Ayoob AI

AI Automation for Amsterdam Finance and Fintech

·11 min read·Husain Ayoob
AI automationAmsterdamfintechfinancial services

Amsterdam is a financial centre of unusual range. It is the headquarters of one of the world's largest payment processors, two of Europe's larger banks, and the administrators of one of the biggest pension pools on the planet, and since Brexit it has gathered a cluster of EU trading venues as well. All of it runs on document-heavy, heavily-regulated work, and all of it now sits under the EU AI Act, DORA, and GDPR at once. That combination is what makes the city a serious automation market, and it is also why the architecture matters more here than the model does.

The useful way to read the Dutch and EU rulebook is not as a barrier to AI but as a specification for how to build it: auditable, explainable, under human control, and with the data kept where the law can see it. This guide is the blog's full treatment of that EU stack, applied to Dutch finance. For the neighbouring markets, see the Luxembourg fund-administration guide and the Singapore financial-services guide; for how the same private-architecture argument plays under a UK sector regulator rather than the EU AI Act, the UK automation hub.

Payments and fintech at scale

Start with the pillar Amsterdam owns outright. Adyen, headquartered in the city, processed on the order of 1.4 trillion euros of payments in 2025, and the wider Dutch payments and fintech cluster, Mollie, bunq, and several hundred others, makes the Netherlands one of Europe's densest fintech ecosystems. Payments at that scale generate enormous, repetitive, document-shaped work: reconciling settlements against unstructured statements, handling chargebacks and disputes, onboarding merchants with know-your-business checks, and triaging transaction-monitoring alerts. None of that is the glamorous end of AI, and all of it is exactly where a private system earns its keep, because the volume is high, the rules are explicit, and the data is sensitive.

The Dutch twin-peaks model

Two regulators sit over this market, and confusing them is a quick way to lose a Dutch reader's trust. The Netherlands runs a twin-peaks model: De Nederlandsche Bank, the DNB, is the prudential supervisor, and the Autoriteit Financiële Markten, the AFM, is the conduct supervisor. They supervise the use of AI jointly, and their 2024 joint report on AI in the financial sector flags data quality, data protection, explainability, biased or incorrect results, and rising third-party dependence, while stressing human-in-the-loop controls and audit trails. The DNB's longer-standing SAFEST principles, Soundness, Accountability, Fairness, Ethics, Skills, and Transparency, make the same point: accountability stays with the institution, the AI must be explainable, and staff must understand the system's limits. Ayoob AI is neither of these regulators and makes none of their judgements; we build the system they expect you to be able to control and audit.

The EU AI Act, explained properly

Because this is the blog's reference treatment, it is worth being precise. The AI Act sorts systems into tiers: a small set of prohibited practices, a defined band of high-risk uses, limited-risk uses with transparency duties, and everything else as minimal-risk, plus a separate regime for general-purpose AI models.

For finance the load-bearing fact is that AI used to evaluate the creditworthiness or credit score of individuals is high-risk under Annex III, with fraud detection specifically carved out. High-risk status brings real obligations: a conformity assessment, data-governance controls, logged human oversight under Article 14, post-market monitoring, and a nine-part technical-documentation pack under Article 11 and Annex IV. The system cannot be fully autonomous, and there is no human-oversight derogation even for an already-regulated bank. Separately, Article 5 prohibits social-scoring-style assessment that judges people from unrelated social behaviour.

The timeline matters and is partly in flux. The prohibitions and the AI-literacy duties applied from February 2025, and the general-purpose-AI obligations from August 2025. The high-risk obligations for uses like credit scoring were originally due in August 2026, but under the EU's Digital Omnibus package they are reported to be deferred toward December 2027. That deferral is provisional and not yet final law, so the responsible way to read it is as additional runway to build properly, not as a settled deadline to relax against. The practical conclusion is the same either way: a high-risk financial AI has to be explainable, documented, overseen by a human, and auditable, which is what a private, full-code system is built to be.

DORA, explained properly

The second pillar of the EU stack is the Digital Operational Resilience Act, which has applied to financial entities since January 2025. DORA covers ICT risk management, incident reporting, resilience testing, and, most relevant here, ICT third-party risk. Every financial entity has to keep a Register of Information listing all its ICT third-party arrangements, with provider, scope, locations, subcontractors, and criticality, alongside dependency maps and exit strategies, and it has to watch for concentration risk where a critical function leans on a single, hard-to-replace provider.

This is where a careful claim matters. Engaging a vendor, even for a private on-premise build, is itself an ICT third-party arrangement that belongs in the register, so no one should tell you that a private system makes DORA go away. What it does do is change the dependency calculus. A hosted SaaS model is a run-time dependency on someone else's service and a flow of your data out to it; a private system that runs inside your environment removes that run-time reliance and keeps the data in. It helps with concentration and resilience, and it leaves the DORA duties, correctly, with you. The broader case for that architecture in a regulated, auditable setting is set out in private AI for UK regulated businesses.

GDPR and where the data can go

The third pillar is the one people forget until an audit. Regulated EU financial data cannot simply be routed through a US-hosted cloud AI API; it needs a lawful basis for the transfer and a hard look at GDPR and the sector rules. A private, on-premise build answers this by construction, because the customer data is processed inside the client's environment and never leaves it, with the access logs and audit trail that a supervisor will ask for. The full architecture is covered in private AI on-premise; the Dutch point is that data residency here is achieved by where the system runs, not by a vendor's promise about where it stores things.

The bunq lesson

One Dutch story captures the whole argument. The challenger bank bunq won an appeal in 2022 when the central bank tried to block its use of AI and machine learning for anti-money-laundering monitoring; the Trade and Industry Appeals Tribunal held that Dutch AML law sets open standards and does not dictate exactly how a bank screens, so AI-driven monitoring is permitted. That is the green light. The amber light came in 2025, when the DNB nonetheless fined bunq 2.6 million euros for AML failures over an earlier period, with regulators noting that over-reliance on automation had bred complacency in the manual review of alerts. And the red light is the shared industry utility for transaction monitoring, wound down in 2024 after the pooling of customer data across banks ran into data-protection and legal-basis problems.

Read together, the three tell you precisely what to build and what to avoid. AI-driven monitoring is allowed; automation without rigorous human review is punished; and pooling or outsourcing the underlying data is fragile. Auditable, human-in-the-loop, and kept inside each institution's own walls is not a marketing posture here, it is the shape the Dutch experience has already validated.

The Dutch pensions surge

The least-discussed pillar may be the largest. The Netherlands holds one of the biggest pension pools in the world, among the highest pension-asset-to-GDP ratios in the OECD, administered through giants like APG for ABP and PGGM for PFZW. On top of that steady-state load sits a once-in-a-generation event: the Wtp pension reform is migrating an estimated eleven million participants onto a new system, a vast exercise in records reconciliation, data-quality work, and member communication concentrated into a short window. That is an enormous, structured, document-heavy workload of exactly the kind a private system handles well, performed on deeply personal data that has every reason to stay in-environment.

Automating the paperwork the rules create

Here is the angle most automation pitches miss. The EU stack does not only constrain AI, it generates work, and that work is itself automatable. A DORA Register of Information has to be assembled and kept current. An AI Act high-risk system needs its nine-part technical documentation drafted and maintained through every change. Incident evidence has to be collected and structured. These are recurring, templated, document-shaped tasks, and a private system can draft and maintain them, with a person reviewing and signing. To be clear about the line: this is automating the compliance paperwork, never the compliance decision. The pattern is the one in our AI compliance automation guide, applied to the specific artefacts the Dutch and EU regimes demand.

Where a private build pays back first

For an Amsterdam finance or fintech operation, the highest-return, lowest-risk starting points are:

  • Payments and settlement reconciliation from unstructured statements
  • Merchant and client onboarding, with know-your-business and know-your-customer collation
  • Transaction-monitoring alert triage and case-file drafting, with the MLRO deciding
  • Pension participant data-quality and benefit-administration document processing
  • DORA Register of Information and incident-evidence assembly
  • AI Act technical-documentation drafting and upkeep
  • Private internal search across your own documents and records, the retrieval-augmented pattern kept entirely in-environment

Each of these is assembly, extraction, or drafting, and each keeps a qualified human at the decision point. The general shape is covered in AI for finance teams; the value here is doing it privately, against the Dutch and EU specifics.

The cost case

The return scales with the cost of the people whose routine load it removes, and Amsterdam is an expensive, scarce market for exactly the talent this work consumes. Senior engineering and analytical pay is high, the competition for machine-learning hires runs against London and New York, and a majority of Dutch companies report that skill shortages are slowing their AI delivery. Against that, the cost of building a private automation does not move with the local salary. We work that calculation in full, in any currency, in the true cost of your most expensive roles; our retainers run from GBP 4,000 to GBP 6,000 per month as of June 2026, set against the cost of a year of in-house hiring and building.

