A Conversational Clinical Copilot
A conversational copilot embedded in the notes workflow: clinicians ask in plain language, summarize free text in one step, and review AI-drafted note entries backed by transcript evidence and confidence scores.
jobs in one command bar
confidence shown on each suggestion
verify against the transcript
stays human-in-the-loop
The brief
Clinical documentation is where care quality quietly leaks away: clinicians juggle several tools, re-key the same details, and still leave notes incomplete or unsigned at the end of the day. The problem was never knowledge; it was the time and friction between a conversation and a finished, trustworthy note.
We built a conversational copilot embedded directly in the notes workflow. Instead of adding another screen to learn, it sits in a single command bar: clinicians ask in plain language, summarize free text in one step, and review AI-drafted note entries that arrive with the transcript evidence behind them. The goal was a colleague that drafts decisively, explains its reasoning, and always leaves the clinician in control.
A colleague that drafts decisively, explains its reasoning, and always leaves the final word to the clinician.
One command bar for the whole note
Rather than another screen to learn, the copilot lives in a single conversational bar. Ask in plain language, or launch a focused job from the feature menu.
Hi! I’m your smart assistant. Let’s get things done quickly.
Free text into a clean summary
Paste any block of clinical findings, history or observations and get a concise, structured abstract back, ready to edit, copy, or save.
The patient reports frequent shortness of breath, especially after minimal physical activity. No signs of fever or chest pain. Using prescribed inhalers regularly. Reports minor wheezing in the evening. No hospital visits or ER admissions in the past month. Patient denies smoking and follows a low-sodium diet.
Patient reports exertional shortness of breath and mild evening wheezing. No fever, chest pain, or recent hospital visits. Inhaler use is consistent. Denies smoking; follows a low-sodium diet.
Every suggestion shows its work
After a scribe session, draft note entries arrive paired with the exact transcript turns they came from, a plain-language reason, and a confidence score, so review takes seconds, not minutes.
You stay in control
Nothing reaches the chart on its own. The copilot prepares the work; the clinician decides what lands.
After each scribe session, the system prepares draft note entries for you. You can:
- Review them with linked transcript references.
- Apply all at once, or select the ones you want.
- Edit freely; you stay in control of your note.
- Review
each draft with its linked transcript references.
- Apply all, or select
apply everything at once, or pick only the entries you want.
- Edit freely
the note is always yours to shape before it’s signed.
What it runs on
The copilot is a thin, fast conversational layer over a clinician’s existing notes and transcripts, with every generated suggestion grounded in retrievable evidence rather than free-form inference.
Conversational Orchestration Layer
A lightweight intent layer turns natural language into safe, scoped actions:
- Intent Router. Classifies each message into a feature action (Start Note, Custom Note, Find Duplicates, Summarize) or a general query, then dispatches to the right handler.
- Streaming Responses. Token-streamed output keeps the assistant responsive while long summaries and drafts are generated.
- Mode Scoping. Context modes constrain the model to the active task, reducing drift and keeping prompts tight and predictable.
Evidence-Grounded Generation
Recommendations are retrieved and cited, never hallucinated, to earn clinician trust:
- Transcript Store. Scribe-session transcripts are indexed by speaker and timestamp so every turn is individually addressable.
- Retrieval-Augmented Drafting. Draft note entries are assembled from the most relevant transcript turns rather than the model’s own memory.
- Provenance & Confidence. Each suggestion carries a confidence score and a link back to the supporting transcript evidence for one-click verification.
Run it safely, improve it continuously
Trust & Safety Guardrails
The copilot is designed to defer, never to overstep:
- Human-in-the-Loop. Nothing reaches the chart automatically. Clinicians apply all, select individual entries, or edit before saving.
- Confidence Thresholding. Low-confidence suggestions are surfaced for closer review rather than presented as settled fact.
- No Silent Failures. When evidence is thin, the copilot says so instead of guessing, keeping the clinician aware of the gap.
Feedback & Continuous Improvement
Every interaction is an opportunity to make the assistant better:
- Inline Signals. Copy, regenerate, and thumbs up / down on each response capture lightweight quality signals in the flow of work.
- Override Learning. Clinician edits and rejections are aggregated (stripped of PHI) to track where drafts fall short and tune prompts over time.
Tuned for clinical relevance
To keep the copilot helpful without overstepping, prompts are tuned to draft only from what the clinician actually said and to make its reasoning visible.
System Role
A clinical documentation copilot: decisive on routine drafting, deferential on anything ambiguous.
Primary Data
Indexed scribe-session transcript turns plus the existing note context.
Constraint Guardrails
Summarize and draft only from provided transcript evidence; invent no clinical facts; defer when uncertain.
Strategy
Evidence-grounded reasoning that attaches a confidence score and a plain-language reason to every suggestion.
The outcome
The copilot reframes documentation from a chore into a conversation. By meeting clinicians in a single command bar, summarizing in one step, and drafting note entries backed by transcript evidence and confidence scores, it behaves less like autopilot and more like a trusted colleague, one that drafts quickly, shows its work, and always leaves the final word to the clinician.
Frequently asked questions
What is the conversational clinical copilot?
A copilot embedded directly in the notes workflow. Clinicians ask in plain language, summarize free text in one step, and review AI-drafted note entries that arrive with the transcript evidence and a confidence score behind them.
How does it avoid hallucinations?
Draft entries are retrieved from the actual scribe-session transcript turns (retrieval-augmented), each carrying a confidence score and a link to the supporting evidence. When evidence is thin, the copilot says so instead of guessing.
Does the copilot act autonomously?
No. Nothing reaches the chart automatically: clinicians apply all suggestions, select individual entries, or edit before saving. The copilot drafts; the clinician always has the final word.