Doctolib ships an AI agent that lets clinicians build their own note templates by chatting

The problem

Doctolib offers an AI scribe, the Consultation Assistant, that listens to a patient visit and drafts the clinical note. The time saving depends on the note reading as that doctor would have written it. Two cardiologists will not write alike, so every mismatch sends the doctor back to rewrite. Doctolib's product let doctors add instructions to note templates, but writing a good template took time and technical skill most practitioners did not have.

What they built

The Consultation Template Manager swaps a settings screen for a chat. A clinician says what kind of note template is needed, with its specialty and structure. The agent asks multiple-choice questions, drafts a template, and refines it as the clinician reacts, for example "add an allergies section". It can also restructure an old free-form template in one pass. Behind it, a small group of cooperating AI agents is coordinated by one that routes each request to an interviewer, a builder or a converter.

How they did it

The idea came from an internal hackathon at the start of March 2026. About three weeks later, the agent was live in production with external beta practitioners. The team stayed small: two full-time engineers (one software, one machine learning) with part-time help from product management, design and product marketing colleagues.

The speed came from an existing AI platform. It supplies an agent framework built on Google's Agent Development Kit, a gateway that lets Doctolib switch language model providers without changing application code, automated deployment, and built-in monitoring and testing. Doctolib's Medical Excellence committee reviewed the product before release, and staff tested an alpha in week one.

Before release, the team ran automated tests where a model played doctors with different specialties and goals, and other models scored whether the agent repeated questions, adapted to the specialty and produced a coherent template.

Results

Doctolib reports figures from the roughly 6,000 practitioners with access. Among users who began using the agent, 78% finish by saving a customized template. Doctolib says that about nine times out of ten the saved template carries minor edits or none. Roughly half of those who try the agent go on to use their own template in a real consultation summary, usually within a few days. The agent runs in France and Germany.

These are adoption measures from an early rollout. The post does not report time saved per note.

What to copy

Pick the step that stops people from using a feature well, and let a conversational agent do it for them. Reuse a shared platform for deployment and evaluation so the first release takes weeks, and track whether people keep what the agent produces.