Skip to content

Fraud detection with data and AI

We put rules and a model side by side and give your investigators a score with an explanation. For GVB, we built a model to detect unusual hours.

Our anomaly detection model gives GVB’s Rail Materiaal department insight into unusual patterns in logged hours

This was about planning, not fraud. The Rail Materiaal department wanted to understand unusual patterns in logged hours. Our anomaly detection model now provides project-level insight into productivity, the completeness of time records, overtime and sick leave, so each project can be planned more effectively.

Read the case study

How does our team spot the few suspicious expense claims among all the routine ones?

Reviewers use fixed rules and spot checks. The rules generate many alerts that turn out to be false alarms, while new patterns slip through. A model used alongside those rules ranks transactions, invoices or claims by how unusual they are, so the team starts with the ones that stand out most.

Book a call

Every transaction gets a score and a reason; the investigator decides

Discuss your checks

Anomalies in your own data

Transactions, expense claims, invoices, hours or insurance claims: the model learns what is normal and flags what stands out. For GVB, we did this with logged hours.

Rules and model side by side

Known patterns remain fixed rules. The model catches what the rules miss. For each alert, you can see which rule or feature triggered it.

A manageable review queue

Your team investigates a limited number of cases per period. We set the threshold so the list fits that capacity, with the highest scores at the top.

Investigation outcomes fed back

Every investigation outcome, whether fraud or not, is fed back as a label. This helps the model learn from false alarms and makes it possible to refine the list.

What your investigators already know becomes the first version of the model

  1. Signals and investigated cases

    We work with investigators to document the patterns they recognise and the cases already investigated, including their outcomes.

  2. Backtesting on historical data

    We run the rules and model on older data. We measure how many genuine cases appear near the top and how many false alarms there are.

  3. Signals in the work queue

    Scores and their explanations feed into the team’s system. Every outcome feeds back into the process, and we monitor quality.

Do we record enough investigation outcomes to train a model?

Talk it through with us

A good fit if

Many transactions, reimbursement requests or claims pass through your organisation, a team checks them using rules or sampling, and the outcomes of previous investigations are recorded somewhere.

Not a good fit if

There are few transactions and an investigator can review them all, or nobody has time to investigate flagged cases. A score would then only create a longer list.

What determines the scope

How many sources we connect, whether investigation outcomes are already recorded, whether you need a result immediately for each transaction or in a periodic list, and how much explanation the investigator needs for each flagged case.

Will the model make decisions about people?

No. The model provides a score and a reason; an investigator reviews the case, makes enquiries and decides. You should assess how to record that human review with your data protection officer.

Fixed rules, anomaly detection or a model trained on investigated cases: where do you start?

Fixed rules

Choose this when the patterns are known and can be clearly defined, such as duplicate invoices or expense claims above a threshold. The results are easy to explain, but you may miss new patterns.

Anomaly detection

Choose this when you have few investigated cases. The model looks for behaviour that differs from the norm. Not every anomaly is fraud, so an investigator is still needed.

A model trained on investigated cases

Choose this when you record the outcomes of investigations. Our advice: keep the rules and run the model alongside them, so you can see what produced each alert.

We’ve already highlighted anomalies in operational data, including for GVB

Data Science Lab has been building data and AI applications since 2016. Based in Amsterdam, we are a Databricks and Microsoft partner. For GVB, we used an anomaly detection model to highlight unusual time entries.

Book an introductory meeting

The cases you’ve investigated and your current control rules set the agenda for our first conversation

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your point of contact. In our first conversation, we’ll discuss which stream you monitor and which rules are currently running.

Which transactions or claims would you like to check more closely?

Describe the flow, the rules you currently use and who investigates the alerts. We’ll reply by email, or we can arrange an initial call.

Enter a valid email address.

Let us know how we can help you.

Frequently asked questions

Fraud detection identifies unusual transactions, expense claims, invoices or insurance claims in large volumes of data. Rules catch known patterns; a model finds cases that deviate from those patterns or resemble previously confirmed fraud. Using data or AI to detect fraud does not produce a verdict. It gives your investigators a prioritised list of signals.

It is one of the AI applications we build.

Anomaly detection looks for cases that differ from what is normal, without defining fraud in advance. Techniques such as isolation forest or an autoencoder give each case an anomaly score. But unusual does not mean wrong: a large order, a new supplier or a seasonal peak can also stand out. That is why we discuss with the team beforehand which anomalies matter.

Every signal takes an investigator’s time, and a list full of false alarms will be ignored. So we set the threshold with the team: how many cases can it investigate in a given period, and how many missed cases are you willing to accept? When we backtest, we measure how many genuine cases appear at the top of the list: the precision for that workload.

For each signal, we show the features that contributed most, such as an unusual amount for that supplier or an expense claim outside working hours. For tree-based models, techniques such as SHAP can help. An investigator must be able to check the reason and explain it to a colleague; we do not deliver a score without an explanation.

We record it as a label: fraud, error or legitimate. This lets you train a model on investigated cases and see which rules mainly produce false alarms.

Be aware of the bias: you only know the outcomes of cases someone investigated. That is why we also include a small random sample, so the model does not learn only from what the old rules already found.

If people are involved, the GDPR applies. Article 22 gives people the right not to be subject to a decision based solely on automated processing that has legal effects or significantly affects them. That is why our model provides a signal and a person makes the decision. Your legal counsel and data protection officer determine what constitutes meaningful human review.

For what a DPIA requires in this context, read DPIA for AI.

Any sector with more transactions or applications than a team can review individually: insurance claims, invoices and payments in financial services, and expense claims and timesheets at large employers.

Read more about data and AI for insurers and for financial and business services.

When the volume is small enough for an investigator to review everything, when a few fixed rules already catch the known patterns, or when nobody has time to follow up on signals. Sometimes it helps more to get the data in order first, so duplicate suppliers or empty fields no longer occur.

If you want to understand more broadly what a model can do, read about developing a machine learning model.