a clinician can step in earlier. MHRA chief executive Lawrence Tallon and trust chief executive Mark Cubbon are both named as sponsors, and a roundtable with clinicians, regulators and technology developers is planned later this year before the MHRA opens the sandbox to further device developers. Source: MHRA and Manchester University NHS Foundation Trust, announced 10 September 2026, reported by Intelligent CISO, 21 September 2026.

The timing is not a coincidence. The National Commission into the Regulation of AI in Healthcare recommended on 10 September 2026 that new AI devices move through staged authorisation, restricted use first, full approval only once real-world evidence backs it up, a model one regulatory forecaster has already nicknamed "L-plates" for AI. The Manchester sandbox is the first place that model is actually being run, rather than just proposed.

The so-what: this sandbox will not touch your practice directly, it is NHS-only, but it is worth watching. A device that has been tested inside a live NHS sandbox and shown to work under real clinical conditions is starting from a stronger evidence base than one still "MHRA registered" on paper alone. We looked at who actually carries the risk when a device falls short of that bar in 0 AI malpractice cases. Insurers move first; sandbox evidence is one of the few things that could start closing that gap before a claim forces the question. Add sandbox participation to your own checklist for any device that reaches your shortlist next year.

2. Eight NHS trusts put AI in 104 operating theatres, and the target is dead time, not diagnosis

Proximie, working with AWS and Deloitte, launched what is described as the largest NHS AI theatre evaluation to date on 10 September 2026: eight trusts, six of them in London, running AI across 104 operating theatres, endoscopy suites and catheterisation labs, covering more than 20,000 procedures over a 12-month evaluation. The tool being tested, Proximie's Intelligence Suite, is not making clinical calls. It measures the dead time between cases, the gap between one patient leaving theatre and the next being anaesthetised, which is where most unused capacity actually sits. Source: Proximie, AWS and Deloitte NHS surgical AI evaluation, 10 September 2026.

The so-what: you do not run an operating theatre, but the underlying finding travels. The AI opportunity most independent practices skip past is not diagnosis or documentation, it is the scheduling gap between one patient and the next. A physiotherapy or dental practice with three or four rooms has its own version of dead time, and it is measurable with far less infrastructure than 104 theatres. If your booking system cannot tell you your own gap-between-patients figure today, that is a cheaper place to start than a clinical AI tool.

3. An insurer's own numbers say AI scribes and coding tools added $1bn to hospital bills, not sicker patients

Blue Cross's internal analysis, reported by HIStalk on 25 September 2026, found that hospital use of AI coding tools and ambient scribes generated an estimated $1 billion in additional charges during 2024 and 2025. One Blue Cross medical executive put it plainly: the disconnect between diagnoses and treatment suggests AI is identifying more billable conditions, not sicker patients. Separately, McLaren Health's chief financial officer told the same outlet that its AI coding vendor now captures $1 million a month in additional revenue the system says it was previously missing, while insurers deploy their own AI to scan the same charts for grounds to deny the claim. Source: Blue Cross internal analysis and McLaren Health, via HIStalk, 25 September 2026.

The so-what: this is a US billing story, but the mechanism is not US-specific. Any AI tool sold on its ability to "capture more revenue" or "reduce under-coding" is, by definition, changing what gets billed, not only how fast the note gets written. Before you adopt a scribe or billing tool marketed on a coding-uplift promise, ask what proportion of that uplift your own indemnity insurer or funding body would recognise as a genuine correction rather than an aggressive one. Get that answer in writing before the tool goes live, not after your first billing query.

4. Heidi Health's $900m valuation is a bet on AI agents, not better notes

Heidi Health, the AI scribe already weighed against Dragon Medical and Nabla by practices in this sector, raised $100 million in new equity plus a separate $240 million customer-acquisition financing facility, at a $900 million valuation. Heidi says the money will fund expansion beyond documentation into what it calls clinician-supervised AI agents: tools that do not just write up a consultation, but take some of the steps that follow from it. Source: HIStalk, "Healthcare AI News 9/23/26," 23 September 2026.

The so-what: every AI scribe vendor is being funded to become something bigger than a scribe. That is worth knowing before you sign a multi-year contract for "just the notes", because the tool you buy today is likely to look and price differently within eighteen months, once agent features arrive as an upsell. We ran the head-to-head on today's note-taking tools in Heidi vs Nabla vs Dragon Medical: physio note AI verdict; the comparison is worth revisiting once any of these vendors adds agent features to what you already pay for.

5. Physicians striking over AI is now a labour issue, not just a workflow one

Around 150 unionised physicians at Allina Health's Mercy and Unity hospitals staged a four-day strike in September 2026, and physician control over how AI is used in patient care, diagnosis and billing decisions was named as a major sticking point. Source: HIStalk, "Healthcare AI News 9/23/26," 23 September 2026.

The so-what: most independent practices in this sector have no union to negotiate through, which cuts both ways. You can move faster than Allina Health, but you also carry the whole decision yourself, and there is no collective bargaining process standing between a bad rollout and the staff who have to live with it. If you introduce an AI tool that changes how a clinician documents, diagnoses or codes without asking the clinicians who will use it first, you are choosing to find out the hard way whether they agree with the switch. A ten-minute conversation with your clinical team before go-live, covering what the tool changes and who checks its output, is cheaper than finding out afterwards.

What this week adds up to

Put the five stories next to each other and a pattern holds. The UK is still building the scaffolding: one sandbox, one theatre trial, both closely supervised, both measured before anyone calls them a success. This week's US numbers show what happens once the scaffolding comes down and the incentives take over: a billing system that rewards more coding gets more coding, and a workforce that was not consulted eventually says so, out loud, on a picket line. Neither future is fixed for a UK independent practice. What is fixed is the shape of the decision in front of you: adopt carefully, measure what the tool actually changes before and after, and put your own clinical team's view on record before you switch anything on, not after.

If you want a structured, independent read on where your own practice sits before you commit to any of this, the AI Opportunity and Growth Assessment benchmarks you against the CARE Framework in two weeks. Or book a 20-minute discovery call and we will work through what this week's five stories mean for the tools you are actually considering.