AI for AML/KYC: From Onboarding to STR/SAR

This one-day course rebuilds four manual AML tasks, maximising trusted chatbot assistance while keeping every answer tied to a named document or data field and an accountable sign-off.

Participants document each manual process, request the complete chatbot deliverable in one go, then apply the Four-colors method:

• Green, verified by a repeatable test

• Red, refused with reasons when evidence fails or sources conflict

• Orange, a judgement signed by a named person

• Grey, explicitly incomplete, with nothing established

Confidence means traceability, not guaranteed correctness: unsupported claims never appear supported.

AI can support and partly automate customer due diligence, screening and suspicious-activity analysis; regulatory responsibility remains with the institution. Material screening decisions and suspicious transaction/activity report submissions require institutional governance and accountable human approval.

Recommended prerequisites:

• AI Driving Licence in Companies (ref 012301) or AI Driving Licence for Financial Sector Employees (ref 012859)

• How to Navigate the EU Artificial Intelligence Act (ref 011838)

• Generative AI for Back-Office Operations (ref 012303) or Generative AI for Finance and Insurance Professionals (ref 012302)

• Open a customer file a reviewer can check instead of redo: the as-of date and the governing version fixed first, a document and page on every field, quotes confirmed by exact search, and what the document says kept apart from what was concluded

• Close a list, show the arithmetic and judge an absence: the sweep behind a beneficial-owner list, the multiplication chain behind every ownership percentage, and an absence recorded as resting on a search rather than as a fact

• Keep the chase list and stop the queue turning into answers: every open item carries what was already tried, and an answer moving between the four tasks travels with its state attached

• Make the rulebook the source of the rule: thresholds, designations and typology names quoted from the authority with its version, and a deadlock between two governing sources refused and escalated rather than reconciled

• Redesign the onboarding task rather than the prompt: the process as it really runs with its times and signatures, the steps sorted into machine, assisted and human-only work on evidence, and where the freed time went

• Sign for work somebody else prepared, in the order that catches things: the unresolved list, the test for an unresolved item dressed as a verified one, the quotes and one recomputed figure, then the judgements

• Get the work out of the chatbot on purpose: finding separated from writing, the deliverable split into extraction, arithmetic, drafting and judgement, and what changes on a corporate deployment

• Grade every answer green, red, orange or grey, and know what work each state calls for

• Produce a one-page procedure for this workflow and map it against relevant AI governance requirements, including AI literacy, human oversight, traceability and record-keeping where applicable under the AI Act and financial-sector governance requirements

• Decide task by task whether to use the tool at all, including whether the customer data concerned may be provided to the approved AI environment under the institution's data-protection, confidentiality, professional-secrecy and information-security rules, and distinguish proportionate checking from going through the motions

Contents (8 hours), built around one synthetic customer carried through four tasks:

1. The AML file as a checklist, and the four-step method:

• the manual process

• the far extreme

• what it would take to trust it

• and where the line falls

Regulatory context includes the Luxembourg AML/CFT framework and the transition to the EU AML Single Rulebook (AMLR/AMLA).

2. The manual process written down, then the far extreme: the beneficial ownership walk through a layered structure, asked for in one go (Exercise).

3. What it would take to trust it: the four-colours method, and grading the output green, red, orange or grey (Exercise).

4. Building to the requirements: finding separated from writing, a source for every field, absence reportable, visible working for every figure (Exercise).

5. The line and the supervision protocol: what the AI hands over, what the human owns, and the three checks that keep supervision bounded, with relevant AI Act and financial-sector governance requirements mapped onto them, including human oversight, traceability and record-keeping where applicable (Exercise).

6. Screening hit adjudication: the named discriminating field, and the comparison table as the supervising check (Exercise).

7. Alert review and the disposition narrative: computed figures, and measured facts kept apart from inferred ones (Exercise).

8. Preparing an STR/SAR for reporting to the Luxembourg FIU (CRF): completeness against the applicable reporting schema, provenance back to the earlier cases, and the human decision gate (Exercise).

9. Afterwards: approved enterprise GenAI, access controls, logging, model/vendor governance, testing, human approval gates and periodic review.

Designed for :

• AML KYC analysts and approving officers,

• Onboarding and client lifecycle staff,

• Transaction monitoring teams,

• Compliance officers and MLROs who review and approve their work.

The course uses employer-approved chatbot accounts and synthetic data generated in the room.

No programming, no installation or product demonstration is involved.

• Short taught blocks alternating with hands-on work in the participants' own chatbot

• One case built with the group, then three run independently

Packs are exchanged and reviewed, so every participant both builds and approves.

After the course, participants can download their House of Training certificate of attendance from the learner portal.