Know what a tool is approved for
An AI system's regulatory clearance covers a specific use. Using it, or teaching people about it, outside that scope is an adopter decision with adopter risk attached, not something the clearance already covers.
My approach
Governance-first AI for regulated learning
AI can compress weeks of production into hours. It can also quietly introduce errors, bias and data exposure into training that people's safety, livelihoods or care depend on. My approach treats AI as a production accelerator that a human always signs off, never as a substitute for professional judgement. This is how I use it in the NHS, in CQC-aligned care and in a chartered professional body, without lowering the bar.
Where I sit
An adopter of AI in regulated learning, not a builder of AI tools
The UK's health and social care system draws a clear line between the organisations that develop AI products and the organisations that deploy them. I sit firmly on the adopter side. My job is not to build or validate an AI model. It is to decide, case by case, whether and how AI has a safe place in a piece of learning, and to keep a human accountable for everything that gets published.
- Human-in-the-loop by default. AI drafts, restructures and translates. A designer or subject matter expert reviews and signs off before anything reaches a learner. This mirrors the CQC's own position that AI should support professional judgement, not replace it.
- Privacy by design. Sensitive learner data and identifiable records do not go into AI tools. Where personalisation uses learner data at all, only the minimum necessary goes in, and it is governed like any other regulated data set.
- SME review as a hard gate. AI-generated content and assessment items are checked by a subject matter expert against professional and regulatory standards, in line with the direction Skills for Health has set out for compliance eLearning.
- Accessibility as compliance, not polish. Every AI-assisted module is built and checked to WCAG 2.2 AA, because UK public sector accessibility regulations apply to intranets and extranets as well as public-facing sites.
- Transparency and choice. Learners and stakeholders know where AI has been used in producing a module, and organisations keep the choice about how far AI is allowed into a given piece of learning, rather than that choice being made by default.
Human-in-the-loop
AI does the mechanical work. A person signs off the judgement calls.
The distinction I hold onto in every AI-enabled workflow is between mechanical work and judgement work. AI is very good at the first and unreliable at the second, so I never let it own the second.
What AI is allowed to do
Draft first-pass content from an approved outline, restructure existing material, translate text while preserving instructional intent, and suggest assessment items from a source document. All of it starts as a draft, never a final version.
What a person always does
Checks factual accuracy against the source, confirms tone and terminology fit the profession or clinical context, verifies that assessment items test the right thing, and takes accountability for publishing. This gate is never skipped, whatever the deadline.
The CQC has set out its own expectations for AI in health and social care in similar terms: it should support the people delivering care, not substitute for their judgement. I hold learning design to the same standard, and I have written more on what that looks like in practice in AI to support, not replace. If AI has touched a piece of content, a named person still decided it was fit to publish.
The framework I work within
The NHS AI and Digital Regulations Service, and the adopter's job
The NHS AI and Digital Regulations Service brings together guidance from NICE, the MHRA, the CQC and the HRA into a single reference point for anyone using AI in health and care. It is aimed mainly at people developing and regulating AI products, but the developer-versus-adopter distinction it draws is exactly the one I use to frame my own responsibilities as someone deploying AI-enabled learning inside regulated organisations.
Keep oversight local and ongoing
A regulatory mark tells you a product met a bar at a point in time. It does not tell you it is safe in your context, forever. Local governance, local review, and a person accountable in the room.
Build the same discipline into learning
If clinicians and carers are expected to apply this discipline to AI in patient care, the training that teaches them to do it has to be built to the same standard: governed, reviewed, and honest about where AI sits in the pipeline.
The AI Fundamentals for Imaging and Oncology module teaches clinicians this exact distinction between regulatory approval and clinical validation. It is the same discipline I apply to my own use of AI in design.
Where this shows up in the work
Governance is not a slide. It is a pipeline decision.
The CIPS AI localisation pipeline
Localising a course used to mean rebuilding it by hand for every language. I built a governed pipeline that lets AI handle translation and restructuring in context, with every module still going through a human review gate for accuracy, tone and cultural fit before it is published. Speed came from the pipeline. Trust came from never skipping the sign-off.
AI Fundamentals for Imaging and Oncology
A live Articulate Rise 360 module for radiologists and oncologists that teaches AI literacy for clinical practice: how these tools produce their outputs, why regulatory approval is not the same as clinical validation, and how to critically appraise an AI result before it touches patient care.
Accessibility does not soften
WCAG 2.2 AA applies whether a human or an AI drafted the content
UK public sector accessibility regulations require websites and apps, including intranets and extranets, to meet WCAG 2.2 AA. That requirement does not have an AI exemption. If anything, AI-generated first drafts need more scrutiny at this stage, not less, because a model has no concept of a screen reader, a caption or a colour contrast ratio.
Every module I produce, AI-assisted or not, is checked for keyboard operability, sufficient contrast, captions and transcripts, logical focus order and plain language before it goes near a learner. See how this plays out in practice on the accessibility page, and how it applies specifically to compliance content on the compliance training insight.
Common questions
Questions I get asked about AI and governance
How do you use AI in learning without losing quality control?
I only ever give AI the mechanical work: drafting, first-pass translation, restructuring. A designer or subject matter expert reviews and signs off everything before it is published, so accuracy, tone and instructional intent are never delegated to the model. I built the CIPS localisation pipeline exactly this way, and the human review gate is never skipped.
What does the NHS AI and Digital Regulations Service require of adopters?
The NHS AI and Digital Regulations Service brings together guidance from NICE, the MHRA, the CQC and the HRA, and it distinguishes developers of AI products from adopters who deploy AI within their own services. I sit on the adopter side, so I take responsibility for understanding what a tool is approved for, keeping human oversight in place, and never treating a regulatory mark as a substitute for local, ongoing judgement.
How do you keep learner data safe when using AI?
Privacy by design, not by exception. I keep sensitive learner data, performance records and identifiable information out of AI tools entirely. Where AI supports personalisation, I put in only the minimum data needed, and I handle it with the same care as any other regulated data set, not as something harmless because a model is involved.
Does AI-generated training still need accessibility compliance?
Yes, without exception. Public sector websites, intranets and extranets must meet WCAG 2.2 AA under UK accessibility regulations, and that requirement does not soften because AI produced the first draft. I check every AI-assisted module against WCAG 2.2 AA before it goes anywhere near a learner.
Sources and further reading
Get in touch
This is how I think about AI in learning.
If you would like to talk about any of it, or see more of my work, I would be glad to hear from you. Learning that changes behaviour, not just tick boxes, is still the only measure that matters.