NHS productivity and AI: what health leaders should do before the review lands

Beyond the copilot: why assistants do not move NHS productivity

The first wave of NHS AI adoption has been assistants. Ambient documentation tools that draft a clinic letter from a consultation. General-purpose copilots that summarise a thread or draft a policy. These are genuinely useful, they are cheap to trial, and they are the easiest thing to put in a board paper.

They are also the category where the productivity argument is weakest, for a structural reason. An assistant gives time back to an individual. Whether that time converts into additional clinical capacity depends on how the service around that individual is organised, and an assistant changes nothing about that organisation. A consultant who saves twenty minutes on documentation in a clinic that is rate-limited by room availability has not created a single extra appointment.

Suppliers in the sector make a version of this point. Darren Kilroy, chief medical officer at RLDatix, put it plainly in reaction to the Budget: "Sustainable productivity won't come from asking clinicians to do more." Richard Craven, managing director of Dedalus UK and Ireland, framed the test as whether "every pound of digital investment should be judged on the capacity it unlocks and the outcomes it improves".

Capacity unlocked is a harder measure than time saved, and it is the measure a review feeding a spending review is likely to apply.

Time saved by an individual is not capacity created by a service. The gap between those two measures is where most NHS productivity business cases quietly fail.

Where bespoke AI moves the needle on NHS productivity

The workflows where AI changes throughput rather than individual workload share a characteristic: the decision logic is specific to the organisation. That is precisely what a packaged product cannot ship, and it is why this category needs work built around the trust rather than around the tool.

Discharge

Delayed discharge is a flow problem with a dozen inputs: therapy assessments, pharmacy turnaround, social care capacity, transport, family availability, equipment delivery. AI is useful here not for writing the discharge summary, though it can, but for predicting which patients are likely to be delayed and why, early enough for someone to act. That prediction depends on the trust's own historical data, its own definitions of a medically fit patient, and its own relationships with the local authority. None of that is portable between trusts.

Rostering

Rostering is where productivity and staff experience either align or collide. The optimisation problem is well understood in the abstract and almost never standard in practice, because every trust carries its own mix of substantive, bank and agency staff, its own skill-mix rules, its own local agreements and its own escalation thresholds. A model that cannot encode those constraints produces rosters that a roster manager overrides, which is the same as producing nothing.

Referrals

Referral triage sits at the boundary between primary and secondary care, which means it depends on how a particular ICB has configured its pathways. Getting referrals to the right clinic first time reduces onward referrals, repeat appointments and the administrative tail behind both. The clinical thresholds, however, are local, versioned and contested, and they are the part that has to be right.

This is not an argument that every trust should commission bespoke development. Where a workflow is genuinely standard, a packaged product is the right answer, and our view on where off-the-shelf software still wins has not changed. The argument is narrower: the workflows with the largest upside for productivity in the NHS are, disproportionately, the ones where the logic is institutional. That is also why so many of these projects stall between pilot and rollout, a pattern we have written about in moving AI from proof of concept to production.

Evidence the review will want to see

A review that reports into a spending review is, in substance, a value-for-money exercise. Jennifer Dixon, who heads the Health Foundation, argued after the Budget that the first job is building ways of testing whether AI and other new approaches actually improve patient care and offer value for money. That is a pointed observation, because it is the thing most NHS AI pilots cannot currently demonstrate.

The practical difficulty is that evidence of improvement requires a baseline captured before deployment, and a trust that starts measuring when the review is announced has already lost the comparison. Four NHS productivity metrics are worth establishing now:

• The pre-deployment baseline. Cycle time, unit cost and exception rate for the specific pathway, measured before anything is switched on. Retrospective baselines are always disputed.
• The capacity conversion. Not hours saved, but what those hours became: appointments delivered, beds released, referrals resolved at first contact. This is the number the review is likeliest to ask for and the one trusts most often cannot produce.
• The exception queue. What proportion of cases the system cannot complete without a person, and whether that proportion is falling. A tool with a large and static exception queue has moved work rather than removed it.
• The safety and governance record. Clinical risk management evidence under DCB0129 and DCB0160, the DTAC assessment, DPIA and the audit trail of what the system did and who reviewed it.

That last point is where the "built in from the start" argument stops being a slogan. The clinical safety standards require a named clinical safety officer, a hazard log and a clinical safety case report. A system designed without them can sometimes be retrofitted, but the retrofit usually means re-opening architectural decisions about logging, human review points and explainability, which is expensive at exactly the moment a trust wants to scale from one ward to ten. The same logic applies to information governance: a model that was never designed to record what data it touched cannot later be made to prove it.

Data sovereignty has become part of this conversation too. Nick Wilson, chief executive of System C, responded to the Budget by noting that "it matters who owns the technology that holds our most sensitive health data", a question boards are increasingly asked to answer in procurement rather than in principle.

Preparing a bid for the AI Challenge Prize

Alongside the review, the government announced a £15 million AI Challenge Prize on 24 October 2026. HM Treasury described it as rewarding "innovative AI that fixes people's common problems", with the first award targeted at firms and innovators solving health issues with AI.

The detail that matters for anyone considering a bid is not yet published. At the time of writing, no administering body has been named, and no application process, deadline or eligibility criteria have been set out. Any adviser telling you how to apply today is guessing.

What can be done now is preparation that holds regardless of how the competition is structured, and which is largely the same preparation the review will reward:

• A named pathway with a measured baseline
• Evidence the problem is common, not local
• A clinical safety case already started
• A route to evaluation, not just deployment
• Clarity on data ownership and hosting
• An integration plan for existing systems

There is also a route that is already open. Under the £100 million Sovereign AI R&D Procurement Scheme, announced on 31 August 2026 and updated on 23 October 2026, the government launched an NHS productivity challenge run with the Department of Health and Social Care. For an organisation that wants to move before early 2027, a live procurement competition is a more concrete prospect than a prize whose rules do not exist yet.

Two further Budget measures are worth noting for planning purposes. £190 million was allocated to help manage winter pressures in England this year. A separate £100 million for robotic surgery comes from the existing £300 million NHS Transformation Fund rather than new money, with the stated ambition of moving from around 70,000 robot-assisted procedures a year to 500,000 by 2035, supported by a national registry using AI to track outcomes. And £24 million will expand the public sector AI defence programme across local authorities, hospitals and schools, which matters because an AI system inside a trust is a new security boundary as well as a new clinical one, a point we made in our piece on AI agent security after SalesBleed.

How Imobisoft approaches NHS productivity work

We work on the assumption that a clinical AI system has to be governable before it is useful, which means the safety case, the audit trail and the human review points are designed alongside the model rather than added once it works. Our work across multiple industries follows this pattern: AI that drafts, a specialist who reviews, and a record of both. The regulated-sector constraints that shape that kind of build are the same ones we discuss in our article on AI in pharmaceutical MLR review.

Where a trust's pathway logic is genuinely institutional, that work sits in AI solution design and system integration rather than in a licence. Where it is not, we will say so.

Frequently asked questions