Choosing

Can a clinic let an AI assistant near patient data?

The question every clinic asks first, and the one most vendors answer too confidently. The honest answer has a shape: some of a patient's message can be masked automatically, some cannot, and the difference decides what you are allowed to put the assistant in front of.

What is masked before anything leaves, and what is not

Structured identifiers are replaced with tokens at intake, before any model is called: national ID numbers, telephone numbers, card numbers, medical record numbers. The originals are held encrypted and the replacement is reversible by the clinic, so staff still see the real values. Names written in free text are NOT detected automatically. That is a real limit, not a footnote: a message saying "my mother Sarah has chest pain" carries a name through.

Where the data goes, and how to ask about it

Processing involves sub-processors, and the only honest things to say are which categories of them there are, that they are bound by data-processing agreements, and that a specific hosting jurisdiction can be agreed per clinic on request. A vendor who answers "your data never leaves" is either not using a model at all or is not telling you the truth. Ask for the sub-processor list, not a reassurance.

The safety that matters most is refusal, not encryption

Encryption protects the message in transit and at rest. It does nothing about the real clinical risk, which is an assistant that answers a question it should have escalated. Emergency symptoms, whether a treatment suits a particular patient, dosage, and anything about a named colleague or another patient must be refused and routed to a person. A clinic should require that behaviour in writing and test it before going live.

Three questions worth more than any certificate

What exactly is masked, and what is not. Which sub-processors receive the text, and under what agreement. What the assistant does when it does not know - verbatim, with an example. A vendor who answers all three plainly is telling you how it fails, and that is the only information a clinic can act on.

Frequently asked questions

Is it safe to let an AI assistant handle patient messages?
Only for published, non-clinical questions - opening hours, preparation instructions, what a procedure involves, how to get a result. Anything that is a clinical judgement must be refused and passed to a person. The safety that matters is what it refuses, not what it encrypts.
What patient information is masked before the model sees it?
Structured identifiers: national ID numbers, telephone numbers, card numbers and medical record numbers are replaced with tokens at intake, reversibly, with the originals held encrypted. Names written in free text are not detected automatically, so a message naming a relative carries that name through.
Does patient data leave the clinic?
Yes - processing involves sub-processors bound by data-processing agreements, and a specific hosting jurisdiction can be agreed per clinic on request. Any vendor claiming nothing leaves is either not using a model or not being straight with you. Ask for the sub-processor list.
Can the assistant give medical advice if we configure it to?
No, and a clinic should refuse a vendor who offers it. Emergency symptoms, suitability of a treatment for a particular patient, and dosage must be escalated to a clinician. A confident answer given remotely reads as clearance and the clinic owns the outcome.
What should we ask a vendor before signing?
What exactly is masked and what is not, which sub-processors receive the text and under what agreement, and what the assistant says when it does not know - word for word. A vendor who answers all three plainly is describing how it fails, which is the only part you can act on.

Get in touch

Leave your details and we will reply. You do not need a sales call to get an answer.