Why every AI agent needs a control layer
AI agents are ready for regulated work. What decides whether they reach production is identity, policy, approval and evidence: the control layer.
When a clinician opens a shared record, the most dangerous thing on the screen can be what is not there. Reimea Exchange is designed around one idea: never hide where information came from, or that some of it is missing.
Many shared records try to produce one tidy list: one medication list, one problem list, one timeline. Tidy is reassuring. But merging hides two facts a clinician needs to know: which organisation recorded each item and when, and which sources were not available at the moment they looked.
An out-of-hours clinician at 2am who sees no recent admissions needs to know whether that means there were none, or that the acute trust's feed was down. Those are very different situations, and a merged view can make them look the same.
Exchange keeps the originating organisation, the date it was recorded and its status with every item. Items from different sources are shown side by side, not silently reconciled.
If a source is unavailable, out of date or not participating, the view says so, in the same place the data would have been. Missing coverage is shown, never implied away.
Exchange access is designed around explicit permission, identity review and a stated purpose of use. It is a separate product from Meridian: a Meridian login never implies access to Exchange records.
A shared record should help a clinician see what they know, and also what they do not know.
As agents start to read records and draft summaries, provenance becomes more important, not less. An agent that summarises a merged list cannot tell you where a fact came from. An agent working from a source-attributed record can cite the source, and show the reviewer which sources it could not see.
The same principle runs through Reimea's control layer. Evidence sits beside every claim, whether the claim is made by a person, a record or an agent.
Taking Exchange into live service needs work that only partners and assurance bodies can complete: production identity, national connectivity, a clinical safety case signed by a clinical safety officer under DCB0129, and each deploying organisation's own DCB0160 assessment. We will publish progress here as each step is evidenced.
If you are working on shared care records in an integrated care system and want to influence the design, we would like to hear from you.
AI agents are ready for regulated work. What decides whether they reach production is identity, policy, approval and evidence: the control layer.
"Human in the loop" is easy to promise. How to design approval of AI agent actions that staff can sustain and auditors can trust.