Most boards still treat confidence in AI as a matter of model accuracy plus a signed ethics policy. It is neither. Trust in a deployed system is an operating property: it comes from the controls that run on every request, the monitoring that continues after launch, and the record that someone outside the team can inspect. A heavily regulated part of the economy has just published what that looks like in practice, and the design is more useful to a chief technology officer than most of what is sold as trust infrastructure.
On 10 September 2026 the National Commission into the Regulation of AI in Healthcare published its recommendations, as GOV.UK reported. The commission was set up by the MHRA in September 2025 and is led by two practising NHS doctors, Alastair Denniston as chair and Henrietta Hughes as deputy chair. It drew on evidence from more than 12,000 people over a year, and the Health Foundation ran deliberative workshops with 78 members of the public in Cardiff, Milton Keynes and York. GOV.UK describes the exercise as the largest engagement of its kind in the UK on regulating healthcare technology. Denniston's summary of what he heard, quoted by GOV.UK, is that innovation has to "earn its place and earn our trust".
That is the argument compressed to seven words, and the recommendations are what it looks like when a regulator takes it seriously.
How staged authorisation and continuous monitoring ensure AI safety
It did not ask for a harder gate at the start. It asked for permission to be staged, conditional and reversible, which is a different design.
New AI models would go through staged authorisations, which the report compares to the learner plates carried by new drivers: each model runs under close supervision and narrow limits while it demonstrates real-world safety, before fuller authorisation is granted. AI-enabled medical devices would be monitored for as long as they are in use. Members of the public would be able to search for safety information about specific devices, including adverse incidents, building on tools such as the MHRA's interactive Drug Analysis Profiles. The MHRA would gain stronger enforcement powers. Patients would be told, proportionately, when AI is being used in their care.
The Health Foundation's deliberative research, published the same day, set out five principles. Among them: accuracy is the public's top priority, human oversight is a condition of use, regulation should be proportionate to risk, and AI should not leave any group with worse care. The government and the MHRA will respond formally in due course. The direction, though, is not in doubt.
A single approval assumes the system you approved is the system running six months later
That assumption fails for exactly the models enterprises are now deploying.
Generative systems behave differently in different circumstances and change after deployment. GOV.UK quoted Neil Lawrence, who chaired the commission's technology working group, warning that "we can't rely on a single point of approval and assume the job is done". A model that passed an evaluation in March may be answering a different distribution of questions in September, through a retrieval layer that was reindexed in June, with a prompt template that was edited in July. The approval covers none of that.
OvalEdge's enterprise AI trust framework sets out the corrective: governance that runs across a data layer, a model layer and a decision layer, with a trust signal layer on top that surfaces data quality, lineage coverage, model confidence and policy status. The point of that top layer is that trust stops being a judgement and becomes a measurement someone can check.
Trust between people runs on a separate ledger, and it decides adoption
A model can pass every technical threshold and still fail, because the people asked to use it are running their own assessment.
CCL's work on relational trust in AI transformation separates trust of capability, trust of communication and trust of character. AI strains all three at once: it unsettles expertise, it makes communication feel automated, and it opens gaps between stated values and staffing decisions. CCL's advice is blunt and correct. Leaders who perform certainty about a technology nobody has mastered lose trust faster than leaders who name the uncertainty and describe how it will be resolved.
Signal AI's reframing of the trust equation makes a related point with commercial consequences. Once competence is adequate, reducing perceived self-interest buys more trust than another point of accuracy. A vendor that hides its incentives will be discounted no matter how good the benchmark looks.
Policy belongs on the data path, where every agent failure lands
Every way an agent can go wrong, a poisoned prompt, an over-broad credential, a misconfigured tool, ends in the same event: a request to read or write data.
Everpure makes this case in its account of trust in the agentic era. Agents act, they accumulate reach across systems, and they operate at machine speed, which necessitates real-time oversight. The argument follows: enforce policy in the data path, on every operation, and you no longer have to trust the agent or the application to behave. Everpure describes building this into its Data Stream product with NVIDIA's DOCA capabilities, including runtime visibility, zero-trust file access and network policy enforcement.
You do not need that vendor's stack to accept the architectural claim. If the control sits outside the path, it is always one step behind the action.
