Skip to main content

Health organisations are accountable for the AI they buy and use, but there has been no shared, structured way to assess it. The AI Risk Profiler Method (AiRP) Method, developed with 40+ organisations, now has its own website. Here's what it is, how it works, and how to get involved.

A new home for the AiRP Method

The AI Risk Profiler Method (AiRP Method) now has its own website: airiskprofiler.org.

The AiRP Method is an open method for profiling the risks of AI in health and care. It has been shaped by more than 40 organisations, including patient groups, health authorities, universities, professional bodies, providers and industry, across several countries. ORCHA acts as secretariat for the AiRP Research Programme.

The new site brings everything into one place: the full method, short videos on each step, and the ways to help test it. We've given it a home of its own on purpose. The method belongs to the programme, and it's open for anyone to read, use and challenge.

Why it's needed

Health organisations are already responsible for the AI they buy and use. What they haven't had is a consistent way to look at an AI feature and answer three questions. What could go wrong? How bad would that be in this setting? Has enough been done about it?

Traditional assurance assumes a product is designed, tested, approved and then stays broadly the same. AI doesn't work like that:

  • Volume. AI features are arriving faster than anyone can review them.

  • Pace. Models change in weeks, not years.

  • Opacity. A learned system doesn't always behave the way its specification says it will.

  • Drift. Performance in real use can move away from performance in testing, over time and across different groups of people.

Regulation and guidance set the principles. The EU, UK and US each approach this differently, but underneath they ask the same things. What does the AI do, what can go wrong, what's in place to stop it, and who answers if it does? The missing piece has been a structured way to apply those principles to one product, in one setting.

What the AiRP Method is

The AiRP Method works out which AI-specific risks are active in an AI feature, how serious each one is in its clinical setting, and where the controls fall short. It's designed so that people who aren't AI specialists can apply it consistently.

The 6 steps

The first three steps describe the AI. The last three assess the risk.

  1. Clinical context. What's at stake clinically, and for whom?

  2. Computational function. What does the AI actually do?

  3. AI technique and pipeline. How does it do it, and on what data?

  4. Risk activation. Which of ten known AI risks are in play? These include hallucination, bias, drift, overconfidence and opacity.

  5. Controls. What's in place to reduce each risk?

  6. Gap analysis and findings. Where do uncontrolled risks remain?

 

Three design choices

One feature at a time. A product with three AI features gets three profiles, because each can fail in a different way.

Risks before controls. Risks are identified first, so a well-controlled product is never mistaken for a low-risk one.

Nothing is averaged. Every risk is reported on its own, including the ones that were checked and ruled out. There's no single score, because a single score can hide the one risk that matters.

 

What it isn't

The AiRP Method doesn't say a product is safe or approved, and it doesn't replace regulation. It builds on existing frameworks, and its results can be read across to the ones buyers and regulators already use, such as the EU AI Act, IMDRF categories and the NICE Evidence Standards Framework. In the UK, it's designed to sit alongside the NHS clinical safety standards DCB0129 and DCB0160.

 

What comes out

Each AI feature gets a profile in the same format: all ten risks shown, how well each active risk is controlled, and the supplier's own explanations beside the method's result. It ends in a numbered list of actions. For a supplier, that's a to-do list. For a health system, it's the questions to ask before signing.

 

Who it's for

Health systems and commissioners. See where the risk sits in a specific AI feature, in your setting. Get the questions to ask a supplier, grounded in evidence rather than marketing, and one structured view that works across different regulatory regimes.

Digital health suppliers. See your AI feature the way a careful buyer will, and get a clear list of actions to close the gaps before your buyers find them.

Clinicians, safety leads and patient representatives. You don't need to be an AI expert. The method is built to be applied by non-specialists, and your view on clinical risk, patients or governance is exactly what it needs.

 

Built in the open, and being tested now

The method was introduced in a methodology paper at HHAI 2026 in Brussels. It's now at version 3.3, and every change is published with a version history.

It's being tested in two ways:

  • Reviewers look at one step at a time: a short video, a reading page and a 20-minute survey. We want clinicians, safety and governance leads, researchers, patient representatives and anyone else with a view.
  • Suppliers run the profiler on one AI feature of a live or pilot product, flag where it doesn't fit, and get their own risk profile back.

Feedback goes to the advisory group, which decides what changes in the next release. Everyone who reviewed is told what changed, what didn't, and why. 

 

Get involved

There are three ways to take part at airiskprofiler.org:

  1. Review the method. Pick a step, watch a short video, and tell us where it's right and where it's wrong.
  2. Test it on your product. Run it on one AI feature and get your own risk profile back.
  3. Follow the programme. Hear about new method releases, papers and events.

Questions? Contact us or send us a note to hello@orchahealth.com and we'd be happy to chat!