5 AI stages to 2033, Colorado day one: this week in AI

The most useful document in UK health AI this week was not a regulator's guidance note or a vendor's white paper. It was a hospital board pack. Mid Cheshire Hospitals put its seven-year AI plan into the public domain, and inside it sits the thing practice managers keep asking suppliers for and keep not getting: a ranked sequence, with the governance each step actually requires written next to it. Five stages, four domains, published under NHS transparency rules and free to lift. That leads this week, alongside a therapy AI law that went live on Wednesday, a national safety investigation opened one day after a trust switched on a full rollout, and a rejected funding bid that explains more about the year ahead than any of them.

1. An NHS trust published a five-stage AI ladder, and stage one is where you already are

Mid Cheshire Hospitals NHS Foundation Trust launched a seven-year digital and data strategic plan running to 2033, set out in its public board papers of 30 July 2026 and picked up on 10 August. It maps AI deployment across four domains (clinical, operational, workforce and patients) and five progressive stages. Stage one, "support", covers summarisation, documentation and ambient voice on the clinical side, reporting and basic analytics operationally, admin support and transcription for the workforce, and information delivery for patients. Decisions at this stage are made by humans only, and the governance named against it is three things: transparency, privacy, and mandatory human review. Stage two, "recommend", adds clinical decision support and triage prompts on a human-in-the-loop basis, governed by clinical safety assurance, validation and audit. Stage three, "enable", lets AI prepare or initiate actions (pre-populating notes, orders, care plans, scheduling, patient communication) but still requires human approval, and brings in clinical safety officer oversight, role-based accountability and bias monitoring. Stage four automates defined actions inside set thresholds with human overrides. Stage five is autonomous decision-making, described as a future state. Source: Mid Cheshire Hospitals NHS Foundation Trust, Board of Directors bundle, 30 July 2026, reported by HTN on 10 August 2026.

The number: five. Five stages, each with its own governance controls, published by an organisation with no product to sell you.

The so-what: almost everything an independent practice can realistically buy today (a scribe, a recall tool, a transcription assistant, an analytics dashboard) sits in stage one. That matters because it tells you which governance burden is proportionate. You do not need a bias monitoring programme or a clinical safety officer to run an AI scribe, because those controls belong to stages two and three. You need three paragraphs: what the tool is and who knows it is in use, what happens to the data, and who reviews the output before anyone relies on it. Write those three paragraphs against each tool you run, date them, and you have documented governance matched to what you are actually doing rather than to what a vendor's compliance pack assumes. Our longer walkthrough of what an inspector looks for sits in the 10 AI governance checks CQC will look for in 2026.

2. Colorado's AI therapy law went live on Wednesday, and the sponsors named the carve-out

Colorado HB 26-1195 took effect on 12 August 2026. In the joint release announcing commencement, Representative Gretchen Rydin stated that the law "does not hamper licensed professionals' use of AI for administrative purposes, but does require all clinical treatment and psychotherapy to be administered by a licensed, human provider". The same release confirms two provisions that most coverage skipped: AI chatbots may not be marketed to patients as equivalent to a licensed psychotherapist or counsellor, and they may not imply that patient input is covered by ordinary confidentiality protections such as HIPAA. Source: Colorado Senate Democrats joint release, 3 August 2026, and Colorado General Assembly HB 26-1195.

The number: 12 August 2026. The first US state law with a bright line inside the therapy session is now in force, four days before this briefing.

The so-what: notice what the operative test actually is. Colorado is policing the claim a product makes about itself, not the code inside it. That is precisely the test the MHRA applied on 29 July when it confirmed that an ambient scribe's medical device status turns on the manufacturer's stated intended purpose rather than its features, which we worked through in why your AI scribe's regulatory status is set by marketing. Two regulators, two continents, five weeks apart, converging on the same rule. For a UK practice the practical action is identical: pull up your supplier's public website and read what it claims the product does, because that page, not your internal policy, is what a regulator will read first. We covered what the Colorado statute requires in full, including the notice duty that binds practices using no AI at all, in Colorado's AI therapy law and the provision the headlines skipped.

3. A trust reported real admin savings, then said it cannot afford to move

Leeds and York Partnership NHS Foundation Trust gave a board update on 12 August 2026 confirming two ambient voice technology pilots, with Heidi and with Anatham, running through to October 2026, with formal evaluation commencing in November. The trust reported real reductions in clinical admin time from the pilots. The same update recorded that its Frontline Productivity Fund application had been rejected and that capital funding "remains the most significant challenge", with its electronic patient record tender not due until September or October and business case approval targeted for January 2027. Source: Leeds and York Partnership NHS Foundation Trust board update, 12 August 2026, reported by HTN.

