Approach · 22 November 2026 · Martijn van Grieken

Detecting fraud and anomalies

Fraud that was never discovered is not in your examples. How a model learns what is normal, ranks anomalies and explains every signal.

Analistenbureau met twee schermen vol puntenwolken waarin enkele punten oplichten, en een teal bureaulamp

An insurer, bank or benefits agency processes thousands of claims, payments or applications every day. Its assessors can only look at a small share of them. Which share that is, is usually decided by a set of rules drawn up after earlier fraud cases. Those rules catch what has been caught before. A new method slips through until someone happens to spot it.

This page describes how we approach detecting fraud and anomalies. Examples from government are described under public sector.

Your examples lack exactly what you are looking for

An ordinary predictive model learns from labelled examples: this case was fraud, that one was not. Fraud that was never discovered is not in those examples. A model trained on them mainly learns to recognise the known cases.

So we turn it around. The model learns what an ordinary claim, payment or application looks like, and flags what deviates from it. You do not need a list of fraud cases for that. Outlier detection finds cases that fall far outside the usual pattern. Clustering groups what is similar, so that the cases that fit nowhere stand out by themselves.

A ranking, with a reason for every signal

The model does not say ‘fraud’ or ‘no fraud’. Every case gets a score, and the assessor starts at the top of the list. That way your investigation capacity determines where the cut-off lies: a hundred correct signals that nobody looks at deliver nothing.

Each signal shows why it stands out. An assessor who cannot see that will set the alert aside, and a decision about a customer or applicant has to be explainable afterwards.

Rank Score Why this case stands out
1 0.94 Four applications under different names to one bank account
2 0.91 An amount of €1,250, where comparable applications are around €12.50
3 0.87 Maximum amount claimed, two days after the policy started
4 0.62 Three applications within an hour from the same address

Illustrative example of a fictitious organisation that pays out benefits.

Number 2 is probably not fraud. Someone put the decimal point in the wrong place.

Many anomalies are input errors

For a music rights organisation our specialists examined claims on music rights. Together with the client they drew up twenty indicators that combine into a single ranking. Many anomalies turned out to be input errors.

For anyone looking for fraud, that seems a disappointment. Yet it is useful. Every input error points to a place where the administration can improve: a check at input, a drop-down list instead of a free text field. And the less noise there is in the data, the better the real anomalies rise to the top of the list. The ways we can work together are described under Services.

How we approach it

We work in the phases of CRISP-DM:

  1. Gather patterns. We talk to assessors and domain experts about what they consider suspicious and which data exists.
  2. Draw up indicators. Every indicator gets a description in plain language, so that an assessor can explain it in their own words.
  3. Techniques side by side. We compare several techniques and combine them into one ranking.
  4. Your own assessment. Your assessors investigate the top of the list. Whether something is fraud is for them to establish.
  5. Feed back and decide. Confirmed cases and input errors found go back into the model. Then you decide: continue, adjust or stop.

What you can organise as an executive

  • Decide beforehand how many signals per week your organisation can investigate.
  • Ask for an explanation with every signal that an assessor can put into their own words.
  • Treat a high score as a reason to investigate. Suspicion only arises after assessment.
  • Appoint an owner for the input errors that come to light.
  • Record what happens with confirmed cases, so that the model learns from them.
  • Involve your data protection officer from the start. Flagging based on personal data requires a prior assessment.

Discussing your detection question

Bring one process in which you receive more cases than you can assess. That first conversation is without obligation and we aim to respond within one working day.

Discuss your question

Author: Martijn van Grieken, founder of Twentynext. How we make this. AI wrote the first version of the Dutch original, and Martijn read and approved it before it went live. This English version was translated with AI. Twentynext is responsible for the content but cannot guarantee that it is accurate or complete. See our disclaimer.

Related articles