The Five Questions · Worked example
One tool, five questions
What a governance review with me actually produces: the findings a board saw, the conditions it set before approval, and what changed as a result.
The situation
A supplier offered a mid-sized acute trust an AI deterioration-prediction tool for its adult intensive care unit: risk scores generated from vital signs and laboratory results in the electronic patient record, surfaced on a dashboard to prompt earlier review. The executive sponsor wanted a decision within six weeks; the clinical team was enthusiastic; the published performance figures looked strong. The board asked for an independent clinical review before approving deployment. The Five Questions structured that review.
The review, question by question
What I found: the validation studies came from large academic centres overseas. This trust's case mix was older, more deprived and more comorbid, with a different admission profile. Nobody could say how the model performed on patients like these, and the failure modes (silent degradation, alert fatigue, missing-data behaviour) had not been mapped locally.
Condition set: one quarter of silent running, the tool scoring live data but invisible to clinicians, judged against locally agreed performance thresholds, with drift monitoring specified before go-live rather than after.
What I found: the supplier had never reported performance by subgroup. When stratified metrics were requested, performance fell in patients over 80 and in those with incomplete observations, who skewed towards out-of-hours admissions: precisely the patients a deterioration tool exists to protect.
Condition set: subgroup performance published to the trust before deployment, alerting thresholds adjusted for the weaker subgroups, and an equity review written into the six-month evaluation.
What I found: the draft procedure named "clinician judgement" as the only safety net, with no named owner. If the tool missed a deterioration, or prompted one that led to harm, it was unclear who answered: the bedside nurse, the consultant, the informatics team or the supplier.
Condition set: a named clinical owner at consultant level; incidents routed through the trust's existing governance rather than a parallel process; the supplier's obligations on model updates and error notification written into the contract, not left to goodwill.
What I found: there were no patient-facing materials at all, and the clinicians piloting the dashboard could not explain in plain language why a score had risen. A tool influencing escalation decisions was, in effect, invisible to the people it was used on.
Condition set: a one-page clinician explanation, a patient information leaflet in plain English, and an entry in the trust's public register of AI tools in use.
What I found: the business case priced the year-one licence and nothing else. No budget for revalidation after the next electronic patient record upgrade, no integration maintenance, no training for a rotating workforce, and no answer to "how do we switch this off safely if it stops working?" The case failed in year three, which is where most clinical AI actually fails.
Condition set: a three-year total cost of ownership including revalidation and training, and a written exit plan agreed with the supplier before the first patient was scored.
The outcome
The board did not reject the tool, and it did not wave it through. It approved a conditional pilot: silent running first, then a single-unit deployment with a named owner, published subgroup performance, patient information in place and a quarterly review. The supplier delivered the stratified data and the consent materials within eight weeks, faster than the procurement would otherwise have moved, because the conditions were specific and answerable.
That is the point of the framework. Governance done early is not the brake on adoption; it is what makes adoption defensible, to a regulator, to a coroner, and to the patient who asks "why did the machine say that about me?"
This is the scrutiny your product will meet in every serious NHS procurement. Meeting it in your design phase, with subgroup evidence, an accountability model and a year-three answer ready, is the difference between a six-week decision and a stalled deal.
These five questions, asked before approval, are what "we adopted this with proper governance" looks like in evidence. Each condition above is a paper trail your board can point to when it is asked to account for the decision.
Want this lens on your tool or your deployment?
A review like this typically takes a few weeks, not months, and ends in conditions a board can approve against.
Reminder: the example above is an anonymised, illustrative composite. No real trust, supplier or patient is described.