Skip to content

How to use the Case Copilot for a case

When you need this: You're working a case on the Case Investigation page and want a fast first draft — a summary of the alert, the likely typology, a starter STR/SAR narrative, the counterparty network, or a risk read with a recommended next action — without leaving the workspace or hand-rolling the boilerplate. The Case Copilot (#499) is the governed, in-page way to get those DRAFTS.

Prereqs: A run loaded in the dashboard with at least one case. A GenAI backend selected via AML_AI_BACKEND — ollama, openai, or azure_openai for a live model; the default template gives canned scaffolding with no model call (useful offline). program.ai_audit_log (default hash_only) set per your privacy posture.

Time: ~2 min per case.

The Case Copilot is draft-generation only. Every reply is an explicitly-marked DRAFT the investigator reviews — it is never an auto-decision, never an auto-disposition, and it never touches the engine run path. It reuses the same governed assistant as the sidebar AI advisor (get_assistant + AssistantContext/AssistantReply), so every action is audited to the run's append-only ai_interactions.jsonl (event ai_case_copilot_action) with the backend, model, and confidence recorded. This is the SR-26-2 carve-out in practice: GenAI augments a human investigator who stays accountable for the decision — the model proposes, the person disposes.

Steps

1 · Choose a backend (governance-first)

The copilot inherits whatever backend the dashboard is configured for:

export AML_AI_BACKEND=ollama        # or: openai | azure_openai | template
  • template (default) — deterministic canned scaffolding, no model call, no PII leaves the process. Use it offline or to see the action shapes.
  • ollama — local model; PII transit stays on-box.
  • openai / azure_openai — hosted model; confirm program.ai_audit_log and your data-egress posture first (hash_only keeps only a SHA-256 of each draft on disk; full_text retains the draft text).

The backend status (provider + whether a key is set + the spec's ai_audit_log mode) is visible on the AI Assistant page.

2 · Open the Case Investigation page and select a case

Navigate to Case Investigation (page 4) and pick the case you're working. The Case Copilot panel sits in the sidebar, scoped to that case — it sees the same entity profile, alert details, and transaction context the page renders.

3 · Pick a Copilot action (or ask freeform)

Choose one of the structured actions:

  • Summarize — a concise read of the alert and the case so far.
  • Identify typology — the likely typology(ies) the activity matches.
  • Draft STR-SAR narrative — a starter regulatory narrative.
  • Counterparty network — the counterparties and how funds move between them.
  • Risk & recommended action — a risk read with a suggested next step.

Or type a freeform question for anything the structured actions don't cover.

4 · Review the DRAFT

The reply is labelled DRAFT and carries its confidence plus the backend/model that produced it. Read it critically: it's a starting point, not an answer. Confidence is an honesty signal, not a green light — a high-confidence draft still needs a human's eyes before anything is acted on.

5 · Edit and use it — the investigator decides

Take what's useful and discard the rest. For example, copy the Draft STR-SAR narrative into the case's narrative field as a seed, then edit it to match what you actually found. The disposition, the escalation, and the final narrative are your decision — the copilot never writes to the case state, never closes, escalates, or re-queues an alert.

Verify it worked

Two checks:

  1. The reply is marked DRAFT and shows a confidence value plus the backend/model name. If it doesn't, you're not looking at copilot output.
  2. The action is audited — tail the run's ai_interactions.jsonl and confirm a row with event ai_case_copilot_action, the backend/model, confidence, and (under hash_only) the SHA-256 of the draft rather than its text:
jq -c 'select(.event == "ai_case_copilot_action")' .artifacts/run-*/ai_interactions.jsonl

No case in the ledger should have changed disposition as a result — the copilot is dashboard-only and never on the engine run path.

Common problems

Symptom Cause Fix
Replies look generic / canned AML_AI_BACKEND=template (the default) — no model is being called Set AML_AI_BACKEND=ollama (or openai/azure_openai) and restart the dashboard
Draft text isn't in ai_interactions.jsonl, only a hash program.ai_audit_log: hash_only (the privacy-safe default) Expected. Flip to full_text in the spec only after clearing it against your privacy posture — the spec change is itself the paper trail
Copilot panel is empty No case selected, or no run loaded Load a run and select a case on the Case Investigation page first
Worried the copilot might auto-close an alert It can't — it generates DRAFTS only, never touches case state or the engine run path Re-read the audit: every action is ai_case_copilot_action, the disposition stays whatever the investigator set

Next steps