Opinion · Data
AI in healthcare has a plumbing problem
By Element 45 · 2 October 2026

The models are improving faster than the records they depend on. Until health data can move, AI will keep answering from half the picture.
Most of the attention on AI in healthcare goes to the models: how accurate they are, which exam they passed, whether they outperform a specialist on a test set. Much less goes to the question that decides whether any of it helps a patient, which is whether the model can see the patient’s information at all.
In most of the health system, it cannot. We think this is the real constraint on AI in care, and that it will not be fixed by better models.
One patient, many records
A person with a long-term condition has a record at their general practice, another at each hospital they have attended, results held by pathology and imaging providers, and letters from specialists that arrive as scanned documents. Their pharmacist, physiotherapist and psychologist each keep their own. None of these systems was built to talk to the others, and in practice the patient carries the story between them.
Point an AI tool at any one of those records and it will produce a confident answer based on part of the picture. The summary will leave out the hospital admission. The medication check will miss the script written elsewhere.
A model cannot summarise a result it was never sent.
The standards exist
This is not a technical mystery. There are agreed standards for exchanging health information, and national infrastructure built to hold a shared record. The difficulty is that using them is optional for most of the people who would have to do the work.
A software vendor gains little from making it easy for a customer to leave. A clinic is not paid for the hour spent sending a clean summary to another provider. A hospital’s priorities end at its own walls. Everyone agrees sharing would be better, and nobody’s budget depends on it.

What would change it
Buyers have more power here than they use. A practice or health service choosing software can make open export a condition of the contract, and walk away from products that hold data hostage. Funders can pay for information that arrives in usable form. Patients can be given a copy of their own record that they are free to share.
Each of these is dull compared with a new model release. Together they decide whether AI in healthcare works from the whole record or keeps guessing from a fragment.
Where we stand
When we build a service, we treat the ability to send and receive information as part of the care, not an integration to add later. We would choose a modest tool that can see the whole patient over an impressive one that cannot.