NHS's unverified 29%, a dental wiretap win: this week in AI

Two of this week's stories are mirror images of each other. One is a widely quoted AI statistic that turned out to rest on data nobody could check. The other is a legal claim that turned out to rest on a standard the AI vendor could meet after all. Between them sits a lesson this site repeats often but is worth repeating again: the number a vendor or a health system leads with is rarely the number that survives scrutiny, and independent practices are better served asking to see the evidence than taking the headline on trust.

1. The NHS's flagship AI statistic didn't meet its own watchdog's bar

The Office for Statistics Regulation ruled that a 29% reduction in GP phone queues, the figure NHS England used to justify accelerating a nationwide AI triage rollout through the NHS App, did not meet official statistics standards. The underlying evaluation from the Sussex GP pilot that produced the figure was never published, so outside academics could not verify it. Oxford researchers filed the complaint that triggered the review. Source: Office for Statistics Regulation, via GB News, August 2026.

The number: 29%. The queue-reduction figure that reached a national rollout announcement before anyone outside NHS England could check the working behind it.

The so-what: this is the same failure mode Section 11 of our own editorial rules exists to catch, playing out at national scale. A percentage with no published methodology is a marketing line, not evidence, whether it comes from a government press release or a vendor's case study. Before any practice treats an AI tool's headline efficiency claim as a reason to buy, ask the same question the Office for Statistics Regulation just asked the NHS: where is the underlying data, and can someone outside the organisation that produced it check the number. We built exactly that checklist for diagnostic AI marketing claims in optometry's autonomous AI code, and the checklist for telling assistive claims from autonomous ones, and the same discipline applies to any admin or efficiency stat, not just clinical ones.

2. A dental AI phone system beat a wiretap claim, on the opposite reasoning to this year's other scribe cases

On 13 January 2026 the US District Court for the Northern District of Illinois dismissed a wiretapping claim against RingCentral and its customer Heartland Dental. The complaint alleged RingCentral's AI-powered phone platform transcribed calls to Heartland-supported dental clinics and ran sentiment analysis without callers' knowledge or consent. The court held that because AI transcription was core to RingCentral's own phone product, it fell within the federal Wiretap Act's "ordinary course of business" exception. The dismissal was without prejudice, so the plaintiff can still amend and refile. Source: Becker's Dental Review and Troutman Privacy, reporting on the ruling, 2026.

The number: one exception. A single, narrow carve-out in a decades-old federal statute, and the only reason this claim failed where others this year have not.

The so-what: don't read this as a green light. It cuts the opposite way to the Sutter Health and Sharp HealthCare cases we've covered, where a signed BAA satisfied HIPAA but not state wiretap law. Here, a federal wiretap claim failed because the AI function was inseparable from an already-disclosed phone service, a narrow and fact-specific defence that has nothing to say about the all-party consent laws in 13 US states, which operate independently of federal law and were not tested in this case. If your practice uses an AI phone or scheduling system that listens to or analyses calls, the safer assumption is still the one in your AI scribe's BAA covers HIPAA, it doesn't cover this $5,000-a-patient lawsuit: check state consent law directly, don't rely on a federal exception that hasn't been tested against your specific setup.

3. The FDA cleared an AI tool to guide ultrasound scans by clinicians who aren't ultrasound specialists

ThinkSono received FDA 510(k) clearance in August 2026 for ThinkSono Guidance, software that gives real-time guidance to healthcare professionals without ultrasound training while they acquire vascular ultrasound images. Source: FDA 510(k) clearance announcement, August 2026.

The number: zero. The amount of ultrasound-specific training the cleared use case requires of the clinician holding the probe.

The so-what: this clearance covers image acquisition guidance, not diagnosis, and it is not evidence that any allied health discipline can now add point-of-care ultrasound to its scope of practice. But the direction matters. AI tools that extend what a non-specialist clinician can safely do, rather than replacing clinical judgement, are the category most likely to reach physiotherapy, OT and similar settings next. To be clear: this is not clinical advice, and any practice considering point-of-care imaging technology needs its own regulatory and scope-of-practice review, not a vendor's marketing claim, before adopting it.

4. Staff and community protests against a major AI vendor spread across US hospital systems

Nurses, caregivers and community members in multiple US cities publicly called this week for hospitals and health systems to cut ties with Palantir, the data analytics and AI vendor increasingly embedded in large US health systems, citing concerns about how patient data is shared and governed. Source: healthcare trade press coverage, week of 25 August 2026.

The number: multiple cities. Not a single local dispute, but a coordinated set of protests across several health systems in the same week.

The so-what: this is a large-system story, but the underlying concern, that patients and staff don't know what happens to their data once an AI vendor is embedded in clinical infrastructure, is exactly the concern an independent practice can address directly and visibly. A practice that can name its AI vendors, explain what data each one touches, and show it took no commission for the choice is answering the question these protests are asking before anyone has to ask it. That transparency is a genuine point of difference against both large systems and the vendors trying to sell into them.

5. The UK's AI healthcare rulebook still hasn't answered its hardest question

The National Commission into the Regulation of AI in Healthcare, the body tasked with producing a UK regulatory rulebook for healthcare AI, published the findings of its Call for Evidence, drawing on 761 responses from patients, clinicians, providers, academics and industry. The findings frame the Commission's task as making AI in healthcare "safe, fast and trusted," and name how legal liability for AI-related harm should be distributed as one of the questions still to be resolved. A final rulebook has not yet been published. Source: Health Research Authority and GOV.UK, 2026.

The number: 761. The response count behind the evidence base, and still no answer on who is liable when a healthcare AI tool gets something wrong.

The so-what: until that question is answered nationally, the liability position for an independent UK practice using AI defaults to existing professional and regulatory duties, the same ones this site has covered for psychology, dentistry and physiotherapy over the past two months. Nothing about this Commission's work changes what CQC, GDC, HCPC or ICO already expect today. It is worth watching for when it lands, not worth waiting for before getting your own house in order.

What this week adds up to

The throughline across all five stories is verification, not adoption. A government's own headline statistic failed a check it should have passed before publication. A wiretap claim failed for a narrow, fact-specific reason that doesn't generalise to a green light. A device clearance covers a specific, limited function, not a broader scope change. A data-governance concern about one large vendor is a preview of the question every practice using AI should be able to answer about its own. And a national liability framework everyone is waiting for still isn't finished.

None of that is a reason to slow down. It's a reason to make sure whatever you already run, or are about to buy, can answer a direct question about what it actually does, what data it touches, and what evidence backs any efficiency claim attached to it. If you want a structured, independent read on where your own practice's AI use actually sits against those questions, 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 stories mean for the tools you already run.