How to run enhanced due diligence on high-risk customers with an AI agent

Enhanced due diligence on a high-risk customer is a file to build and a story to test. An agent can assemble ownership, source of wealth, source of funds and screening results, check that the documents support what the customer says, and list every gap. An analyst judges whether the story holds. A senior manager approves the relationship.

What is enhanced due diligence, and who owns it?

Enhanced due diligence is the deeper set of checks applied when a customer presents higher money laundering or terrorist financing risk: politically exposed persons, customers linked to high-risk countries, complex or opaque ownership, unusual transaction patterns, or a business type your risk assessment marks as high. It sits on top of standard customer due diligence, which identifies and verifies the customer and understands the relationship.

The FATF Recommendations set the international expectation: a risk-based approach in which higher risk brings more scrutiny, including approval from senior management for certain relationships. Our note on FATF customer due diligence with AI agents explains how that standard reads when an agent does the preparation.

The financial crime team owns the EDD standard. Analysts own each file. Senior management approves the relationship. The agent prepares, consistent with our guide to AI compliance agents.

What are the steps, and at what depth?

1. Agent, depth 1: opens the file from the trigger. The reason the customer is high risk determines which EDD measures apply. The agent records the trigger and the measures required.
2. Agent, depth 1: maps ownership and control. Every layer to the natural persons who own or control the customer, from registry extracts and the customer's own documents, each marked with its source.
3. Agent, depth 1: collects screening results for each person in the structure. Open alerts are attached; clearing them is a separate process.
4. Agent, depth 2: tests source of wealth. The customer's account of how they built their wealth, set against the evidence: sale documents, accounts, payslips, inheritance papers. Each claim is marked supported, partly supported or unsupported.
5. Agent, depth 2: tests source of funds for the money entering this relationship, tracing it to an account and a reason.
6. Agent, depth 2: drafts requests for missing or weak evidence. The analyst sends them.
7. Agent, depth 2: drafts the EDD summary: who the customer is, what they will do with you, the risks, the evidence, the gaps, and the conditions an approver might set.
8. Person, analyst: assesses plausibility. Does the wealth make sense for this person's age, career and country? This is judgement, not matching.
9. Person, senior manager: approves, approves with conditions, or declines the relationship.

Which controls matter, and what evidence does each leave?

Provenance on every document. A registry extract or an auditor's report is signed by a party that answers for it; a spreadsheet the customer made is declared; a forwarded email is external. Evidence: the provenance class against each item.

Unsupported means unsupported. The agent may not upgrade a claim because it is plausible. Plausibility is the analyst's call. Evidence: claim-by-claim marks.

The approval is a held verdict. Establishing or continuing a high-risk relationship always waits for the named approver. Evidence: held verdict and approval record.

No contact with third parties. The agent drafts requests to the customer only. It does not contact employers, lawyers or other banks. Evidence: outbound messages, all sent by the analyst.

Nothing that could alert a suspect. Where the analyst suspects laundering, the drafted requests must not reveal it. The agent stops drafting once a suspicion referral exists. Evidence: drafting halted against flagged files.

What should you measure?

• Files approved with open gaps, and the conditions set.
• Claims marked unsupported that the analyst later accepted, with the reason.
• Time from trigger to approval decision, and where the waiting happens.
• Requests sent and answered, and customers who stop replying.
• Post-approval surprises: relationships later exited or reported, traced back to what the file showed at approval.

Where does it usually go wrong?

Evidence is collected but not tested. A pile of documents looks thorough. The question is whether they prove the claims. Use your enhanced due diligence checklist as the minimum and the agent's claim marks as the test.

The summary sounds more certain than the file. Fluent writing hides weak evidence. Require the summary to list gaps before conclusions.

Ownership stops one layer early. A trust or foreign company at the top of a structure is not a natural person. The agent should mark the chain incomplete until it reaches people.

Approval becomes a signature. Senior managers sign what the file recommends. The approver should see the gaps and the analyst's view, not only the recommendation; why enhanced due diligence still needs a person argues the point.

Reviews forget the original conditions. Conditions set at approval must be checked at the next review.

Where Pryme fits

The Pryme Context Protocol stamps each EDD document as sealed, signed, declared or external, so the analyst sees at a glance which claims rest on the customer's word. The approval step is a held verdict that names the approver, and the full file, with the agent's marks and the analyst's changes, is filed.