What should teams do first when building AI agent prompts for AML investigations?

Start by separating extraction, interpretation, and narration

The first design move is to treat the prompt as three different jobs, not one. Data gathering tells the agent what to collect and from where, summarisation tells it how to compress facts without adding meaning, and narrative tells it how to present the result for investigators. That split reduces ambiguity and gives reviewers a clear place to look when outputs drift.

A single opaque prompt tends to mix evidence collection, judgement, and prose generation in one instruction block. That is where teams usually lose control: the agent starts inventing connective tissue, skipping key fields, or blending raw signals with conclusions. A layered structure makes it easier to define what must be factual, what may be inferred, and what must remain formatting only.

For AML work, this is especially important because investigators need traceability. The same source data may support different outputs depending on whether the agent is extracting transactions, identifying counterparties, or drafting a case narrative. If those responsibilities are not separated early, prompt edits become hard to review and harder to test.

Why this structure improves AML prompt quality

Layering the prompt forces a cleaner control point for each decision. The data-gathering layer can be judged on completeness and source fidelity, the summarisation layer on whether it preserves material facts, and the narrative layer on whether it stays within approved language and tone. That lets teams spot whether a failure is a retrieval problem, a reasoning problem, or a presentation problem.

It also creates a better path for validation. Investigators and model reviewers can compare the agent’s extracted fields against the source record before they ever assess the final write-up. In practice, that means fewer hidden errors, less rework, and a more defensible audit trail when the prompt is used in production case handling.

Teams should also use the structure to keep judgement where it belongs. If the agent is allowed to combine extraction and interpretation too early, it may overstate risk, miss benign explanations, or produce a narrative that sounds conclusive before the evidence supports it. A layered prompt makes that boundary explicit.

How to apply the split in real investigative workflows

One practical pattern is to define each layer with a different output contract. The extraction layer should return only source-backed fields, the summarisation layer should restate those fields in a compact case view, and the narrative layer should turn that case view into investigator-ready prose. This is easier to govern than a single request that asks for "analysis" in one step.

Teams can then review each layer with different acceptance criteria. If the extraction layer fails, the fix is usually scope or source access. If the summarisation layer fails, the fix is usually wording or context handling. If the narrative layer fails, the fix is usually template discipline or over-generation. That separation shortens debugging and makes prompt changes safer to approve.

For AML investigations, this approach also helps with consistency across analysts and scenarios. A good prompt should produce the same evidence structure whether the case involves unusual payments, beneficial ownership questions, or transaction monitoring alerts. The layer model gives the team a repeatable pattern they can reuse without rewriting the whole prompt each time.

Risk and Threat Considerations

Without a layered prompt, an agent can blur evidence and conclusion, which creates avoidable review risk in investigations. The bigger the case volume and the more the prompt is reused, the more likely that hidden prompt drift will produce inconsistent summaries, missed red flags, or narratives that read more certain than the source data supports.

Failure mechanism: A single instruction block lets the model shortcut from raw inputs to a polished answer, so extraction errors, weak inference, and narrative embellishment can become indistinguishable during review.

Impact: Investigators lose traceability, quality checks become harder, and false confidence in the output can delay escalation or force rework later in the case lifecycle.

Practitioner Guidance

What to verify: Before trusting the prompt, test that each layer can fail independently without breaking the others. A strong AML prompt should let reviewers see exactly which facts were pulled, which facts were condensed, and which wording was generated for the case note.

What good looks like: The prompt produces a source-faithful extraction table, a compact case summary, and a narrative that adds no new facts. If a reviewer can trace every sentence in the narrative back to an extracted field or approved interpretation rule, the design is working.

Practitioner takeaway: The first goal is not a cleverer prompt, it is a reviewable prompt, because separable layers make AML output easier to validate, harder to overstate, and safer to operationalise.

Related resources from NHI Mgmt Group

• What should teams do first after an AI agent privilege escalation flaw is found?
• What should teams do in the first 24 to 72 hours after discovering a compromised AI agent runtime?
• Which AML controls should teams prioritise first when building an implementation plan for South Africa?
• How do teams decide which AI agent deployment needs the strongest controls first?