Trust becomes manageable when the signal and the threshold are written down
The practical work is choosing a small number of signals, setting the level each must hold before a system gets more scope, and naming the evidence that would satisfy an auditor.
| Signal | What it tells you | Threshold to hold before widening scope | Evidence that answers the question |
|---|---|---|---|
| Data quality and completeness | Whether inputs are fit for this decision | Every field in the decision path validated against an agreed domain definition | OvalEdge trust signal layer |
| Lineage coverage | Whether an output can be traced to its source | Full lineage from each input field to the decision record | OvalEdge; Galileo on model lineage |
| Model confidence | Whether the system knows when it is unsure | Low-confidence outputs routed to a person, not returned as answers | The AI Journal on confidence gates |
| Policy status | Whether controls held on the last request | No unresolved policy violation older than 24 hours | Everpure on inline enforcement |
| Human review rate | Whether oversight is real or nominal | A weekly random sample of high-impact decisions, with reason codes recorded | Galileo's discussion with IBM's Maryam Ashoori |
| Incident disclosure | Whether an outsider can check the record | A searchable register of adverse incidents, published | GOV.UK's account of the commission |
Signals drawn from OvalEdge's trust framework, The AI Journal's account of confidence gates and Galileo's discussion of lineage; the thresholds are illustrative.
The hard part is not choosing the signals. It is writing down the level each must hold before a system earns more scope, and naming the evidence that would settle the question. A threshold nobody wrote down is not a control.
Staged authorisation is a release gate, and a release gate can be configured
The learner-plate idea translates directly into deployment machinery. A model earns scope in steps, and each step has an exit condition agreed before the step begins.
release_gate: claims_triage_assistant
stages:
- name: shadow
scope: read_only
exit: 2000 consecutive decisions with no unresolved policy violation
- name: supervised
scope: recommend_with_human_signoff
exit: 30 days at or above the agreed concordance rate with reviewers
- name: general
scope: recommend_with_sampled_review
review: weekly random sample of 50 decisions
monitoring:
drift_check: daily
incident_register: published
owner: named_accountable_executive
Two things make this more than a process diagram. The exit conditions are measurable before the stage starts, and the review continues after full authorisation, which is the part most programmes skip.
Confidence gates belong in the same configuration. The AI Journal's account of applied data trust describes routing low-confidence answers to a person for review, and notes that the reviewer needs enough context to see why the system was uncertain. That is a design requirement.
How to achieve compliance within the UK's principles-based AI regulatory model
The UK has no single AI act. OneAdvanced's guide to responsible AI describes a principles-based, regulator-led model, with duties arriving through UK GDPR, the Data Protection Act 2018, ICO guidance including the forthcoming code on AI and automated decision-making, and voluntary standards such as ISO/IEC 42001. A firm that can evidence monitoring, lineage and accountability is already most of the way to compliance, and it can show the same evidence to a customer.
techUK's AI programme makes the organisational version of the point: realising the benefits needs more than technical capability, and depends on governance approaches, adoption pathways and collaboration across the whole company. Trust is not a workstream owned by the AI team.
Why current governance models fail to address the risk of autonomous agents
Agents are being deployed faster than the oversight around them.
The AI Journal cites a Deloitte 2026 enterprise AI survey finding that only about one in five companies has a mature governance model for autonomous agents. That is the number to take to a board. An agent with broad reach and machine-speed action is a different risk class from a chatbot, and the governance most firms have in place was built for the chatbot.
Galileo's discussion of transparency and governance is useful here because it treats observability and governance as preconditions for production. Its point about model lineage, knowing which version of which model was trained on which version of which data, is the minimum needed to answer a question six months after an incident. The Health Foundation's Jennifer Dixon made the equivalent point about the NHS, noting that the harder test is whether the service has the capacity, skills and systems to implement and monitor these tools at the scale required. Substitute your own organisation for the NHS and the sentence still holds.
What to do this quarter
Four moves, in order, none of which requires a new platform.
Name an accountable owner and a review date for every model in production, including the ones bought as a service, because an unowned system cannot be governed. Write the expansion criteria for each one before it gets more scope, and put those criteria in the release gate. Instrument the data path so policy is checked on every request, and log the refusals, since a control that never fires is probably not running. Publish an incident register internally, then decide what portion of it belongs in public, because the commission's argument is that disclosure is what converts compliance into trust.
The healthcare commission was asked a narrow question and answered a broad one. Its recommendations describe a system that grants trust in increments, keeps watching after the grant, and can take the grant back. Any enterprise deploying AI into a decision that affects a person's money, health, employment or legal position is in the same business, whether or not a regulator has said so yet.
Sources
- Building Trust into Enterprise AI for the Agentic Era
- Reframing Trust in AI
- What Is an Enterprise AI Trust Framework? Full Guide (2026)
- Trust & AI: Relational Trust in AI Transformation
- Responsible AI: Principles, Examples & Implementation Guide
- Artificial Intelligence programme
- Building Trust and Transparency in Enterprise AI
- Applying Data Trust in Enterprise AI
Source: Independent Commission led by NHS doctors sets out blueprint to accelerate safe AI adoption in healthcare - GOV.UK, GOV.UK





