AI for healthcare
Search AI for healthcare and you get diagnosis, drug discovery and a photograph of a robot in scrubs. Run a clinic and your problem is that the phone rings during consultations, referral letters take four days to leave the building, and a third of the diary is spent confirming that people are still coming. The interesting question is not whether a machine can read a scan. It is which parts of the week around the care can be handed over without anyone getting hurt.
Three tiers, and only one of them is a software decision
Sorting the work this way is most of the job, because the tiers have completely different rules attached.
- Clinical judgement. Diagnosis, triage, dosage, whether something can wait. In the EU this is regulated as a medical device the moment software influences it, and that is a certification project, not an automation project. I will say so rather than sell you around it.
- Clinical documentation. Notes, letters, summaries. A machine drafts, a clinician signs. The gain is real and the accountability does not move — which is exactly why the signing step is not optional.
- Everything else. Booking, reminders, intake, insurance paperwork, chasing results, answering the phone. No clinical judgement is involved, the volume is large, and this is where the week actually goes.
What gets built in the third tier
- The phone answered every time, with a real slot offered and written into the diary, including out of hours.
- Intake completed before arrival: history, consent, insurance details, arriving as fields rather than as a clipboard someone retypes.
- Reminders and confirmations that stop the moment a patient replies, and escalate to a person when the reply is not a simple yes.
- Referral and insurance paperwork drafted from what is already in the record, then read and signed by the person accountable for it.
- A daily list of what did not happen: results not back, letters not sent, patients not confirmed. Short enough that someone actually reads it.
The rules that decide the shape of the build
In Spain and the rest of the EU, health data is a special category under GDPR. That is not a paragraph at the end of a proposal, it changes what gets built.
- Where the data physically sits, and a written answer to it. EU hosting for anything holding patient records, and a model that does not train on what you send it.
- The narrowest possible access: an automation that books appointments has no business reading clinical notes, and should not be able to.
- A log good enough to answer, months later, what the system did and on what basis. This is skipped in demos and needed the first time anyone complains.
- A clinician's signature on anything that ends up in a record. Not a checkbox — an actual person who read it.
When it is too early
- Records on paper, or in three systems that disagree. Joining them is the project; automation on top of a disagreement just repeats it faster.
- No named owner for the daily exception list. A system that stops and explains itself is worthless if nobody reads the explanation.
- A pending change of clinic software. Build after the move, not twice.
The parts this is assembled from
- Data platform — the joining problem done properly on another industry's data: several systems reduced to one true record per item, with an AI layer that answers from it instead of from four contradictory copies.
- Customer-chat assistant — the pattern that makes an assistant safe to point at the public: one model writes, a second independent one reads every answer and blocks anything the business has not authorised. That second model is what a clinic's version would be built on.
What clinics ask before they start
Can AI for healthcare make a diagnosis?
Software that influences a clinical decision is regulated as a medical device in the EU, and that is a certification route with its own timeline and cost. It is a real path and it is not the one most clinics are asking about. Everything on this page sits deliberately on the other side of that line.
Is patient data safe with this?
It is safe when three things are true: it stays in the EU, the model provider does not train on it, and each automation can only reach the narrow slice it needs. Those are build decisions, and they are made at the start because retrofitting them means building twice.
Can it write clinical notes?
It can draft them from a consultation and put them in front of the clinician to correct and sign. The draft saves the typing. The signature is what keeps accountability where it belongs, and it is not a formality to be automated away later.
What do clinics start with most often?
Answering and confirming, because the volume is high, the risk is low and the effect on the diary is visible within a fortnight. Documentation usually comes second, once people have seen the system be careful about the easy things.
Does it replace reception?
It absorbs the repetitive part — the calls at 21:00, the fourth confirmation, the form filled in twice — so reception spends the day on the patients who are actually in the building. That is the honest version, and it is the one worth paying for.
Bring one week of your clinic's admin
How many calls came in, how many were answered, how long a referral letter takes to leave the building, how many appointments were confirmed by hand. In the free hour I will tell you which of these can move this quarter, which is a regulated project, and which one is simply not worth automating in a clinic your size.
Book that hour →