When we designed Donna for hospitals, the hardest question was not what she should do. It was what she must never do.
Administration, not medicine
Donna handles appointments, reports and payments. She does not interpret symptoms, reports or medicines, and she does not suggest treatment. If a patient asks whether a result is normal, Donna tells them which reports are ready and offers to book a visit with a doctor who can explain them. That line is built into Donna’s instructions; it is not a setting someone can switch off.
Emergencies skip the AI entirely
Some messages should never wait for a language model. Emergency phrases are caught before the message reaches the model at all. Donna stops, and the conversation is handed to hospital staff. Hospitals can add their own phrases, in the languages their patients use.
Nothing happens without the patient
- Every booking, change and cancellation is confirmed by the patient before it runs.
- Reports and payment links are sent only after the patient gives consent.
- Anything Donna cannot handle goes to a staff inbox, where the team replies as the hospital.
The hospital stays in charge
Administrators set what Donna says and which emergency phrases she watches for, and they can test her by chatting as a patient and seeing exactly what she did for each reply. Our wider approach to security, privacy and responsible AI is on our governance and trust page.