AI for Finance & Banking

What AI actually changes inside a Dutch financial institution

The headline use cases — fraud, risk, AML, reporting — are easy to name and hard to land. The gap is rarely the model. It is whether the system fits a regulated workflow: whether a flag arrives in the analyst's queue with enough context to act, whether a credit decision can be defended to the AFM, whether the audit trail holds up two years later when DNB asks why a customer was declined. We build for that reality, not for a demo. Below is what genuinely shifts on the floor when AI is built properly for banking and financiële dienstverlening — and where the value is, in concrete operational terms.

If you are weighing whether to start at all, the cheapest way to find out is a short, fixed-scope AI audit and strategy engagement. It puts a number on the use case before you commit to building anything.

Fraud detection: the economics are in the false positives

Every Dutch bank and payment provider already runs rules-based fraud screening. The problem is not catching fraud — it is catching it without freezing thousands of legitimate transactions a day. A static rule like "block any first transfer over EUR 5,000 to a new IBAN" stops some fraud and infuriates a lot of real customers, each of whom calls your contact centre. The cost of fraud screening is mostly the cost of being wrong in the other direction.

Machine-learning models score each transaction against the customer's own behavioural baseline — device, time of day, beneficiary history, typical amounts, channel — instead of one-size-fits-all thresholds. In practice that means two things finance teams feel immediately: a meaningful share of fraud that slipped through static rules gets caught, and the volume of blocked-legitimate transactions drops. Industry benchmarks for well-tuned anomaly models put false-positive reductions in the 40–60% range against legacy rule sets; the exact figure depends entirely on your data, so we measure it on yours rather than promise a number.

The operational payoff is concrete: fewer outbound "is this you?" calls, less revenue lost to abandoned legitimate payments, and analysts who spend their hours on cases that are actually suspicious. We design these systems with a human in the loop on high-value holds and a clear reason code on every flag, so the decision is reviewable. This sits squarely in our machine learning and data engineering work — the model is only as good as the feature pipeline feeding it.

Credit and risk scoring under the EU AI Act

Here is the part most vendors skip: under the EU AI Act, creditworthiness assessment of natural persons is explicitly classified as high-risk. That is not a footnote. It means a model used to score consumer or SME loan applications carries documented obligations — risk management, data governance, technical documentation, human oversight, transparency to the affected person, and logging. If you deploy a black-box scoring model without these, you are not just exposed to the AFM; you are exposed under the AI Act regime that phases in across 2025–2027.

That changes the engineering choices. We favour models whose decisions can be explained — gradient-boosted trees with SHAP-style attribution, monotonic constraints so that "higher income never lowers your score", and reason codes mapped to plain-language explanations a customer and an auditor can both read. A consistent, explainable scoring model does more than satisfy a regulator. It removes the drift between individual underwriters, so two applicants with the same profile get the same answer, and it shortens decision time on the straightforward cases so your credit team focuses on the genuine edge cases.

• Consistency: the same inputs produce the same score, every time — defensible to both the applicant and the regulator.
• Speed on the easy cases: clean applications clear faster, freeing analysts for the files that actually need judgement.
• Bias testing built in: we test for disparate impact across protected groups before launch, because a model that quietly discriminates is both an AI Act failure and a reputational one.
• A real audit trail: every score is logged with its inputs and rationale, so you can reconstruct any decision long after it was made.

We design this compliance posture from day one rather than retrofitting it after a finding. If you want the regulatory framing in depth, our note on AI consulting in the Netherlands covers how the AI Act and AVG land on Dutch deployments.

AML transaction monitoring that respects your compliance team's time

Anti-money-laundering monitoring is where Dutch banks have poured the most resource and felt the least relief. The Wwft obligations are strict, the penalties are real — Dutch institutions have paid nine-figure settlements over AML failures — and the standard answer has been to hire more analysts to clear an ever-growing alert backlog. Most of those alerts are noise. Legacy scenario rules routinely run false-positive rates above 90%, which means a compliance analyst spends the overwhelming majority of their day closing alerts that were never going to be a SAR.

