1,000 NHS scribes, Colorado's AI therapy ban: this week in AI

Five stories mattered in healthcare AI this week, and only one of them came from the NHS. Colorado just told psychotherapists exactly what an AI tool is legally allowed to do inside a session, a Hertfordshire trust scaled ambient voice technology past a threshold nobody else in community care has reached, and the AMA drew a line insurers cannot cross with automated coverage decisions. None of the five need action from you by Monday morning. All five narrow the gap between what independent practices assume the rules allow and what those rules are about to actually say.

1. Colorado bans AI from running therapy sessions on its own

Colorado governor Jared Polis signed HB 26-1195 on 3 June 2026. The law applies to every licensed, certified, or registered psychotherapy provider in the state, and it prohibits allowing an AI system to independently engage in therapeutic communication with a client, meaning the AI generates the therapeutic content itself rather than assisting a human clinician who remains in control of the session. Source: Colorado General Assembly, HB 26-1195, via Snell & Wilmer legal alert, June 2026.

The number: 12 August 2026. That is when the law takes effect, ten days from this roundup's publication. It is the first US statute to draw an explicit line between AI assisting a clinician and AI acting as the clinician, rather than leaving the distinction to guidance or professional codes.

The so-what: Colorado's law does not reach UK psychology practices or US practices outside the state, but the underlying distinction it makes, AI assists, a licensed human decides and delivers, is the same one HCPC's binding consent standard already implies without naming AI, as we set out in where HCPC and BPS guidance actually stand on AI consent. Expect more states to legislate this line explicitly rather than leave it to professional bodies to infer.

2. One NHS trust just gave 1,000 clinicians an AI scribe

Hertfordshire Community NHS Trust has chosen Accurx Scribe, powered by Tandem, after a six-month pilot the trust describes as the largest of its kind in England's community and mental health services. The rollout reaches more than 1,000 clinicians and is expected to be used in around 250,000 appointments a year. The software transcribes consultations in real time and drafts a clinical note, summary, or letter, which the clinician reviews and approves before anything reaches the patient record. Source: Clinical Services Journal, July 2026.

The number: 1,000 clinicians. That is a materially larger single-trust deployment than the four-trust south-west London rollout covered in an earlier briefing, and it is the first at this scale specifically inside community and mental health services rather than acute hospital care.

The so-what: the review-and-approve step before anything enters the record is the same governance pattern that matters for any independent practice choosing a scribe, and the same pattern we weighed for a physiotherapy practice in the bundled-versus-standalone AI scribe decision. A mental health deployment at this scale also raises the safeguarding stakes directly, worth remembering whatever specialty your own vulnerable caseload sits in.

3. The AMA tells insurers what AI can and cannot decide

At its June 2026 Annual Meeting, the American Medical Association adopted new policy opposing the use of autonomous or semi-autonomous AI systems as a substitute for physician review in coverage determinations. The policy calls for AI-enabled prior authorization and utilization management tools to disclose the clinical logic, data sources, and guidelines behind any adverse decision. Source: American Medical Association, June 2026, via Fierce Healthcare. It follows a CMS FAQ, published May 2026, clarifying that Medicare Advantage organisations may use AI and algorithms to help inform coverage decisions, but the final determination must still rest with the plan, not the algorithm.

The number: zero. That is how much regulatory force the AMA's policy carries on its own, it is professional advocacy, not law. What it signals matters more: organised medicine is now on record demanding the same transparency independent practices already have to build into their own AI-assisted workflows.

The so-what: this is the physician side of the same story covered in Medicare's WISeR AI prior authorization pilot, where payer-side AI already reviews adjacent procedures. If insurers face growing pressure to explain AI-driven denials, practices that already document their own clinical reasoning clearly are better placed to contest one.

4. Brussels quietly pushes the medical AI deadline back two years

The EU's Digital Omnibus package, agreed in June 2026, delays the compliance deadline for high-risk AI embedded in devices regulated under the EU's Medical Device Regulation or In Vitro Diagnostic Regulation, from August 2026 to August 2028. The deadline for standalone high-risk AI systems classified under the AI Act's Annex III moves to December 2027. Source: Digital Omnibus, June 2026, via IntuitionLabs and Certivo regulatory analysis.

The number: two years. That is how far the deadline moved for AI embedded in a regulated medical device, from a date that had already arrived to one three years out.

The so-what: this mainly affects practices using EU-regulated diagnostic or imaging AI, which is a smaller slice of the independent practices this site covers than HIPAA or CQC questions. The wider point holds regardless of jurisdiction: regulatory timelines for medical AI are sliding in more than one direction at once, so any vendor claim of being "AI Act compliant" is worth a direct question about which deadline they mean.

5. Eye care AI stops just reading images and starts doing the admin

Optometry AI coverage this month shows a shift from passive diagnostic support toward tools that actively manage practice workflow, screening patient charts and automating recall and staff training tasks rather than only flagging findings on a scan. Optivate, a platform built specifically for ophthalmology and optometry practices, opened a new US headquarters this month to support its growth. Source: Optometry Times, July 2026.

The number: two. That is roughly how many optometry-specific vendors, per this site's own comparison work, currently score well on the CARE Framework, a reminder that specialty-built tools remain the exception rather than the rule even as the category grows.

The so-what: if an optometry group is evaluating scheduling or recall AI, the questions worth asking have not changed, only the number of vendors claiming to answer them has. See the optometry AI scheduling comparison for how Weave, Solutionreach, and RevenueWell actually score against each other.

What this week adds up to

Every story above draws the same line in a different setting: a human, not the AI, makes the call, and that human has to be able to show their working. Colorado wrote it into statute for therapy sessions. Hertfordshire built it into a 1,000-clinician workflow as a review-and-approve step. The AMA is demanding it from insurers on the coverage side. Even the EU's shifting deadlines are, underneath the delay, still about proving a human oversight process exists before an AI-embedded device reaches a patient. None of that changes what an independent practice should do this month: keep documenting where your own AI tools assist and where a clinician still decides, because that distinction is what every regulator above just spent the week reinforcing.

If you want a clear picture of where your practice sits against that gap, the AI Opportunity and Growth Assessment benchmarks your practice against the CARE Framework in two weeks. Or book a 20-minute discovery call to talk through what this week's changes mean for your specific practice.