AI for finance AI Act high-risk DNB/AFM supervision DORA & MiFID II

AI for banks, asset managers, fintech and insurers

AI in the financial sector touches fraud detection, AML/KYC automation, credit scoring, document AI on contracts and KIID/PRIIPS generation all at once. We build models and integrations that keep working within the framework of the AI Act, DNB/AFM supervision, EBA Guidelines and DORA, with audit trails, model risk management and verified explainability as the starting point, not a document produced after the fact.

Schedule a call View use cases
TRANSACTION MONITORING AML score 0.92 Anomaly HIGH Audit-trail v3.1 model-risk-management · SR 11-7 · EBA Guidelines

AI in finance: opportunities and frameworks together

The financial sector processes enormous volumes of structured transaction data, unstructured contracts, KYC files and market information. That makes it one of the sectors best suited to AI, and at the same time one of the most heavily regulated. For banks, asset managers, fintech players and insurers, every AI system needs a design that covers not only accuracy but also traceability, explainability and governance from day one.

We develop AI solutions for financial organisations that both the CTO/CIO and the second and third lines (risk, compliance, internal audit) must be able to defend. That means model risk management based on SR 11-7 principles, EBA guidelines for model validation, a working integration with the existing DNB/AFM reporting landscape, and the operational resilience that DORA requires. No lab experiments without an end goal, but production AI with a sustainable governance layer underneath.

On this page we explain the main use cases: from transaction anomaly detection and AML/KYC to credit scoring (where the AI Act's high-risk rules apply), document AI on credit agreements and ISDA paperwork, automated KIID/PRIIPS generation, customer service AI within MiFID II limits, sentiment analytics on market data, and the hard limits around robo-advisory.

AI use cases in finance

Eight concrete domains where AI delivers direct value within the prudential frameworks of banks, insurers, asset managers and fintech players. Each domain requires its own risk classification and governance approach.

Fraud detection and transaction anomalies

Real-time scoring of payments, card transactions and account openings against anomaly patterns. A combination of rule engines, gradient-boosted trees and sequential neural models, embedded in a feedback loop with fraud analysts so that false positives actually fall rather than staying structurally high.

AML/KYC automation

Document extraction from identity documents, UBO investigations, sanctions-list screening and ongoing transaction monitoring within a single pipeline. AI speeds up the human review, but every high-risk decision remains a documented human check, in line with the Wwft (Dutch AML Act) and the EBA Guidelines on ML/TF risk management.

Credit scoring and credit acceptance

Under the AI Act, creditworthiness scoring is explicitly classed as high-risk. That requires a conformity assessment, a risk management system, training-data governance and logging that you can present in detail. We design scoring models that deliver a full audit package alongside strong performance.

Document AI for credit agreements

Extraction and classification of clauses from credit agreements, ISDA Master Agreements, CSAs and intercompany loans. Automatic detection of clauses that deviate from a bank standard, and integration with the credit administration system for downstream reporting and covenant monitoring.

KIID/PRIIPs document generation

Generation and periodic updating of Key Information Documents and PRIIPs fact sheets based on fund and product data. Built-in validation of mandatory sections, language variants and change detection against the previous version, so compliance can review rather than draft by hand.

Customer service AI within MiFID II

Chat and voice assistants for retail and private banking clients that stay strictly within the boundaries of non-advice. The line between providing information and giving investment advice is a hard one under MiFID II, and we design the prompts, guardrails and escalation paths explicitly around that line.

Sentiment analytics on market data

NLP on news feeds, earnings calls, broker research and social data to aggregate sentiment signals at instrument level. Intended as input for portfolio managers and quantitative teams, not as an automated order trigger without human review and trading limits.

Robo-advisory within clear boundaries

Fully automated investment advice falls under MiFID II, with the same suitability assessment as human advice. We build advisory workflows with a transparent factor engine, suitability checks, escalation to a human adviser and an audit log that the regulator can use to reconstruct decisions.

Insurance AI: claims and pricing

Claims triage, fraud detection on claims, automated first notice of loss and risk pricing support. EIOPA expects insurers to explicitly safeguard AI governance, fairness and explainability, especially for products that directly affect premiums or coverage.

Compliance first: AI Act, DNB/AFM, EBA, EIOPA and DORA

An AI system in finance is not "done" when it scores accurately on a test set. It is done when it demonstrably meets the combined requirements of the AI Act, sector supervision and operational resilience rules. We build those frameworks into the architecture instead of mapping them afterwards.

