Financial crime teams need a competency matrix that matches each skill to the decisions agents now prepare: alert triage, screening hits, case narratives and draft suspicious activity reports. Analysts must judge held cases on evidence, investigators must test the agent's reasoning, and the reporting officer alone decides what is filed. The template below sets levels and owners.
Why does financial crime need its own matrix?
Because the decisions that matter most here are the ones an agent must never make. Under the FATF Recommendations, a firm takes a risk-based approach, carries out customer due diligence and reports suspicious transactions. An agent can gather the evidence for each of those steps quickly. It cannot form the suspicion, and it must not decide to file or not to file.
That changes what people need to know. An analyst who used to work alerts from scratch now reviews an agent's draft disposition. The risk is no longer slowness; it is agreement. A tired analyst who approves a well-written narrative without checking the transactions behind it has handed the judgement to the agent in all but name. The matrix exists to prevent that, and it sits within the broader map of AI agent skills by role.
How should you fill in the levels?
Use three levels, named for what the person does:
- Knows: can explain depth, the four verdicts and provenance classes, and why an agent's narrative is a draft.
- Applies: decides held alerts and cases in a live queue, checks each claim in the narrative against its source, and records reasons.
- Signs off: holds the decision rights that stay with people: escalation to investigation, customer exit, and the filing decision.
For each skill, write the minimum level per role and the evidence. Then check the matrix against your approver authority matrix for the financial crime team: nobody should hold an approval right for a skill they have not reached at "Applies" or above.
The template
| Competency | Why it matters | Alert analyst | Investigator | Team lead | Reporting officer | Evidence |
|---|---|---|---|---|---|---|
| Vocabulary: depth, verdicts, provenance | Shared language for every case | |||||
| Checks each narrative claim against its source | Stops approval on trust | |||||
| Weighs provenance of customer explanations | Declared statements are unverified | |||||
| Recognises when the agent should have refused | Catches scope creep | |||||
| Avoids tipping off in agent-drafted messages | Customer contact during a case | |||||
| Decides screening hits held for review | Names, dates, locations | |||||
| Escalates to full investigation | Decision stays with a person | |||||
| Decides whether to file | Never delegated to an agent | |||||
| Reconstructs a closed alert from the record | Supervisory and audit challenge | |||||
| Calibrates reviewers on the same cases | Keeps dispositions consistent |
A worked example
Consider a hypothetical payments firm with eight alert analysts, three investigators, one team lead and a reporting officer. Its triage agent reads alerts at depth 1 and drafts dispositions and narratives at depth 2. Nothing closes without a person.
| Competency | Analyst | Investigator | Team lead | Reporting officer | Evidence |
|---|---|---|---|---|---|
| Vocabulary | Knows | Knows | Knows | Knows | Foundations test |
| Checks narrative against source | Applies | Applies | Applies | Applies | Sample of 50 decisions re-marked |
| Weighs customer explanations | Applies | Applies | Applies | Applies | Scenario test |
| Recognises missed refusal | Knows | Applies | Applies | Applies | Calibration session |
| Avoids tipping off | Applies | Applies | Applies | Signs off | Message review |
| Decides held screening hits | Applies | Applies | Signs off | Signs off | Live queue sign-off |
| Escalates to investigation | Knows | Applies | Signs off | Signs off | Case file review |
| Decides whether to file | Knows | Knows | Applies | Signs off | Named in policy |
| Reconstructs a closed alert | Knows | Applies | Applies | Signs off | Audit walkthrough |
| Calibrates reviewers | Knows | Knows | Signs off | Applies | Monthly session record |
Two entries carry most of the weight. Analysts must reach "Applies" on weighing customer explanations, because a customer's account of where funds came from arrives as declared data: stated, not verified. The agent will quote it fluently. The analyst's job is to see that nothing sealed or signed backs it up.
The second is tipping off. If the agent drafts a request for documents, the analyst must read it as the customer will. A message that mentions a review, a report or an unusual pattern can do harm, and it is the person who sends a depth 2 draft who answers for it. The approver rule for the financial crime team should name who may send customer-facing drafts during an open case.
Where the Academy fits
Pryme Intelligence Academy's Operator track is built for the people in the "Applies" column: how a governed agent decides, human review, and what to do when something goes wrong. Its module tests are set in the learner's own industry and role, so analysts practise on financial crime cases. The practicum then marks two weeks of real review decisions, with no missed red line allowed.
Questions people ask
Can an agent close false-positive screening hits on its own?
Only if the business has decided that a narrow class of hits may close at depth 3 and the agent has earned that depth on the firm's own cases. Many teams keep every screening disposition held for a person, at least until reviewers have seen enough agent work to trust the pattern.
What should a new analyst learn first?
Provenance. Knowing which facts in a case are sealed by the firm's own systems, which are signed by a named party, and which are only declared by the customer is the fastest route to better judgement on agent drafts.
How does this compare with the risk team's matrix?
The risk team's competency matrix focuses on testing controls and setting appetite. This one focuses on case judgement. The two meet at reconstruction: both teams must be able to rebuild any closed alert from the record.