Skip to main content

Clinical Workflow Copilot

Clinical Corvus is a workflow copilot, not a chatbot. It helps synthesize, draft, and check — but does not execute for you. Every output is a clinician-reviewed draft.

Check-to-Accept

The core pattern is:

  • Draft, don't execute: the system proposes items (text, action list, handoff)
  • Explicit control: you accept, edit, or discard each suggestion
  • Human review in the flow: this reduces automation bias

When the copilot is most useful

  • Building a quick picture before rounds
  • Turning messy text into a short, usable summary
  • Generating a consistent I-PASS handoff
  • Requesting evidence when the case requires a guideline or literature check
  • Analyzing labs with deterministic interpreters
  • Expanding differential diagnosis with evidence research

Quick Agent Actions

ActionDescription
ICU OverviewSummary + red flags + daily goals
Lab AnalysisLab test interpretation
Quick ResearchOn-demand evidence with citations
Expand DifferentialLimited differential expansion
Diagnostic TimeoutDeliberate pause to reconsider
Deep ResearchIterative autonomous research
Generate PlanTherapeutic plan suggestion

Research actions (Quick Research, Deep Research) show real-time progress because they can take 30 seconds to 3 minutes depending on provider latency.

Traceability

Each agent response includes traceability:

  • Citations: title, URL, date, journal for each source
  • Admitted/rejected evidence: the system records which sources were used and which were excluded and why
  • Quality badge: indicating whether the response has adequate evidence ("verified") or is useful but with limited evidence ("helper")
  • Severity scores: NEWS2 and SOFA with delta when available

Agent Team

A team of specialized agents operates behind the scenes:

Clinical Discussion Agent (CDA)

The primary agent. Handles reasoning, summarizes the case, and decides when to invoke research. Uses a structured question profile to classify each request.

Clinical Research Agent (CRA)

When the CDA identifies an evidence need, it delegates to the CRA. The CRA searches PubMed, Europe PMC, OpenAlex, Lens and other sources, synthesizes a referenced brief, and returns to the CDA.

Compass Controller

Orchestrates the CDA reasoning loop: Plan → Verify → Pivot → Stop. Controls token budget and decides when there's enough evidence to answer.

Verifier

Before any response is shown, the verifier checks if citations are valid, claims are evidence-backed, and risks are stated. If verification fails, agents correct the response.

Response States

StateMeaningClinician Action
ReadyResponse with adequate evidenceReview and accept/edit
PartialPartial response with stated uncertaintyReview carefully, supplement
InsufficientInsufficient evidence to answerSeek more information
BlockedRequest out of beta scopeReformulate or use another resource
ReviewRequires mandatory clinical reviewReview before use

Boundaries

The copilot operates with bounded support:

  • Supported: clinical question with context, differential synthesis, plan drafting, evidence retrieval, lab analysis
  • Warned: questions with insufficient evidence — the system states uncertainty
  • Blocked: requests for autonomous diagnosis, automatic treatment planning, unrestricted advice

Clinician review is part of normal workflow, not an exception.