AI Act high-risk: what it requires in practice

The AI Act explicitly classifies credit scoring, life and health insurance pricing and certain HR applications as high-risk. That means a working risk management system across the entire lifecycle, data governance for training and test data, technical documentation, logging, human oversight, robustness and cybersecurity, plus a conformity assessment before market introduction.

DNB and AFM supervision of AI

DNB has published General Principles for the Use of AI in the Financial Sector: soundness, accountability, fairness, ethics, skills and transparency. The AFM looks at duty of care and market integrity from a conduct supervision perspective. Both require demonstrable governance, explainable models and a working human override.

EBA Guidelines and model risk management (SR 11-7)

The EBA guidelines on model risk management, ICT and outsourcing are the de facto benchmark for banks. SR 11-7 (Federal Reserve) is treated as the de facto standard by large international players: development, implementation, use and validation must be kept independently separate, with independent model validation.

DORA: operational resilience for AI systems

DORA sets requirements for ICT risk management, incident reporting, resilience testing, third-party risk and information sharing. AI systems fall fully within its scope: this affects your choice of cloud, model hosting, vendor lock-in with LLM providers and the way we design fall-back paths into the architecture.

EIOPA and insurers

For insurers, the EIOPA AI governance principles framework applies alongside Solvency II. Fairness and explainability in pricing and claims handling in particular are under scrutiny, because indirect discrimination through proxy variables is a real and demonstrable risk.

GDPR and purpose limitation

Training data drawn from customer files falls directly under the GDPR: lawful basis, purpose limitation, retention periods and the right to an explanation in automated decision-making (Article 22). We keep training, test and production data strictly separate and document the purpose limitation for each dataset.

How we build AI systems for finance

Our approach always starts with the question: which framework will this system fall under, and which risk classification fits it? Only then do we move on to model choice, dataset and architecture. This prevents a promising proof of concept from later stumbling on compliance.

Discovery with risk and compliance involved

In an AI discovery process for finance, we involve second-line risk and compliance from day one. The use case, data basis and initial risk classification are fixed before we move to the build phase. Also read about our approach at /ai-discovery-workshop.

Model validation and challenger models

For high-risk models, we work with an independent validation track and challenger models. Performance measurements, stability, fairness tests across relevant segments and stress tests on extreme scenarios are documented in a model validation report that can be handed over to the second and third lines.

Audit trails and logging by design

Every model decision that has an impact on a customer or a supervisory report is recorded: input features (hashed where necessary), model version, score, decision and the human override. Audit trails are a first-class architectural component, not a log file that can be rotated away by accident.

MLOps with model risk control

Pipelines with automated data quality checks, drift detection, retraining triggers and clear promotion gates from staging to production. A model only goes live after a documented sign-off from the model owner, validation and compliance. Comparable to our predictive analytics dashboard approach.

Privacy architecture and data segregation

Data classification, pseudonymisation, key management and strict separation between training, test and production. For LLM applications, that means: no customer data goes to an external provider without a DPA, an update to the processing register and a DPIA on the table.

Integration with core banking and policy administration

Our AI rarely lands on a greenfield site; it usually has to connect to a core banking system, policy administration, ERP or GRC tooling. We build event-driven integrations for this, with idempotency and dead-letter queues. See building a core banking platform.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

Tools and stack we use

A combination of classical ML, gradient boosting, sequential models and, where appropriate, large language models. Not every use case calls for an LLM; for scoring, a transparent boosted tree often works better, while for document extraction an LLM with retrieval is the better fit.

Models and frameworks

Python with scikit-learn, XGBoost and LightGBM for scoring, PyTorch for sequential models, and Hugging Face and OpenAI/Azure OpenAI for document and language applications. For explainability we use SHAP and counterfactual explanations, and for fairness checks Fairlearn and internal metrics per protected segment.

Data infrastructure

Snowflake, Databricks or BigQuery as the data platform, dbt for transformations, Apache Kafka for event streams and Airflow for batch pipelines. Strict separation between development, acceptance and production environments, with separate keys and roles.

MLOps and monitoring

MLflow or Vertex AI for the model registry and lineage, Evidently or WhyLabs for data and model drift, and Prometheus and Grafana for service metrics. Alerts on drift, latency and exceptions are forwarded immediately to the model owner team and the on-call SRE.

Cloud, security and DORA fit

Azure, AWS and Google Cloud are all workable, provided you set up the cloud region, encryption at rest and identity management correctly. For DORA third-party risk, we document the chain of sub-processors, the exit strategy and concentration risk explicitly in the solution architecture.

