Mercurie Concierge — your AI clinical assistant
A chat-style assistant that lives in the top-right of every provider page. Ask about your day, your patients, or recent clinical literature — the concierge queries your chart with the same access checks as the rest of the app, drafts referrals or notes, and never writes to the record without your approval.
Open the concierge
Click the sparkles icon in the top-right of any provider page. A panel slides in from the right edge with a chat box at the bottom and a transcript area above. The concierge stays open across navigation — it remembers the conversation as long as you're on the platform.
Click New conversation at the top of the panel any time you want a fresh context — handy between patients, or when you want to start from scratch on a different question.
Ask in plain language
Type your question and press Enter (Shift+Enter for a new line). The concierge picks the right tools to answer it — no slash commands, no menus to learn. Examples that work today:
- "Who am I seeing today?"
- "Summarise my appointment requests."
- "What was the last note for Priya Sharma?"
- "Pull up John's recent vitals and labs."
- "Draft a referral for Mrs. Khan to a cardiologist — uncontrolled BP despite triple therapy."
- "Recent literature on persistent dry cough in adults?"
Tools the concierge can use
Each tool runs with your session — the same row-level security gates that guard every other provider action apply unchanged. The concierge can't read records you wouldn't be allowed to read manually.
- Today's schedule — your appointments for any date with patient names + reason.
- Patient search — fuzzy name match across patients in your organisation.
- Patient summary — demographics + active conditions + current medications + allergies + recent vitals + recent labs.
- Recent records list — latest Compositions / DocumentReferences / Observations for one patient.
- Open a specific note — full title + sections of any Composition.
- PubMed search — top abstracts from NCBI's PubMed for evidence questions.
- Web search — when enabled at the org level (see step 5 below), the concierge can search the public web for clinical guidelines.
Each tool the concierge invokes shows up as a small chip in the chat bubble. Click a chip to expand it and see exactly what the concierge asked the tool and what it got back — useful for verifying a clinical claim before acting on it.
Drafts (referrals, notes, prescriptions)
Ask the concierge to draft a referral letter, a clinical note, or a prescription and it produces a structured draft inside the chat with an Approve card. You decide what happens next:
- Approve & send (referrals) — opens an in-chat recipient picker (a clinic on Mercurie OR an off-platform email). On submit the referral goes through the same send path the existing Referrals UI uses, so audit, consent, and notifications work identically.
- Approve (prescriptions, notes) — saves as a draft Composition with status preliminary. Open the draft in your Drafts inbox to review, edit if needed, and click Finalize when you're ready. Patients only see the document once it's finalised.
- Copy — copies the rendered text to your clipboard so you can paste it into another tool or another window of Mercurie.
- Discard — drops the draft. The chat keeps the conversation history but the registered draft is gone.
Drafts auto-expire 30 minutes after the concierge created them. Approve, copy, or discard within that window. Once approved as a Composition draft, the document lives in your chart indefinitely under preliminary status.
Schedule X drugs (opioids, stimulants, benzodiazepines, ketamine, codeine, etc.) are non-negotiable: the concierge refuses to draft a prescription containing any of them. Write those manually so the medico-legal authorship stays unambiguous.
How the concierge is powered
The concierge runs on an enterprise-grade large language model hosted by our cloud provider. Cost is metered per-message under your existing AI budget. (The processors we use are listed on our Terms page.)
Web search is opt-in (off by default). When on, the concierge can search the public web for clinical guidelines and surface citations. Patient identifiers are never sent in a search query — the concierge abstracts the question first ("recent guidelines on persistent dry cough" rather than the patient's name).
Privacy & safety
Two design rules the concierge can't break, no matter how it's prompted:
- Same access as you. Every tool runs under your session. The concierge cannot read a patient's chart you wouldn't be allowed to read in the regular app — the database row-level security refuses cross-organisation reads silently.
- Reads anything, writes only with your approval. The only way the concierge writes to the record is if you click Approve on a draft card. Even then, the actual write goes through the same handler the manual UI uses, so audit + consent + notification all run identically.
Every action the concierge takes on your behalf carries an agent run id in the audit log so an admin can trace exactly what the concierge read or wrote during any single conversation. Your name remains the recorded actor — the concierge is a tool you used, not a separate identity.
Tip: The concierge is best at the structured questions ("who's on my schedule," "what's on Sarah's chart," "draft a referral"). It's not a substitute for clinical judgement — for diagnostic decisions, treat the concierge's output the way you'd treat a useful junior who hands you a summary, not the way you'd treat a senior who hands you a verdict.
Found a hallucination? The expander on each tool chip shows what the concierge actually asked the tool and what it got back, which usually surfaces the issue. If you suspect a real bug, share the conversation through Account → Support so we can investigate.