Ask "Are we ready for CQC?", "What's missing in Caring?", "Draft me a Duty of Candour letter for incident #42." The Copilot answers from your own evidence stream — never invents — cites every claim back to the source record.
Every claim the Copilot makes points at a source record: an evidence event [E12], a complaint [C3], an incident [I7], an open risk [R2], a policy [P5]. Click any chip to open the underlying record in the dashboard.
The system prompt instructs the model to refuse questions it can't answer from the supplied context. Never invents. Never guesses. If the answer isn't in your evidence stream + records, the Copilot says so plainly.
Every claim references a bracketed source id (E1, C3, I7, R2, P5) that maps back to the underlying record. The UI renders these as chips. Click any chip to open the source — the user verifies before acting.
The system prompt explicitly forbids: clinical decisions, legal advice, regulator-facing statements without human review. For high-stakes outputs (DoC letters, statutory notifications) a human always reviews before send.
Outstanding services don't just capture — they change. Each governance huddle, the Copilot answers these three. Inadequate services can't.
Lists every change made last month because of an incident, complaint, or audit finding — with each change cited back to its trigger record. If the list is empty, you're capturing without improving.
For each change, surfaces the post-change metric: recurrence count, audit pass rate, training completion, complaint trend. Data, not vibes — Outstanding services prove the change embedded.
Cross-references the change list against training completions + read-receipted policy circulars. If clinicians haven't acknowledged the change, you've changed paperwork, not practice.
Returns a readiness score (0-100), per-KLOE evidence density, weak areas, and the top 3 actions that would boost readiness fastest. Sourced from your live evidence stream.
Surfaces the under-evidenced KLOE + tells you specifically what categories of evidence are thin (e.g. "you have 3 complaint outcomes but 0 dignity-in-care training completions this quarter").
Generates a Reg 20-compliant notification letter from the incident summary. Apology without admitting legal liability. Plain English. Under 400 words. Reviewed and edited before send.
Pattern-matches across the recent complaints + incidents corpus. Surfaces themes you can lift directly into the next governance committee report.
Recommends specific tracker actions that would shift the Effective KLOE — typically completing a paused audit cycle, recording mandatory training catch-ups, or surfacing outcome data.
Ranks the open risk register by combined likelihood × impact × age, plus surfaces any risks whose evidence base contradicts their current rating.
The Anthropic adapter is the platform default. Best-in-class reasoning over compliance corpora. No training on your data per Anthropic's enterprise terms.
Adapter pattern means one config change. Use Azure OpenAI if your procurement requires Microsoft. Use local Ollama if you need zero third-party data egress. Per-tenant choice.
Our enterprise agreement with Anthropic excludes training on customer data. Per-tenant API key isolation. SAR export covers every Copilot interaction so subjects can request their full history.
Every Copilot call logs to HR_EVENTS with the question, the context counts, and the model+stop-reason. SOC 2 auditors get a complete record on request.
Group-tier feature. 14-day trial includes the Copilot at full power so you can interrogate your own evidence stream from day one.