Automating the AML process of an accounting firm with AI
An accounting firm can use AI to prepare much of its AML work (collecting documents, mapping beneficial owners, preparing sanctions and PEP screening, proposing a risk rating) while the firm and its AMLCO keep every decision. Here is how Belgian accounting and tax firms (fiduciaires) set it up.
Get in touchWhat AI prepares, what the firm decides.
In Belgium, much of the anti-money-laundering work that the law of 18 September 2017 imposes on accounting and tax firms (fiduciaires), and that the ITAA (the Belgian Institute for Tax Advisors and Accountants) oversees for its members, can be prepared by AI: gathering and reading the documents, reconstructing the ownership chain, preparing the screening, proposing a risk rating, tracking review deadlines. The decision itself is never automated. Accepting or refusing a client, qualifying a suspicion, reporting to the CTIF-CFI (the Belgian Financial Intelligence Processing Unit): that stays with the firm and its AMLCO (the officer responsible for anti-money-laundering compliance).
Keyser Soze's method starts from that line: first work out what the AMLCO can actually review, then automate within that limit. Otherwise, the AMLCO's signature becomes a rubber stamp.
- collecting and reading ID documents, articles of association and register extracts
- the ownership chain and a proposal of beneficial owners
- the comparison with the UBO register, with discrepancies flagged for the AMLCO
- preparing sanctions and PEP screening in your tools, and listing the matches to examine
- a proposed risk rating, with its reasons
- a dated, sourced file, ready for review
- tracking periodic reviews according to the risk level
- validating the identification and the beneficial owner
- validating the risk rating
- examining any screening match (sanctions, PEP)
- reporting a discrepancy to the UBO register
- ambiguous cases
- accepting, continuing or ending the relationship
- qualifying a suspicion and reporting to the CTIF-CFI
- signing off the file
The method, in five steps.
-
Start from your obligations and your portfolio, not from the tool.
Your overall risk assessment, your internal procedures, the real make-up of your client base: individuals, companies, multi-layered structures. That starting point shows where AI saves days and where it adds nothing.
-
Size the human review before building.
Every case the AI sends back to the AMLCO takes review time. If the queue exceeds what the AMLCO can actually examine, files end up signed without being read. And a history of unexamined approvals is worse than no automation at all: it documents a check that never took place. So the review workload is calculated first, then the portfolio update is split into batches that fit the AMLCO's capacity.
-
Send back to a person what belongs with a person.
Some cases always go back to a person: control exercised by means other than ownership, a discrepancy with the UBO register, being unable to identify a beneficial owner or having doubts about one, the question of falling back on the senior managing official, activity inconsistent with the known profile. These are signals to handle, never default values the system fills in silently.
-
Review the first batches in full, then measure.
Whatever the test results, the first files produced from real documents are all reviewed. Errors are then measured at real volume, against criteria the AMLCO understands and approves. The right indicator is not “so many files processed” but “every file carries an up-to-date, supported and defensible record”. Review is scaled back only on preparation tasks where the measurement allows it; decisions stay human.
-
Log, hand over, stay adaptable.
The workflow is designed so that every AI proposal and every human decision is logged. Your team is trained to run the workflow on its own. Thresholds, time periods and lists are parameters, not code: when European rules change, much of the adaptation is done through settings.
Your data, your control.
Installed in your environment, under the firm's control: no client file passes through a third-party service without your explicit decision — and if an external component is justified, it is covered by an art. 28 GDPR processor agreement.
No message goes to the client automatically. Nothing in the workflow may reveal that a suspicion is being examined or that a report to the CTIF-CFI is being considered: the tipping-off prohibition also applies to what your tools send or display.
What this is not.
You keep your screening tools: the workflow is organised around them. No decision is automatic, and responsibility does not transfer: the firm and its AMLCO remain responsible; AI saves them the preparation time.
Where to start.
A first 30-minute call, free, to see whether your portfolio lends itself to this. Then the AI Diagnostic Sprint (€1,500–2,500): your obligations, your portfolio and your review capacity examined, and a costed plan for the set-up.
Book 30 minutesGeoffrey Michiels, founder of Keyser Soze (Brussels, 2014), leads the set-up on site, as project manager and translator between the business and the technology: he scopes it with your AMLCO, steers the build, checks it and hands it over. More than 25 years of digital project management, in English, French or Dutch.
How can an accounting firm automate AML with AI?
An accounting firm automates AML by letting AI do much of the AML preparation (collecting and reading documents, mapping the ownership chain and proposing the beneficial owners, preparing sanctions and PEP screening in your existing tools, proposing a risk rating, tracking periodic reviews) while the firm and its AMLCO keep every decision, including accepting the client, validating the risk rating, examining any screening match, reporting a discrepancy to the UBO register, qualifying a suspicion and reporting to the CTIF-CFI. Keyser Soze's method starts by calculating how much the AMLCO can actually review, then automates within that limit.
Can AI file the report to the CTIF-CFI for us?
No. AI detects and prepares. Qualifying a suspicion and reporting to the CTIF-CFI remain human acts, carried out by the firm. The workflow is designed so that the AMLCO receives a complete file, not to decide in the AMLCO's place.
How long does it take to update our whole portfolio?
It depends first on how much review time your AMLCO has, not on the AI. That is the method's first calculation: it sets the batch size and the schedule, so that every file is actually examined.
Do we need to change our AML software?
No. Your screening and practice-management tools stay in place. The work is on the workflow around them: what is prepared automatically, what is escalated to a person, and how it is logged.
What if the AI gets it wrong?
It sometimes will, like any process. That is why the first batches are reviewed in full, errors are measured at real volume, and the decision stays human. Review is scaled back only where the figures allow it.
Does the EU AML Regulation (AMLR) change things?
Regulation (EU) 2024/1624 will apply from 10 July 2027: a large part of the due-diligence rules becomes directly applicable across the EU, and Belgian law will be adapted accordingly. The workflow is built so that thresholds, time periods and lists are parameters: much of the adaptation will be done through settings rather than by rebuilding.
Talk to me.
I read every message myself. Response within 2 business days. You bring the scope, I bring the plan.
Based in Brussels — working in Brussels, Wallonia and Flanders. Missions in FR, NL and EN.
Book 30 minutes