KYC/AML AI workflows: documents, candidate matching and alert review

Choose one KYC/AML preparation task: extract fields from approved documents, organize candidate matches or assemble inspectable context for alerts. Separate these outputs from account opening, case closure, reporting and other action authority.

The FATF report introduction on AML/CFT technologies discusses risk-based enabling conditions, privacy and inclusion. It is not product approval or local filing authority. The institution establishes actual permitted actions and responsibilities.

Split the path by evidence and responsibility

Three agents are not a required architecture. Compare parsing, rules, search and fixed orchestration first, then identify where model assistance is useful. Explicit scope helps trace failures, source ownership and review effort.

Separate case state from reference labels

Historical closure reasons can be inspected, but similar patterns may involve different data, times or later outcomes. Citations, similarity or no suspicious-transaction report do not independently establish a safe current case or a negative label.

Define evaluation target, case/customer/transaction unit, reference-label source and observation window. Keep unresolved, pending and delayed outcomes as separate states, updating versions when evidence changes. Field quality, identity candidates, alert classification and reporting work are different tasks rather than one accuracy metric.

Match denominators to the target

On one labelled binary evaluation set, TP/FN are reference positives flagged/missed; FP/TN are reference negatives flagged/unflagged. The classification evaluation documentation defines:

Report an undefined measure when its denominator is zero. Track unresolved outcomes separately instead of treating them as negatives. Fewer alerts or reports alone do not establish better detection; inspect label coverage, missed targets and workload.

Accept preparation quality and reviewer work

Use representative cases separate from tuning data. Inspect fields, support, wrong candidates, omissions and unauthorized access. Record lookup, review, correction, escalation and complete handling effort to establish capacity.

Source support aids verification rather than replacing authorized judgment. The FSC's 2024 AI guidance offers administrative risk/responsibility guidance; acceptance follows actual use, controls and applicable requirements.

Retain inspection records and handoff

For the agreed investigation purpose, record input provenance, model/rule versions, output, actions, reviewers and changes. Set necessary content, access and retention. Logs do not establish deterministic model replay, regulatory approval or complete detection.

Agree stopping and handoff for missing data, service failure, uncertainty and resends. Evaluate affected cases after changes and retain recovery versions. The NIST AI RMF supports lifecycle evaluation.

Read the financial adoption guide for intended use, or the existing agentic workflow service for approved tools, review and scope.