Run it on live alerts the agent has already prepared, with a senior investigator deciding each one independently and a log of both decisions. Over ten working days, the investigator should get at least 95 dispositions in every 100 right and miss no red line, such as closing an alert with clear indicators or tipping off a customer. Nothing they decide is final until the senior reviewer agrees.
What is the practicum, and who owns it?
The practicum tests whether an investigator can work the agent's output safely: read a case file, challenge a draft narrative, decide close or escalate, and know when an account restriction is warranted. A written exam can test the rules; only live alerts test judgement under volume.
The head of financial crime operations owns the practicum. A senior investigator with a current certificate supervises. The money laundering reporting officer, or equivalent, sets the red lines and sees any practicum case that leads towards a suspicious activity report. The investigator owns their reasoning, written down for every alert.
How do you build the queue?
AML alerts take longer than most review work, so volume is lower: eight to twelve decisions a day gives around 100 across the fortnight. Within that, the queue should cover:
- transaction monitoring alerts with a draft disposition to close;
- alerts the agent drafted for escalation;
- partial sanctions or screening matches, held for a person;
- proposed account restrictions, which are depth 4 and held;
- cases where the agent refused to contact the customer during an open investigation;
- at least a few alerts where the draft narrative omits a link the investigator should find.
The supervisor seeds anonymised past cases for any type the live queue does not supply. A practicum without a single true positive tests nothing that matters.
For example, one seeded case might show a small business receiving many round-sum transfers from unrelated individuals and paying them out the same day to a single overseas account. The agent's draft narrative describes the volumes accurately but proposes closure because the business's declared activity is "import trading". The investigator should notice that the declared activity is unverified and the pattern fits a known typology, and escalate.
The steps
- Person (head of operations): brief the investigator on the bar, the red lines and the log.
- Agent, depth 1: route alerts to the practicum queue, tagged by type.
- Agent, depth 2: present each case file with the draft narrative, proposed disposition, deciding bound and sources, each with its provenance class.
- Person (investigator): decide and write the rationale: close, escalate, recommend a report or request a restriction.
- Person (senior investigator): decide independently, then compare.
- Person (senior investigator): action the agreed outcome. Restrictions and closures happen only on the senior's authority during the practicum.
- Agent, depth 1: log both decisions, the match and any red-line flag.
- Person (senior investigator, then assessor): review the log after day ten.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Routing | Sends alerts to the practicum queue | 1 | Seeding types |
| Case file | Assembles transactions, customer and history | 1 | Judging relevance |
| Narrative | Drafts narrative and disposition | 2 | Investigator rewrites |
| Decision | Nothing | none | Investigator, then senior |
| Restriction | Prepares and holds | 4, held | Senior approves |
| Log | Records both decisions | 1 | Confirming |
How do you set and apply the bar?
A disposition is correct if it matches the senior investigator's and the rationale would satisfy a quality reviewer. A right answer with a thin rationale is not correct; in AML the reasoning is the record.
The red lines: closing an alert that shows clear indicators without escalation; any step that would tip off the customer; requesting or releasing a restriction without grounds; clearing a screening match without checking the identifiers. One miss fails the practicum.
Customer explanations are declared, not verified. An investigator who accepts one as proof without corroboration has not met the bar on that case, even if the outcome happens to be right.
Which controls matter, and what evidence do they leave?
- Senior authority on every outcome. Evidence: the senior's name on each closure, escalation and restriction in the append-only record.
- Independent decisions. Evidence: the senior's decision timestamped before comparison.
- Written rationale. Evidence: the investigator's rationale per alert.
- Daily debrief. Evidence: notes on each mismatch.
- Full log. Evidence: the practicum log template for AML investigators.
Pryme Intelligence Academy's Certified Pryme Operator practicum uses the same bar, and its sign-off by a Pryme assessor gives the result a certificate anyone can check.
What should you measure?
Correct dispositions out of the total, split by alert type. Red-line misses (zero). Narratives the investigator rewrote and whether the rewrite added a material fact. And escalations the senior downgraded, which can show over-caution worth coaching rather than a failure.
Where does it usually go wrong?
The queue contains only false positives. The investigator scores well and has never seen a real case. Seed true positives.
The senior reviews the investigator's rationale before forming their own. Agreement then means little.
Restrictions are released by the trainee under pressure from a relationship manager. During the practicum, only the senior acts.
The practicum ends without a decision on approval rights. Whether investigators need a certificate to approve restrictions is the firm's choice; see do AML investigators need a certificate to approve AI agent actions, and on responsibility, are AML investigators accountable for an AI agent's mistakes. Other teams run the same structure; compare running a practicum for collections agents and the guide to AI literacy and certification for people who work with agents.
Questions people ask
Can a practicum alert lead to a real report?
Yes. If the case warrants it, the senior investigator and reporting officer act as they would on any alert. The practicum does not delay a filing.
How many decisions are enough?
Enough that the 95-in-100 bar means something. Around 100 is a sensible floor for AML; fewer makes one error decisive.
Should the investigator see the agent's gate results?
Not during the practicum. They should judge each case on its evidence, as they will afterwards.