To overview
Without this foundation, every AI project will fail in terms of compliance

Without this foundation, every AI project will fail in terms of compliance

September 14, 2026 · 5 min read

They implement a new AI model for alert processing. The first few weeks go well, and the numbers look promising. Three months later, someone notices that the model isn't based on the documented process, but has systematically adopted shortcuts and inconsistencies that have crept into the existing process over years.

Such errors often go unnoticed for a long time because a model has no sense for gaps. If an experienced analyst is presented with an incomplete dataset, they will hopefully notice it and ask for clarification. An AI model working with the same incomplete dataset won't notice, but will still make a decision—quickly and with remarkable confidence.

Why is that? Rarely is it the model itself, but usually it's because no one looked closely enough before the project started: Does the process really work as everyone believes? Is the data the model uses reliable? And will anyone even be paying attention if something goes wrong later? These three questions about processes, data, and governance are often dismissed too quickly or even omitted entirely before an AI project begins.

Processes: The difference between what is documented and what happens

Many institutions have a documented process for handling alerts. And then there's the experienced employee who's been doing it differently for years – faster, more pragmatically, with tricks that aren't documented anywhere. Both processes exist in parallel: the written one and the one actually used.

If an AI system is now trained based on historical data, it learns the process as it actually happened – that is, what really happened, not what's written in the documentation. At first glance, this seems good: the system reflects reality. On closer inspection, less so, because it also adopts the shortcuts, inconsistencies, and errors that have crept into this process over the years, often without ever having been checked.

As an obvious reaction, many institutions are updating their process documentation: setting up a project, holding workshops with the relevant departments, and finally producing a new document. This takes time, and by the time the new documentation is finished, reality has long since changed again – whether due to new employees, new responsibilities, new regulatory requirements, or a new tool. In the AML context, this is even more critical than in other areas, because the auditor ultimately sees a document that may no longer describe what actually happens in your company.

One solution approach that we often implement in practice is process mining.

This involves reconstructing how a process actually runs using system logs, instead of blindly relying on documentation. For AI projects in the AML environment, this should be the first step, even before discussing suitable models.

However, there is one limitation: Process mining only works if the process and data run in systems that generate logs. This is far from universally the case. In many organizations, parts of the KYC or KYB processes still rely on Excel spreadsheets, or documents are printed, signed, scanned again, and stored in a third system. As a result, there are no continuous logs, and the process exists partly outside the system, namely in employees' minds, emails, or local files.

This is where GenAI can help by reading and structuring this analog information – scanned documents, Excel spreadsheets with free-text columns. AI doesn't have to be the end of a digitization project; it can also be the first step, for example, in process mining, to shed light on the analog world.

Data: Three systems, three truths

Besides poorly documented processes, many institutions face another challenge: the data their AI models are meant to work with. It must be complete, consistent, and traceable. A real-world example that many will recognize: During customer onboarding at a bank, an address was entered into the CRM system three years ago. The customer has since moved, but the new address was only updated in the account management system. A third version of the customer record exists in the AML monitoring system because the last synchronization was eight months ago.

If you now want to build a model that can detect changes in the record in combination with transaction patterns (typical in identity theft), you have a problem. The model needs reliable traceability: When did what change, and where did this information come from? This is

precisely where data lineage comes in. It allows you to trace the origin of every data point, when it was last modified, and by which process or system. Without this information, firstly, you cannot train a reliable model, and secondly, you cannot explain to an examiner why a model made a particular decision at a particular time, because in case of doubt, you yourself do not know which version of the data the model saw at that time.

A working solution requires two things

These include technical lineage tools that automatically track how data flows through your systems and where it is transformed, as well as organizational clarity about who is responsible for which data. Before investing in tooling, it's therefore worth asking: Which data points in your organization are so critical that you need to know their origin – and who in the company actually knows this?

Governance: Who notices when AI is wrong?

Who is actually responsible for ensuring that an AI system functions well in your day-to-day operations? Who checks it regularly, and who is informed if the quality declines?

For example, the IT department says the model is technically flawless, but compliance criticizes the results. What happens in such a case at your company? Another example: An analyst notices that a certain type of alert has been occurring significantly more often for weeks than before. Is this a new fraud pattern, or is it due to the model? Or conversely: There are suddenly noticeably fewer hits for customers from a specific country – is this truly less fraud, or is it more likely a model error?

In all these cases, a defined escalation path is needed: Who listens to these observations? Who has the authority to say, "We're going to investigate this now"? Who can make changes to the system, and who documents that this question was even asked and answered? Someone must have the authority and the ability to ask and answer such questions – not just at go-live, but continuously.

The path to success: Use case instead of fundamental reform

Documenting processes clearly, establishing a transparent and traceable data foundation, and clarifying or building governance: This sounds like a major project that can take months, if not years. Understandably, no one wants to wait that long before even starting with AI.

Adding to the time pressure is the new EU Anti-Money Laundering Regulation (AML Regulation), which comes into effect on July 10, 2027, and requires traceable data and robust processes. Those who only start building the foundations then will be under time pressure, and that can quickly become expensive.

The more practical approach: You don't have to create the foundation for the entire company all at once. It's sufficient to start with the one specific use case you want to begin with and clearly answer three questions for it.

  • Which part of the process is actually relevant for this use case, and does it really work the way you expect?
  • Which five to ten data fields does this use case depend on, and are they good enough for your test cases?
  • Who checks in the first few weeks whether the results are plausible?

This takes weeks, not months or years. And you learn exactly where the biggest challenges lie in your company – crucial knowledge when it comes to scaling later on. 

Visible costs are lower than hidden costs

Anyone still hesitating because a solid foundation for AI projects seems complex and expensive should turn the question around: What does it actually cost to have no foundation? Every pilot project that never goes into production, every vendor configuration whose default settings are never questioned, and every hour an analyst spends in a system whose results they don't trust, probably already incur immense costs today – not as an item in a project budget, but as hidden costs in terms of inefficiency and risk.

Would you like to know how robust your own foundation is for AI projects? Feel free to contact us via This email address is protected against spambots! JavaScript must be enabled to view it. or on LinkedIn.

FAQ

Why do AI projects fail in terms of AML compliance?

The problem rarely lies with the model itself. Usually, before implementation, it wasn't checked whether the process actually runs as documented, whether the data is reliable, and who continuously monitors the results. A model trained on historical data, for example, adopts the shortcuts and inconsistencies of the actual process without noticing data gaps, yet still makes quick and confident decisions.

What is process mining and why does it help with AI projects in AML compliance?

Process mining reconstructs how a process actually runs based on system logs, instead of relying on documentation. For AI projects in the AML environment, this should be the first step, even before choosing a model. It only works where processes run in systems that generate logs; analogous information such as scanned documents or Excel spreadsheets can be read and structured by GenAI beforehand.

What is data lineage and why is it important for AI in compliance?

Data lineage makes it possible to trace the origin of every data point, when it was last modified, and by which process or system. Without it, neither a reliable model can be trained nor can an auditor be informed which data version a model decision was based on. A typical problem is customer data that exists in three different versions across CRM, account management, and AML monitoring.

How do you start an AI project in compliance without months of preparation?

With a concrete use case instead of a fundamental reform. For this case, three questions need to be answered: Which part of the process is relevant and does it actually work as intended? Which five to ten data fields does the case depend on, and are they sufficient? Who checks whether the results are plausible in the first few weeks? This takes weeks, not months.

 

Note: This article does not replace legal or regulatory advice

also interesting

Do you have a specific problem in mind?