A Ceiling Is Not a Roll-Out

The Slide That Quotes £900m

Picture a programme board asked to approve a first call-off against a national AI framework. The paper opens with £900m. The second slide says the regional scheme covers 70,000 clinicians. The third reports a neighbouring trust's £1.9m contract. This is an illustrative composite, not a reported case, but every figure on those slides is real and in the public record. The difficulty is that none of them says how much AI anyone is using, or what it has achieved.

That is the argument of this paper. The public record of health service AI procurement documents purchasing authority with some care. It does not document deployment or outcome. Leaders who size programmes, benchmark suppliers or set benefit baselines from the first set of numbers are using evidence built for a different question.

A ceiling says what may be bought. An award says what one buyer agreed to pay. A coverage claim says who may use a scheme. None says what is in use.

Three Numbers, Three Questions

Three kinds of figure dominate the English coverage. They are easy to run together because all are large and all are public.

The ceiling is a legal envelope. NHS Shared Business Services describes a framework with eight lots, running from 12 May 2027 to 11 May 2035, with awards expected in March 2027 and call-offs by mini-competition or direct award². The earlier market-engagement notice, published in January 2025, stated an intention to establish a framework, listed six indicative lots, gave no value and committed no spend¹. The Register reported that the same procurement had been priced at a £150m maximum at that stage, and that NHS SBS later said suppliers and customers had indicated a higher threshold was appropriate⁷. I could not find the £150m figure in the January notice itself, so it rests on that report. Even taking it at face value, the fivefold rise tells us how much room was left for future buyers. It does not tell us how much AI the health service expects to use.

The coverage claim has its own softness. One trade account says the Midlands framework covers 70,000 clinicians across 15 acute and community trusts and all integrated care boards, with the contract value undisclosed³. A second, summarising a paywalled HSJ story, says the framework is opt-in, closed to new joiners, and that five trusts had begun deployment, with ten others unnamed⁴. These accounts differ on scope. A framework that covers a population and a service that has reached it are different things, and the sources do not tell us which one 70,000 refers to in practice.

The award values are the most solid numbers and the most narrowly scoped. The Croydon contract is for five months of implementation support, ending on 31 October 2026, and is not the product licence⁵. The Leicester and Northamptonshire award followed a pilot and a five-tender evaluation weighted 60 per cent to quality and 30 per cent to cost⁶. Neither number is an adoption figure.

Where the Record Goes Quiet

Looking for the missing layer, I read the national material. NHS England's guidance on ambient scribing describes a self-certified supplier registry that asks suppliers for evidence of real-world benefit in NHS settings, states that NHS England does not endorse listed suppliers and performs only preliminary completion checks, and says nothing in the portions I could read about how many trusts or clinicians are using these tools⁸. A joint NHS England and MHRA letter of 29 July 2026 clarifies that products which only transcribe, summarise, draft correspondence or suggest codes for clinician review are not regulated as medical devices under current rules, while those supporting diagnosis, treatment or automated clinical action are⁹. That is a useful boundary. It is also a reminder that for part of this market no regulator is collecting post-market data.

Independent datasets count purchases, not use. BidSkim records 252 confirmed AI-related awards by health bodies, 229 with a published value totalling £302.8m, of which the Federated Data Platform's £182.2m is 60 per cent; ambient voice and clinical documentation accounts for 45 awards and £24.1m¹⁰. The same report states plainly that its values are award notices rather than payments, that it excludes framework ceilings, and that published values are minimums¹⁰. It is the right kind of caution, and it is also the point: the best public dataset cannot see deployment.

The outcome claims that do exist are self-reported. Leicester's trust-reported figures, eight minutes saved per patient on post-clinic documentation and about an hour of administrative time a day, come with no sample size or method⁶. The Dudley figures, an 80 per cent reduction in emergency documentation time and a six-month letter backlog cut to 14 days, are the supplier's claims as relayed by a patient-safety site⁴. They may be right. They are not independent evidence.

The Federated Data Platform shows what happens to headline values over time. BidSkim records £182.2m, flagged as a published estimate based on use, with the notice stating expenditure is not committed¹⁰. The Register reports a £330m award, a government review ahead of a contractual break point next February, and benefit figures published by the supplier, including 110,078 additional patients through theatres, which the article does not independently verify¹¹. The two contract values may reflect different bases. The sources do not reconcile them.

The Sceptic's Case, Taken Seriously

The strongest objection is that none of this is a failing. Procurement notices exist to show that a buyer acted lawfully, not to report adoption. Ambient voice go-lives in the Midlands began only in May 2026³. The registry already asks for evidence of benefit. Frameworks are sized generously so they do not run out before the term does. Where a buyer measures, as Leicester did in its pilot, numbers appear. And an operator that wants to count deployment can: Kaiser Permanente's early report said 3,442 physicians had used its ambient scribe in as many as 303,266 encounters within ten weeks, though it gave no quantified time saving¹².

I accept most of this, and it narrows the claim. The paper does not say that the NHS is hiding deployment data, that the framework ceiling was unjustified, or that ambient tools save no time. Silence in notices written for another purpose, six months after the first go-lives, is not evidence of failure. What it does establish is a limit on what a leader can responsibly infer today. The Kaiser example sharpens that: deployment is countable when someone is asked to count it. The gap in the English record is a reporting design choice that can be revised, not a law of nature.

From Notice to Board Decision

The mechanism is mundane. A notice reports authority and price because that is its job. Trade press and datasets quote those figures because they are the only systematic public numbers. A board paper quotes the press or the dataset because it needs a scale for the investment. Deployment, meanwhile, is local and optional: trusts join frameworks or do not, go live in phases, and measure benefit to their own methods if at all. No document aggregates it, so the purchasing numbers stand in.

Take the illustrative board again. It is asked to approve a call-off partly on the strength of £900m and 70,000. If the trust's own deployment ledger shows two wards live and a handful of active users, the true baseline is small, and a benefits case built on population coverage will overstate the base substantially. If the ledger shows broad use but no outcome measure, the case is still unevidenced. In neither case would the headline numbers have warned the board.

A Decision Test for the Next Call-Off

A programme or portfolio leader can apply a short test before any further commitment against a framework of this kind.

1. Name the figure. Is the number in the business case a ceiling, an award, a coverage claim or a measured use, and is it quoted with or without VAT?
2. Ask what "covered" means. Contracted, permitted or live, and how many people used it in the last month?
3. Find the measure. Who recorded the claimed benefit, by what method, with what comparison?
4. Tie the money to the evidence. Make deployment counts, active use and an agreed outcome measure a condition of each call-off tranche and of any extension.
5. Reconcile before quoting. Where two datasets or two reports give different totals for the same thing, state the basis of each or leave both out.

None of this requires the national framework to change. It requires that the question a ceiling cannot answer is asked of whoever is asking for the money. Until the record can answer it, the honest description of the health service's AI position is one of authority to buy, with use still to be counted.