The number: November. Not when the trust finishes evaluating an AI scribe that is already saving its clinicians time, but when the evaluation starts.

The so-what: this is the clearest picture available of the gap you are competing across. A mental health trust has evidence in hand that a tool works and will still take until 2027 to act on it, because the money sits in a capital process that just said no. A three-clinician practice can run the identical software (Heidi is sold directly to independent clinicians) for four weeks, read its own notes, and decide. Speed of decision is the structural advantage independents hold right now, and it is worth more than any feature comparison. If you have been waiting for the sector to settle before choosing, this is what waiting looks like from the inside.

4. The safety investigation opened one day after the rollout started

Kingston and Richmond NHS Foundation Trust began a trust-wide ambient voice technology rollout on 5 August 2026. On 6 August, the Health Services Safety Investigations Body opened a national patient safety investigation into the use of ambient voice technology in hospitals. HSSIB's own notice states that "adoption of AVT is accelerating while the safety implications are not fully understood", that national literature and implementation activity "has focused more strongly on efficiency benefits than on patient safety risk", and that "routes for recognising and reporting AI-related incidents are not yet mature enough to provide confidence that emerging risks are being identified". The investigation covers acute adult secondary care and reports in summer 2027. Source: HSSIB investigation notice, 6 August 2026.

The number: one day. One day between a trust switching a tool on across its services and the national safety body opening a formal investigation into the category.

The so-what: read the third quotation again, because it is the one that reaches you. If NHS organisations with incident reporting systems, clinical safety officers and Datix do not yet have mature routes for spotting AI errors, an independent practice has none at all. Build one this month, and keep it deliberately crude: a shared spreadsheet with four columns, being date, clinician, what the AI got wrong, and what was done about it. It costs nothing, it takes seconds per entry, and it is the single artefact that converts "we use AI carefully" from an assertion into evidence. When HSSIB reports next summer, and when your own regulator starts asking, the practices with eleven months of logged entries will be in a different conversation from the ones with a policy document.

5. The clearest explanation of item 3 was published in January

A study that circulated again this week, originally published in the American Journal of Managed Care in January 2026, measured ambient AI adoption across US hospitals. Of 6,561 hospitals, 2,784 were identified as Epic users, and 1,744 of those (62.6%) had adopted an ambient AI documentation tool as of June 2025. The adjusted figures are the interesting part. Adoption ran at 70.2% among nonprofit hospitals against 28.8% among for-profit hospitals, at 67.6% in the strongest operating-margin quartile against 58.0% in the weakest, and at 64.7% in metropolitan against 54.3% in non-metropolitan hospitals. Three products, DAX Copilot, Abridge and ThinkAndor, accounted for more than 80% of implementations. Source: Yang F and Graetz I, American Journal of Managed Care, January 2026, volume 32 issue 1, pages e25 to e30.

The number: 62.6%. The more useful number is the 9.6-point gap between the best and worst operating-margin quartiles, and the 41-point gap between nonprofit and for-profit ownership.

The so-what: adoption tracked the balance sheet, not the evidence. The authors say so directly, concluding that cost and uncertainty about return on investment are likely barriers for financially constrained organisations. That is exactly what Leeds and York demonstrated in real time this week. Large organisations buy AI through capital processes with committees and funding rounds. An independent practice buys it as a monthly operating cost per clinician, typically between £15 and £150 depending on whether it is bundled into your practice management system, a comparison we ran in the bundled versus standalone AI scribe decision. You are not disadvantaged on price. You are advantaged by the absence of a capital gate, and that advantage has a shelf life.

What this week adds up to

Put the five together and a trade becomes visible. The NHS is building the governance scaffolding for clinical AI in public, at considerable expense, and publishing it for free, while being too capital-constrained to move at the speed its own pilot data justifies. Independent practices have the mirror-image problem: they can decide in a fortnight and have no scaffolding at all. Take one and supply the other.

That means three borrowed documents and one you write. Mid Cheshire gives you the stage-one governance list (transparency, privacy, mandatory human review) and confirms everything heavier belongs to capabilities you are not running. Colorado and the MHRA between them give you the supplier test, which is to read the claim rather than the feature list. HSSIB gives you the gap, which is incident reporting. The one you write is the log, and it starts before the pilot, not after.

Frankly, if the choice is spending September writing an AI policy from a blank page or running a four-week pilot with an incident log attached, run the pilot. The policy can be written out of the log afterwards and it will be accurate. A policy describing tools you have never used is the document least likely to survive a question from an inspector or a patient.

Two dates worth holding: Leeds and York's evaluation begins in November 2026, putting a comparable UK read on ambient voice in mental health roughly a quarter away, and HSSIB reports in summer 2027. If you want a clear read on where your practice sits before you buy anything, 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 which stage-one tools are worth your September.