01

How an agent differs from a medical chatbot

A chatbot primarily answers questions. An agent operates inside a defined pathway: it requests the minimum context, consults approved sources, checks completeness, prepares a draft or creates a task for a professional. Sources, logs and stopping conditions belong to the workflow.

A medical agent is therefore a narrow digital assistant. Documentation, trial search and patient navigation are different jobs; combining them into an all-knowing virtual doctor increases risk and weakens product accountability.

  • the role and permitted task are defined before launch;
  • material conclusions are linked to sources;
  • stopping rules, escalation and human confirmation are explicit;
  • quality is measured on the actual workflow, not only on benchmark questions.
02

Where agents are already used

The most mature deployments cluster around documentation and administrative work: draft notes from encounters, reminders, collection of missing non-clinical information and exception queues for coordinators. Literature and clinical-trial search are advancing but still require expert review.

A 2026 evidence map reports that early adoption is more common in administrative settings while clinical use is expanding. Governance, auditability and clinician oversight remain central requirements.

03

What is not yet proven

A public demo, a high benchmark score or a vendor statement does not prove clinical benefit. Each use case needs its own measures: completed pathways, omissions, unsupported answers, professional time, escalation frequency and consequences of errors.

Autonomous diagnosis, prognosis and treatment selection deserve particular caution. They are outside this public prototype: AI structures information; authorized professionals make clinical decisions.

04

A practical test for a useful agent

A good agent does not try to answer at any cost. It recognizes missing data, shows the source, separates fact from inference and hands difficult cases to a person. Reliability in a narrow task matters more than the breadth of a demonstration.