Security, governance and model risk management

In finance, security is not a separate module bolted on at the end – it is a design prerequisite. That applies twice over when AI models are tied to customer decisions or supervisory reporting.

Three lines model in the architecture

The business line develops and uses the model, risk management and compliance validate it independently, and internal audit reviews the entire chain. We make that separation concrete through separate environments, role-based access and a review flow in which a model never goes to production without the four-eyes principle.

Explainability for customers and regulators

For automated decision-making, customers have a right to an explanation under GDPR Article 22. Regulators want to be able to reconstruct why a model gave a particular score in a specific case. We handle this with SHAP explanation endpoints, a readable reason string for customer communication and a complete feature snapshot in the audit log.

Operational resilience and fallbacks

DORA requires demonstrable resilience: what happens if the LLM provider goes down, what if a feature store fails, and how quickly can we revert to a backup model. We design circuit breakers, fallbacks to a previous model version and, where possible, a manual path so the business never grinds to a halt.

Vendor lock-in and LLM governance

An AI stack that depends entirely on a single LLM provider is a DORA concentration risk. We abstract model providers behind an internal facade so you can switch models, run them locally where required, and keep your vendor choice from dictating your compliance position.

Penetration testing and model attacks

Alongside standard pentests, we look at model-specific attack vectors: prompt injection, data poisoning, model extraction and evasion attacks on fraud detection. For production LLMs, we build an input/output sanitisation layer with explicit detection and logging.

Integration with DORA compliance software

The AI system must supply the right signals for your wider DORA reporting: incident classification, third-party overview and resilience test results. Read about our approach at DORA compliance software and AI Act compliance software.

Frequently asked questions about AI in banking and finance

When is an AI application in finance "high-risk" under the AI Act?

The AI Act explicitly names creditworthiness evaluation and certain insurance pricing applications as high-risk. This brings obligations around risk management, data governance, technical documentation, logging, human oversight, robustness and a conformity assessment before market introduction. A fraud detection model is not automatically high-risk, but it often still falls under DNB or AFM expectations around explainability and model validation.

Can an LLM give investment advice directly?

Under MiFID II, fully automated investment advice is permitted, but exactly the same suitability assessment applies as for human advice: knowledge, experience, risk appetite and objectives must be recorded, and the advice must fit them. A chatbot without those built-in checks must therefore not give advice; in that case we put strict guardrails in place around the difference between information and advice.

How do you handle DNB expectations around AI governance?

We work from the DNB General Principles for the Use of AI: soundness, accountability, fairness, ethics, skills and transparency. In practice, that means a documented model owner, a working model risk framework, fairness tests across relevant segments, and an explainability layer that works not only technically but also towards the customer. Our deliverables are structured so that you can use them directly in a conversation with DNB or the AFM.

How does this fit within DORA?

DORA affects AI systems through ICT risk management, incident reporting, resilience testing and third-party risk. We document the chain of model providers, design fallback models and make sub-processors visible in the third-party register. We carry out resilience testing (TLPT where applicable) not only at the application layer but also on the model itself: what happens if the LLM provider disappears, or if a feature store becomes corrupted.

Can you connect AI models to our existing core banking or policy administration system?

Yes. In almost every project, the model lands in an existing landscape of core banking, policy administration, ERP and a data platform. We work with event-driven integrations via Kafka or a service bus, idempotent endpoints and clearly defined contracts, so the model stays decoupled from the system of record.

What explainability do your models offer?

For scoring models, we deliver SHAP or comparable feature attributions per decision, plus a translated reason code for customer communication. For LLM applications, we deliver the prompt, the retrieved context and a trace so that a compliance officer can reconstruct why the model produced a particular outcome. For high-risk applications, explainability is not optional.

What is your approach to model risk management?

We align with the EBA Guidelines and SR 11-7 principles: separation between development, validation and use, an independent validation report, periodic re-validation triggers and a working governance forum that signs off on production. For banks that do not yet have a mature framework, we help set one up in parallel with the first model.

Do you build it yourselves, or push a platform?

We build custom rather than pushing our own platform. That means we choose open components (Python, MLflow, dbt, Kafka) plus the cloud stack that fits your DORA and data residency choices. No vendor lock-in on a proprietary black box, but an architecture your own team can carry on with.

Ready for an AI project that will pass risk and compliance?

We build AI for banks, asset managers, fintech companies and insurers, with the frameworks of the AI Act, DNB/AFM, EBA, EIOPA and DORA as the starting point, not as a final check. Schedule a conversation for a concrete first exploration, or take a look at our related services on fintech app development and AI in asset management.

Schedule a call

Edit content