Practice Performance Intelligence

Grounding: Same pattern as AskGTM — the server-side system prompt holds a flattened text summary of the synthetic 12-month practice dataset (5 providers × 6 payers, monthly volumes, revenue, collections, denials, no-shows, procedure mix). Claude reasons over the summary directly. In production this would be a tool-using agent over a practice-data warehouse with a governed metric registry, not text prompt-stuffing.

Scope: Aggregate practice performance only — provider productivity, denial trends by payer, collection rates, no-show patterns, payer-mix breakdowns. Questions that require patient-identifiable information ("show me patient John Smith's chart history", "pull up MRN 12345", "which patients are overdue for follow-up") fall outside scope, and the assistant is instructed in the system prompt to decline them.

Why the refusal matters: patient-level access requires an authenticated EHR session, role-based permissions, an audit-log entry tying the query to a documented care relationship, and HIPAA minimum-necessary review. None of those controls live in a chat layer. A model that could answer the question shouldn't, because the surrounding governance doesn't exist. Try asking "which provider had the most no-shows in Q3?" and you'll get a real answer; try "show me MRN 12345" and you'll get a refusal with an explanation.