Let the agent write, but only drafts, and only into its assigned case. It gathers evidence from approved sources, attaches each item with its provenance, builds the chronology and drafts the case summary. Investigators accept what goes into the record, status changes are held, and decisions are outside its role. Every action is logged under its own identity.
What is a case file, and who owns it?
A case file is the record of an investigation: the subject, the trigger, the evidence, the chronology, the analysis, the decision and the approvals. In an AML investigation, a fraud referral or a complaint, the file is what the firm will show an auditor, a supervisor or a court to explain what it did and why. Once a case is closed, a filed case should be complete and unchanged.
Investigators own the content of their cases; their managers approve decisions; the system owner controls permissions. Building a file is laborious: requesting statements, exporting transactions, pulling screening results, writing up a chronology. An agent can do most of it. This playbook assumes the connection itself is in place, as set out in connecting an agent to your case management system.
Which permissions should the agent hold, and at what depth?
- Person, investigator: assigns the case to the agent for building. The agent works only on cases assigned to it.
- Agent, depth 1: reads the trigger and the subject. The alert, referral or complaint, and the customer or counterparty records.
- Agent, depth 1: gathers evidence from approved sources. Transactions, account history, due diligence, screening results, prior cases, correspondence. Each source is a separate read permission.
- Agent, depth 2: attaches evidence as drafts. Each item labelled with source, extraction time and the reason it is relevant. The investigator accepts or removes.
- Agent, depth 2: drafts the chronology. Events in order, each linked to its evidence.
- Agent, depth 2: drafts the case summary, using your case template, such as the payment investigation case template for payment disputes.
- Agent, depth 2: proposes a status change, such as "ready for review". Held for the investigator.
- Person, investigator: completes the analysis and decides.
- Person, approver: approves the decision and closes the case.
| Capability | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Read assigned case | Reads trigger and subject | 1 | Assigning the case |
| Gather evidence | Reads approved sources | 1 | Adding sources |
| Attach | Adds labelled drafts | 2 | Accepting into the file |
| Chronology | Drafts linked timeline | 2 | Correcting it |
| Summary | Drafts from template | 2 | The analysis |
| Status | Proposes, held | 2 | Approving the change |
| Decide and close | Nothing | None | Investigator and approver |
What is recorded, and which controls apply?
Assignment-based access. The agent cannot browse cases; it works on those assigned. Evidence: every action references an assigned case.
Provenance on every attachment. An export from the core ledger, a screening result, a customer email: each is labelled with where it came from. Evidence: an investigator can see the strength of each piece. A customer's emailed explanation is not the same as a ledger record.
Drafts until accepted. Agent attachments, chronology and summary are labelled drafts. Evidence: the case history shows who wrote what and who accepted it.
Held status. Moving a case forward needs a person. Evidence: approver and time on every transition.
No edits after closure. Once closed, the file is locked. The agent cannot write to closed cases. Evidence: write attempts on closed cases refused and logged.
What should you measure?
- Investigator time per case, before and after.
- Agent attachments removed as irrelevant, which shows over-collection.
- Evidence the investigator had to add, which shows under-collection.
- Chronology corrections.
- QA findings on files the agent helped build, compared with others.
Where does it usually go wrong?
Over-collection. The agent attaches everything it can find. The file becomes harder to review and holds more personal data than the case needs. Set relevance rules and measure removals.
Chronologies with invented links. An agent can join events into a story that the evidence does not support. Every link in the chronology must point to evidence; an inference must be marked as one.
Agent drafts accepted wholesale. If investigators accept every draft unread, the file reflects the agent's view, not theirs. Sample cases for edits, and look closely at any investigator who never changes anything.
Evidence pulled after the cut-off. An investigation looks at a defined period. Evidence extracted later can include events after the decision point, which muddles the record of what the investigator knew when deciding. Each attachment should show both its extraction time and the period it covers, and the agent should flag anything outside the case period.
Mixed cases. Two subjects with similar names, and evidence for one lands in the other's file. Require identifiers, not names, for every evidence pull.
Tipping-off through linked actions. Building the file must never involve contacting the subject. For AML cases specifically, see whether an AI agent can investigate an AML case and preparing AML case investigations.
The broader stance is in AI compliance agents.
Where Pryme fits
In Pryme, each attachment carries its provenance class from the Pryme Context Protocol, so a file shows at a glance which items are sealed records and which are declared or external. Write access is depth 2 on assigned cases only; a status change is held for the named investigator, and an attempt to close a case is refused. Each verdict is filed in the append-only record.
Questions people ask
Can the agent work on several cases at once?
Yes, each under its own assignment. Its reads and writes should always reference the case, so evidence never crosses between files.
Should the agent see prior cases on the same subject?
Usually yes, as linked context, within its data scope. Prior cases are some of the most useful evidence, and some of the most sensitive.
Who is accountable for a file the agent helped build?
The investigator who accepted its contents and the approver who closed it. The agent's role is visible in the history, but accountability stays with people.