Most DHT companies treat NICE ESF Standard 6 as a sentence in the user manual. “Outputs should be reviewed by a qualified clinician before action is taken.” That sentence is not an oversight model. It is a disclaimer dressed up as one — and NICE evaluators, NHS commissioners, and now the EU AI Act all know the difference.
Standard 6 is the standard where the autonomy debate in clinical AI becomes commercial. It asks a deceptively simple question: how much should the human in your loop actually do, and can you prove the model still works when they do less? Get this right and you have built a clinical governance asset that compounds with every deployment. Get it wrong and you have built a liability transfer mechanism — and the system has stopped buying those.
This post is for DHT founders and product teams who have read Standard 6 and assumed it meant adding a sentence to the IFU. It does not. It means engineering professional oversight as proportionately as you engineer the algorithm itself.
What Standard 6 Actually Asks (and Why People Misread It)
Standard 6 sits inside the Design Factors group of the NICE Evidence Standards Framework and applies across all tiers: A, B and C. The wording is short. The implication is not. NICE expects manufacturers to define, at design stage, the level of professional oversight needed when the product is used in practice, and to make that level proportionate to the clinical risk of the DHT failing to perform.
That word, proportionate, does most of the work. It means that a Tier B health diary and a Tier C diagnostic AI cannot share an oversight model. It means a clinician glancing at an output is not the same control as a clinician reviewing it. And it means that whatever oversight you specify must actually be deliverable in the NHS environment your product enters, not the staff-rich, time-rich version that exists in your product spec.
Three misreadings show up repeatedly in submissions and procurement assessments. The first is treating oversight as a disclaimer. The second is designing for an idealised clinical workflow that does not exist on a Tuesday morning in a busy district general. The third is using oversight requirements as a mechanism to push liability onto the deploying clinician. NICE evaluators read every one of these as the same signal, that the company has not done the clinical governance design work, and is hoping the buyer won’t notice.
The Four Oversight Models — Pick One Deliberately
NICE recognises that oversight is not one thing. The framework allows for several oversight models, and the right one depends on the tier, the use case, and the operational environment. Picking the model is the design decision; everything else flows from it.
Table 1 maps the four primary models to the tiers they best fit, the engineering requirements they impose, and a representative example. The point is not that one model is superior — it is that the model must be matched to the risk and engineered into the product, not bolted on as a post-hoc workflow note.
The most common mistake here is choosing case-by-case review by default because it sounds safest. It is the safest design choice only if the deploying environment can sustain it. A Tier C product that requires consultant-level review of every output, deployed into a department with one consultant covering three wards, will be circumvented within a fortnight — at which point your oversight model has failed and your safety case along with it.
Why Automation Bias Means Oversight Must Be Monitored, Not Just Defined
There is a deeper reason Standard 6 needs to be designed, not declared. Even when oversight is well-specified, it degrades in practice. Clinicians become habituated to approving outputs without genuine review. Volume creeps up. Staffing changes leave gaps. The phenomenon has a name — automation bias — and the evidence base for it has firmed up considerably in the last twelve months.
A randomised clinical trial published in NEJM AI tested whether AI-literacy training protected physicians against accepting incorrect AI-generated diagnostic recommendations. It did not. Trained physicians still followed wrong AI suggestions at rates that materially degraded their diagnostic accuracy. Training the user is not a substitute for designing the oversight system.
The implication for product teams is straightforward. If your oversight model says “the clinician will review every output,” you need engineering evidence that the clinician is actually reviewing, not rubber-stamping. That means override capture in the workflow, statistical review of disagreement rates, friction designed into the interface where the stakes warrant it, and a clinical governance loop that closes when override patterns drift. Standard 6 is no longer a documentation standard. It is a monitoring standard.
This is also where Standard 6 connects to Standard 16 (post-market performance monitoring). The companies presenting credible Standard 6 evidence to NICE are the ones who have already wired the override-and-drift telemetry that Standard 16 will ask for anyway. Build it once.
EU AI Act Article 14 Is Pulling the UK Bar Higher
The regulatory environment around oversight has not stood still. Article 14 of the EU AI Act — which applies to high-risk AI systems including most clinical AI — mandates human oversight measures “commensurate with the risks, level of autonomy, and context of use.” That is the same logic NICE wrote into Standard 6, sharpened with legal teeth. For UK companies operating in either market, designing to Standard 6 now is effectively designing to the direction of EU compliance.
On the UK side, the MHRA’s Software and AI as a Medical Device Change Programme continues to harden expectations on human oversight, and the MHRA’s AI Airlock regulatory sandbox — expanded with a £3.6m investment over three years — is increasingly the venue where novel oversight models are stress-tested before they reach broader deployment. Companies positioning AI as “removing the need for clinical review” are running into the institutional culture of the medical profession, which has not shifted on accountability and is unlikely to.
What Good Looks Like — and Where to Put the Engineering Effort
The strongest Standard 6 submissions read as governance design briefs, not compliance documents. They are explicit about three things in particular.
Co-design with the clinical end users. The oversight model has been built with the clinicians who will operate it, not handed to them. There is documentary evidence of workflow piloting and iteration. This same evidence does double duty for Standard 8 (credibility with UK professionals).
An accountability map. Manufacturer responsibility for algorithm performance, deployer responsibility for governance implementation, clinician responsibility for individual decisions — written down, owned, and visible to procurement. Ambiguity here is read as risk transfer, and risk transfer is read as a no.
Override and drift telemetry from day one. If you cannot show, in real time, that override rates are stable, that drift indicators are within bounds, and that the oversight model is functioning as intended, then your safety case relies on assertion. Twelve months of clean telemetry from a live deployment is the strongest Standard 6 evidence you can produce.
None of this is cheap. But the companies that have done it earn something the rest cannot — a credible path to reducing oversight over time as the evidence accumulates. That is the prize. The bar to demonstrate safe autonomy will only rise; the companies that have already wired the monitoring infrastructure are the ones who will eventually clear it.
The Bottom Line for DHT Founders
Standard 6 looks like a documentation requirement. It is actually a clinical governance architecture decision. Pick the wrong oversight model and your product will be quietly abandoned in deployment. Pick the right one and document it as a disclaimer, and NICE evaluators and NHS commissioners will read you as another company that has not done the work.
The work is engineering oversight in proportion to risk, monitoring it in flight, mapping accountability without ambiguity, and treating the resulting telemetry as evidence — both for Standard 6 today and for the autonomy case you may want to make in three years. That is the standard the system is moving toward. Build for it now.
Healthonomix helps DHT companies design Standard 6 oversight architectures that survive contact with the NHS. If you are scoping your evidence strategy, get in touch.