Better transaction monitoring does not throw away your scenarios. It layers a model on top that scores and prioritises alerts by genuine risk, suppresses the patterns that are demonstrably benign for a given customer segment, and surfaces network-level signals — money mule rings, layering across accounts — that a single-transaction rule cannot see. The outcome your AML lead cares about is a smaller, sharper alert queue: the same team clears more genuine risk because they stop drowning in noise, and your reporting to the FIU is backed by a documented, explainable rationale rather than a rule that fired for reasons no one can reconstruct.

Because everything here is reviewable by a regulator, explainability and logging are not optional extras — they are the product. We connect to your existing core banking and payments stack through secure APIs, so monitoring improves without ripping out the systems your operations already depend on.

Regulatory reporting and the document mountain behind it

Finance does not only decide — it reports, endlessly. FINREP, COREP, AnaCredit, transaction reporting, internal risk dashboards: a large share of a finance function's manual hours go into pulling numbers together, reconciling them across systems, and producing documents on a deadline. This is unglamorous and it is exactly where AI plus solid data engineering removes drudgery without touching a high-risk decision.

Document and data extraction models read incoming paperwork — KYC packs, loan files, counterparty documents, statements — and turn them into structured, checked data instead of fields someone keys in by hand. Retrieval-grounded language models let a risk officer or relationship manager ask a plain question — "what is our total exposure to this counterparty across all entities?" — and get an answer cited to your own internal sources, not a hallucination. The benefit is measured in analyst-days returned each month and in fewer reconciliation errors reaching a regulatory submission, where a mistake is expensive. We keep these assistants grounded strictly in your own policies and data, with no customer information leaving your control.

Why a boutique partner fits Dutch finance better than a giant

Financial services is the one sector where the wrong AI partner costs you twice — once in the build, and again when a regulator looks at it. The big enterprise consultancies (Xebia, Xomnia, Capgemini) can staff a finance programme, but you pay enterprise rates and the senior people who sold the work often rotate off it. The weekend-rebranded "AI" web agencies will happily ship a scoring model with no thought to the AI Act's high-risk obligations — which in finance is not a corner you can cut.

Crux Digits is the senior-led middle path. Founded in 2022 and based at Vlierhoeve 100 in Nieuwegein (province of Utrecht), we serve the Utrecht region, the whole Netherlands and Europe, bilingually in English and Dutch. The senior engineers who scope your fraud or AML system stay on it through production, and you end up owning the solution rather than renting it. We are an AI engineering partner, not a marketing agency, and compliance — EU AI Act plus GDPR/AVG — is built in from the first conversation, because in finance it has to be.

Our pricing is transparent and stepwise, so a risk or compliance budget owner can see the path before committing: an AI Audit & Strategy at EUR 2,500 to find and size the right use case, a Proof of Concept at EUR 20,000 to prove it on your real data, and production launch from EUR 50,000 (day-rate guidance around EUR 150/hour). The full breakdown sits on the pricing page. Across the firm we have delivered 13 case studies spanning forecasting, NLP, computer vision and predictive maintenance — client names confidential, as financial clients require.

Where to start

You do not need an AI roadmap to begin — you need one well-chosen use case proven on your own data. The lowest-risk first move is usually the one where the cost of the status quo is easiest to count: the false-positive load in fraud or AML, the underwriter inconsistency in credit, or the analyst-hours buried in reporting. Pick the pain you can already measure and we will tell you, honestly, whether AI is the right tool for it.

If it is, the audit puts a number and a compliant plan on the table for EUR 2,500 before you spend anything on a build. Tell us the problem — fraud, credit risk, AML monitoring or reporting — and we will map a defensible path to value. Reach Tom Joseph at Crux Digits on info@cruxdigits.nl or +31 6 44384676, or read how others have started with our AI consulting in the Netherlands and AI implementation services.