Treat each recertification as a change to the investigator's working rules. Every 90 days the case-building agent is recertified on your own alerts, and what it may draft, close or escalate can change. Investigators need to know which alert types it now handles differently, which narratives they must rewrite, and what they own when the agent drops a depth.
What does recertification move in an AML queue?
An AML investigator's day is alerts: transaction monitoring hits, sanctions name matches, referrals from front-line staff. A governed agent typically takes three jobs. It assembles the case file from the core ledger, the customer record and prior alerts at depth 1. It drafts the case narrative and a proposed disposition at depth 2. And in some firms, for a narrow and well-evidenced class of alert, it may close obvious false positives at depth 3 within the customer's own limits.
Recertification re-runs the agent's cases on real alerts, covering each of those jobs, and scores them against a floor, gate by gate. Fresh alerts matter most here, because criminal typologies move and an agent tuned on last quarter's patterns can fall behind. A capability that fails is not trimmed; the agent keeps only the highest depth the run cleared. If it cleared none, the agent drops to depth 1 the same day, and every alert arrives with a file but no draft.
Some things never move. Deciding to file a suspicious activity report stays with a person. So does any account restriction, which is a depth 4 action and held for a named approver. Contacting the customer about an open investigation is refused, because tipping off is a red line.
Who runs the cycle?
The head of financial crime operations owns investigator readiness. The builder who certifies the agent owns the recertification. The money laundering reporting officer, or equivalent, should see the change note before it takes effect, because a change to auto-closure changes the firm's risk-based approach in practice.
A 90-day plan for AML investigators
- Agent, depth 1: produce the cycle report. Alerts handled, dispositions drafted, auto-closures, escalations and refusals, each with the bound that decided it.
- Person (quality lead): sample auto-closures. Re-open a sample of alerts the agent closed and confirm each closure would survive review.
- Person (builder): recertify. Replay against recent alerts, including the ones the quality sample challenged.
- Person (builder and head of operations): write the change note. Which alert types the agent may close, which it now only drafts, and any new escalation rule.
- Person (reporting officer): sign the note if auto-closure scope changed.
- Agent, depth 2: draft practice cases. Five anonymised alerts that show the change, checked by the quality lead.
- Person (head of operations): run the briefing. Investigators work the cases aloud and explain where they would rewrite the narrative.
- Person (quality lead): sample the first two weeks. Focus on the alert types whose depth changed.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Case file | Gathers transactions, customer data and prior alerts | 1 | Judging relevance |
| Narrative | Drafts the narrative and proposed disposition | 2 | Rewriting and deciding |
| Low-risk closure | Closes a narrow alert class within bounds | 3 | Sampling every cycle |
| Account restriction | Drafts the restriction and holds it | 4, held | Approving or declining |
| Report filing | Drafts supporting text only | 2 | Deciding to file |
| Customer contact | Refuses during an open case | refused | Any contact decision |
What should investigators retrain on, and what can wait?
Retrain on three things each cycle. First, any change to auto-closure: if the agent no longer closes a class of alert, investigators get that volume back and must know the reason. Second, reading the narrative critically. A fluent narrative can omit a link between accounts that the agent never saw; investigators should check the sources listed against each fact, and note that customer explanations are declared, not verified. Third, the escalation rules, because a typology update often adds a new trigger.
Leave alone what did not change. Re-teaching the whole workflow every 90 days wastes the hour and teaches people to skim the change note.
The FATF Recommendations set a risk-based approach to customer due diligence and suspicious transaction reporting. An agent that drafts faster does not alter who forms the suspicion; the investigator and the reporting officer still do.
Which controls hold, and what evidence do they leave?
- Quality sample of auto-closures each cycle. Evidence: the sample list, outcomes and any reopened alerts.
- Signed change note. Evidence: the note with the reporting officer's sign-off when scope changed.
- Briefing attendance and check questions. Evidence: named records per investigator.
- Restrictions held for a named approver. Evidence: each held restriction and its approval in the append-only record.
- Practicum for new investigators approving restrictions. Evidence: the practicum log template for AML investigators.
Whether an investigator needs a certificate to approve those holds is a business choice, covered in do AML investigators need a certificate to approve AI agent actions. Pryme Intelligence Academy's Operator track is one way to provide that evidence, with module tests set in the learner's own industry and role.
Signals worth measuring
Track the share of agent narratives rewritten substantially, the reopen count from the quality sample, and the time from alert to disposition before and after each cycle. A rise in rewrites after a recertification usually means the run passed but investigators noticed something it did not test. Feed those cases into the next cycle.
Also count alerts that waited because no certified approver was on shift for a restriction. That is a staffing gap, not an agent problem.
Where does it slip?
It slips when auto-closure scope widens quietly. If the change note does not say so, investigators stop seeing a class of alert and never learn why.
It slips when investigators treat the draft disposition as the answer. The agent's draft is a starting point; the decision and its reasoning belong to the investigator, as are AML investigators accountable for an AI agent's mistakes sets out.
And it slips when the cycle collides with a regulatory review or a backlog sprint. Move the recertification, not the briefing. Teams that run collections or risk work face the same pattern; see keeping collections agents ready for recertification and the guide to AI literacy and certification for people who work with agents.
Questions people ask
Can the agent ever file a suspicious activity report itself?
No. It can draft supporting text, but forming the suspicion and deciding to file stay with people. Treat any design that automates filing as outside sensible bounds.
What happens to open cases if the agent drops to depth 1?
They stay open. Investigators receive the case files without drafts and work them as they did before the agent, using the fallback rota.
How often should auto-closures be sampled?
Every cycle, and more often after any change in scope. The sample is what tells you whether the agent's closures would hold up under review.