Working with us

Ayoob AI is an engineering firm based in Newcastle upon Tyne with a second office in Dubai, delivering to Dutch clients remotely. We build private, full-code systems on infrastructure you control, where customer and pension data never leaves your environment, we are ISO 27001:2022 and Cyber Essentials certified, and we hold five pending UK patents on the on-device compute that makes the private model practical. We are not a bank and not a DNB or AFM-supervised entity, so credit, anti-money-laundering, and regulatory decisions, and the DORA and AI Act duties that go with them, stay with you. What we provide is the private architecture, the auditability, and the integration into your existing systems, set up so your own people can meet their obligations and the reasoning behind a full-code build over a generic tool is the one that fits a regulated buyer.

If you run a payments business, a bank, an asset or pension manager, or a fintech in Amsterdam and want to map which parts of your regulated document load can be automated without your data ever leaving your environment, that is what an initial discovery call is for, and you can start one through our AI automation service.

Related reading

About the author
Husain Ayoob, Founder & CEO, Ayoob AI Ltd
Husain Ayoob

Founder & CEO, Ayoob AI Ltd

BSc Computer Science with AI, Northumbria University 2024. 5 UK patents pending covering the Ayoob AI stack. ISO 27001:2022 certified (organisation).

Full bio, patents, and press →

Frequently asked questions

Does the EU AI Act stop us using AI for credit decisions?

No, but it governs how. AI that scores the creditworthiness of individuals is classified as high-risk under Annex III of the AI Act, which brings conformity assessment, data governance, logged human oversight, and a technical-documentation pack, and the system cannot be fully autonomous. There is no human-oversight carve-out even for an already-regulated bank. Fraud detection is treated separately and is carved out of that high-risk category. As of mid-2026 the high-risk obligations for uses like credit scoring are reported to be deferred toward December 2027 under the EU's Digital Omnibus package, but that is provisional and not yet final law, so it is best treated as build runway rather than a settled date. We build the auditable, human-in-the-loop substrate; the credit decision and the conformity work stay with you.

If we use your AI, does that take care of our DORA obligations?

No, and we are careful not to imply it. DORA has applied since January 2025, and outsourcing never transfers accountability. Even a private, on-premise system that we build and support is an ICT third-party arrangement that belongs in your Register of Information. What a private build does change is the dependency picture: it removes the run-time reliance on a hosted model and keeps the data inside your environment, which helps with concentration risk and operational resilience. But the DORA duties, the register, the exit strategy, the resilience testing, remain yours.

Can your AI make our AML or transaction-monitoring decisions?

No. It triages alerts, enriches cases, and drafts the case file, but the Money Laundering Reporting Officer decides and signs, and a suspicious-transaction report is a human call. The bunq saga is the cautionary tale. A Dutch court confirmed that AI-driven monitoring is permissible under the open standards of Dutch anti-money-laundering law, yet the central bank still fined bunq 2.6 million euros in 2025, partly because over-reliance on automation had bred complacency in manual alert review. Auditable and human-in-the-loop is the only design that survives that scrutiny.

We are an EU financial entity. Can we put customer data into a hosted AI API?

Regulated EU financial data generally cannot flow freely into a US-hosted cloud AI service without a lawful transfer basis and a hard look at GDPR and DORA. A private, on-premise build sidesteps the question: the data is processed inside your own environment and never leaves it, with full access logs and audit trails. The collapse of the shared Dutch AML utility, wound down in 2024 after data-protection and legal-basis challenges to pooled transaction-monitoring, is the proof that the outsourced and pooled path is the fragile one.

You have no Amsterdam office. Does that matter?

No. A private, on-premise build runs inside your environment wherever we happen to sit, so geography does not change the data picture. We deliver remotely from Newcastle, with a second office in Dubai, and the UK and the EU recognise each other as adequate for data protection, with the UK's adequacy renewed in December 2025, so remote handling of personal data is lawful. We are an engineering firm, not a DNB or AFM-supervised entity, so the regulatory decisions and the regulatory duties stay with you.

Will off-the-shelf regtech not already cover this?

The horizontal KYC and anti-money-laundering platforms are capable, and we do not try to replace them. The gap they leave is the bespoke document and data work around the edges: reconciling unstructured statements, drafting and maintaining DORA and AI Act documentation, reconciling pension records, and private internal search across your own documents. That bespoke, private, full-code layer, kept inside your environment, is what we build.

Want to discuss how this applies to your business?

Book a Discovery Call