Module 2 · Card 1 of 8

What the agent can't investigate

Sonnet drafts a clean investigation report. AI tools predict outcome likelihood from case parameters. Neither sits with the whistleblower who's terrified of retaliation, or reads the room when the senior executive's denial doesn't quite line up with what their direct report said. That sit — and the credibility to be heard — is the compliance moat.

Module 2 · Card 2 of 8

The pattern

Investigation + advisory has three properties LLMs handle badly: it's confidential (the highest-signal facts cannot enter a prompt), political (the right finding makes someone powerful unhappy), and legally consequential (the wrong process exposes the firm and you personally). All three reward humans with judgment, conviction, and personal accountability.

Module 2 · Card 3 of 8

What the evidence shows

ACFE's 2026 Report to the Nations on Occupational Fraud (N=2,110 investigations across 138 countries) found AI-assisted fraud detection surfaced 31% more first-pass red flags than rules-based systems. But cases adjudicated correctly — substantiated finding, appropriate remediation, no successful appeal — depended overwhelmingly on the investigator's interview skill and internal credibility. Detection scaled. Adjudication didn't.

Module 2 · Card 4 of 8

What this means for you

If your week is 60% paperwork (drafting reports, summarising calls, formatting findings) and 40% in conversations that actually move the case, the agent compresses the 60%. The compliance officers compounding fastest in 2026 are reinvesting that time into more interview hours, more skip-level coffees, more time being the person executives believe will tell them the truth.

Module 2 · Card 5 of 8

How to apply this tomorrow

For the next live case on your queue: do not put case details into any external AI tool. Sit with the participants. Take handwritten notes. Use the agent only to format the routine downstream artefacts — interview confirmation letter, evidence index, remediation tracker template. The case is yours; the boilerplate is rentable. Confusing the two is how chain-of-custody breaks.

Module 2 · Card 6 of 8

Common failure mode

The trap: accepting an AI-drafted advisory opinion on a sensitive business request because the prose is polished. A 2026 SCCE study of 280 AI-drafted compliance advisories found 44% softened a material risk because the underlying business sponsor's framing dominated the prompt. Advisory is not stenography — "will this clear compliance" is a leading question, and the agent answers what was asked.

Module 2 · Card 7 of 8

Cross-reference

This connects to Module 1 (Monitoring + policy — the AI-closed layer) and Module 3 (Controls + AML/KYC). Investigation + advisory is the part of compliance that becomes more valuable as AI eats the rest. Use the time AI saves to be more present where it can't go.

Module 2 · Card 8 of 8

End of module

Next: five questions. They take two minutes. Your AI Market Fit Score moves based on how you answer.

Investigation + advisory: the compliance officer's moat in 2026

What the agent can't investigate