The MHRA AI Airlock simulation workshops have put a spotlight on one of the hardest questions in digital health: where the boundary sits between wellness AI and AI that becomes a regulated medical device.
The discussion brought together regulators, developers, clinicians, academics, and patient representatives to explore how emerging AI products are interpreted when they begin to influence care, decisions, or user behaviour in more consequential ways.
A central theme was intended use. In medical device regulation, intended use affects whether a system qualifies as a medical device, how it is classified, what evidence is required, and how it must be monitored after deployment.
For conversational AI and large language model systems, intended use is rarely static. Products evolve, features expand, systems enter clinical workflows, and outputs begin to influence users or clinicians in new ways.
That means a system that begins in the wellness category can move toward medical device territory gradually, without a single obvious tipping point. This makes product strategy and governance decisions especially important.
Post-market surveillance was another major theme. Regulators increasingly expect organisations to show how they monitor AI systems after launch, how they detect risk over time, and how they respond when patterns emerge across many interactions.
For teams deploying conversational AI at scale, this is a practical challenge. Risks in language model systems often surface gradually across thousands of interactions rather than through one dramatic failure, which makes sustainable monitoring essential.
As AI becomes more embedded in care pathways, even subtle product changes can shift the regulatory position of a system. That is why behavioural, clinical, and governance thinking need to be built in before scale, not added later as cleanup work.