How Do You Train an AI Model on Customer Data in a Bank Without Breaking GDPR?

7 min readAbleneo AI transformation team

Short answer. A bank can train an AI model on customer data without consent, but only after it passes the three-step legitimate-interest test the European Data Protection Board set out in Opinion 28/2024 on 17 December 2024. You must document a real and specific interest, prove the training genuinely needs that data, and show the processing does not override what customers reasonably expect. Pseudonymisation, data minimisation, clear notice, and an honoured right to object are what move the balance in your favour.

1What This Means in Practice

Most banks already hold the data that would make an AI model useful: transaction histories, application forms, support transcripts, fraud flags. The legal question is not whether you own the data. It is whether the purpose you collected it for stretches to training a model, and which GDPR legal basis carries that purpose. Get this wrong and the model itself is exposed, because the EDPB confirmed that unlawful processing during development can taint the lawfulness of the deployed model.

In practice, banks rely on one of three bases for training. Legitimate interest under Article 6(1)(f) covers most internal model development where the use is one customers would expect. Contractual necessity under Article 6(1)(b) is narrow and rarely stretches to training a general model. Explicit consent under Article 6(1)(a) is the fallback when the other two do not fit, and it is fragile because it can be withdrawn.

The Belgian data protection authority has already applied this to banking: it accepted that a bank could use transaction data under legitimate interest to train a model for tailored offers, provided the safeguards held. The ruling shows the basis is available to banks, not that it is automatic.

2Why This Matters for Regulated Industries

For a bank, a data protection problem rarely stays a data protection problem. The same training dataset sits inside the EU AI Act, DORA, and your model risk framework at once. A credit or insurance-pricing model built on customer data is a high-risk AI system under Annex III of the EU AI Act, so the data governance obligations in Article 10, which require relevant, representative, and error-checked training data, apply on top of GDPR. The two regimes are not alternatives. You satisfy both or you ship neither.

The timing is tight. GDPR has applied since 2018 and the EDPB opinion is live now. The EU AI Act’s high-risk obligations for Annex III systems phase in through 2026 and 2027, and a supervisor auditing an AI model will ask to see the legal basis for its training data as a first question. Banks that treat the legal basis as paperwork to add at the end find themselves retraining models on a clean dataset, which costs far more than designing the basis in from the start.

A bank can train AI on customer data without consent, but only after passing the EDPB’s three-step legitimate-interest test from Opinion 28/2024.

3Can You Use Legitimate Interest to Train on Customer Data?

Usually yes, but it is never the default. The EDPB was explicit that Article 6(1)(f) cannot be the automatic basis for AI training. You have to run and document a three-step test. First, identify a real, specific, present interest, not a vague future benefit. Second, show the training is necessary and that you could not reach the same result with less data or anonymised data. Third, balance your interest against the rights of the people in the data, judged on their reasonable expectations and the context in which you collected the data.

The balancing step is where banks pass or fail. A model that serves a purpose close to the one the data was collected for, such as fraud prevention or servicing, usually clears it. A model that reuses the data for a distant purpose usually does not, and that is the point at which consent becomes the honest answer rather than a stretched legitimate interest.

Consent is the right basis when the purpose is one customers would not reasonably expect, when you process special category data, or when another law requires it. Consent under GDPR must be freely given, specific, informed, and as easy to withdraw as to give. That last requirement is why banks avoid building core models on consent: if a meaningful share of customers withdraw, the training set shifts and the model can need rebuilding. Consent works best for optional, customer-facing features, and poorly for infrastructure the bank depends on. Choose the basis per purpose, not per project, because one AI programme often contains several purposes with different answers.

5How Do You Handle Special Category and Sensitive Financial Data?

Article 9 of GDPR bans processing special category data, such as health, biometric, or data revealing ethnicity, unless a narrow exception applies. Financial data is not automatically special category, but a model can infer protected traits from ordinary transactions, and an inference of health or religion from spending patterns falls back under Article 9. The practical rule for banks: assume a model trained on rich transaction data can reconstruct sensitive traits, and either exclude the features that enable that inference or secure an Article 9 exception before training. This is also where the EU AI Act and anti-discrimination law meet GDPR, because a model that learns a protected trait can produce a biased credit or pricing decision.

6What Safeguards Make Training Lawful?

The EDPB weighs mitigating measures directly in the balancing test, so safeguards are not optional extras, they change the legal outcome. The measures that count most for a bank are concrete. Pseudonymise training data so direct identifiers are removed before the model sees them. Apply data minimisation by training on the narrowest feature set that achieves the result. Give transparency beyond the bare Article 13 and 14 notice, including what data trains the model and why. Honour the right to object under Article 21, which means a working path for a customer to have their data excluded. Record each of these in a data protection impact assessment, which is mandatory for high-risk processing and is the document a supervisor will read first.

7How Does the EU AI Act Change This?

The EU AI Act does not replace GDPR, it adds a second layer for high-risk systems. Article 10 sets data governance duties for training, validation, and testing datasets: they must be relevant, sufficiently representative, and examined for bias. For a bank, that means the GDPR legal basis answers whether you may use the data, and the AI Act answers whether the dataset is fit for a high-risk model. A single governance stack can satisfy both, because the pseudonymisation, documentation, and impact-assessment work that GDPR rewards is the same evidence an AI Act conformity assessment expects. Building the two together is cheaper than bolting the second on later.

8The Ableneo Perspective

Ableneo shipped 34 production AI projects in 2025, and roughly 4 of 5 of the projects it starts reach production rather than stalling as pilots. That delivery rate comes from treating the data legal basis as an engineering input at the design stage, not a compliance review at the end. In financial services, where Ableneo works with banks and insurers including ČSOB, Erste, and UNIQA, the model that reaches production is the one whose training data had a documented basis and a clean audit trail from the first sprint. Our AI transformation teams build the GDPR basis, the AI Act data governance, and the model together, so the answer to a supervisor’s first question exists before the supervisor asks it.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo