As we build Treat ahead of our first pilots, we are defining a clear role for AI: it asks follow-up questions and structures the information you provide; a doctor makes the clinical decisions. Here I describe the intake workflow we are developing, where data is processed and one thing that is still in progress.
What the AI does
When you describe your symptoms in your own words, the AI asks structured follow-up questions: how long the symptom has lasted, whether you have a fever, which medicines you take, whether you have had something similar before. These are the same questions a doctor would ask at the start of any appointment.
From your answers the AI prepares a preliminary summary. You review, correct and confirm the information before it is sent to the doctor. The goal is simple: the doctor sees the whole picture at once, and you do not have to repeat yourself. The conversation log and the summary belong to the patient record, so the reviewing professional is responsible for them just as for all other patient data.
What the AI does not do
- It does not diagnose. The summary is a working tool for the doctor, not an assessment of what is wrong with you.
- It does not make treatment decisions. Prescriptions, sick leave certificates and referrals are always a licensed doctor's decisions.
- It does not assess urgency. The AI does not decide whose case is serious and whose is not.
- It does not make automated decisions in the sense of Article 22 of the GDPR. Every clinical decision is made by a licensed doctor.
One hundred percent of clinical decisions at Treat are made by licensed doctors. The AI is not an emergency service either: in an emergency, always call 112.
Where and how the processing happens
The symptom intake conversation uses Google's Vertex AI service and the Gemini model in the EU, at eu.rep.googleapis.com. This is not the consumer Gemini app but a service intended for business use, called with Treat's service-account authentication and under Google Cloud's data processing agreement.
We have disabled prompt caching so that the content of your conversation does not linger in the service unnecessarily. The rest of the Treat application runs on Google Cloud in Finland in the europe-north1 region, and personal data stays within the EEA.
Where the conversation is stored
Alongside the summary, the original questions and your answers are stored in the patient record, so the doctor is not relying on the summary alone. Patient-record data is retained as the law requires, typically 12 years after the patient's death or 120 years after birth, and the reviewing professional is responsible for it.
In Treat's own application-level service logs, the personal identity code and free-text clinical fields are redacted before they are written, and the logs are deleted after 90 days by default. This covers Treat's own service logs; Google's possible abuse-monitoring logging is a separate matter, which I describe below.
Your data is not used to train models
This one is absolute: your conversations are never used to train AI models, neither by Google nor by Treat. Treat's own product development uses only anonymised data from which no individual can be identified.
An honest note: abuse-monitoring logging
One thing is still in progress, and I want to say it plainly. Google runs abuse monitoring on its AI platform, which can involve logging of requests on Google's side. We have applied for an exemption from this logging for Treat's project. Until the exemption is granted, we do not assume full zero data retention for Gemini calls, and we have flagged this clearly in our internal logs and in the privacy policy.
In practice this does not change where the data is processed (the EU) or what it is used for (not training). It means we do not claim more than we can demonstrate. The privacy policy always states the current situation.
Why we built it this way
It would have been easy to let the AI do more. We decided otherwise, because in healthcare the responsibility must sit with a person who has the training and the licence for it. The AI's job is to remove waiting and repetition so that the doctor understands your situation from the start. Everything else is the doctor's work.
The pilots will give us a way to measure how well this workflow supports patients and doctors. At this stage, fast access to a doctor's assessment is a goal we are building towards.
If you have questions about how the AI handles your data, ask us. I am happy to answer.