AI in financial support: what is allowed and what is not

Key takeaways

• AI agents are ideal for status information and help with forms: at the Swiss digital bank Yuh, the chatbot covers around 45 per cent of all requests.
• Article 22 GDPR prohibits purely automated decisions with legal effect, for instance the rejection of a loan.
• BaFin warns against algorithmic discrimination and requires strict data governance for high-risk systems.
• Complaints and suspected fraud have to be handed to the human team so the situation can be defused quickly.
• AI in support may not provide investment advice and has to stay strictly within documents the business has approved in advance.

Status information and forms. What AI may do straight away.

Using artificial intelligence in the customer service of financial providers rarely fails on the technology. It fails on unclear regulatory responsibilities. Banks, insurers and fintechs work in an environment where a wrong answer immediately creates liability or supervisory risk. Even so, there is a wide range of customer requests that can be automated without legal concern and without human approval: pure information.

That includes recurring questions about opening hours, account fees, branch locations or how long a standard transfer takes. Navigation to documents is equally well suited to AI agents, such as forms for exemption orders, changes of address or powers of attorney. As long as the system merely quotes and explains existing terms that are public or contractually agreed, no supervisory duty to advise applies.

Practice shows how large this lever is. A survey by the Lucerne University of Applied Sciences (HSLU) on the customer service of 61 Swiss banks describes, using the digital bank Yuh as an example, that its chatbot today covers around 45 per cent of all requests, according to Yuh especially reliably for standard matters concerning features, prices and general topics¹. For support teams that means a considerable relief from routine work, provided that GDPR-compliant AI draws on clean knowledge sources.

The hard limit. Article 22 GDPR.

The legal limit of automation begins where information turns into a decision. Article 22(1) of the General Data Protection Regulation (GDPR) sets out a clear prohibition: data subjects have the right not to be subject to a decision based solely on automated processing which produces legal effects concerning them or similarly significantly affects them.

For financial institutions this means a strict ban on autonomous AI execution in critical procedures. An AI agent may not reject a loan application on its own, refuse to settle an insurance claim, or unilaterally close a current account. Even where risk models suggest a rejection in the background, the final legal decision has to remain reviewable by a person or be taken by staff directly.

• Legal effect: a contract does not come about, is terminated, or benefits are finally refused. This is where the ban on purely automated processes under Art. 22(1) GDPR applies.
• Significant impact: financial restrictions such as lowering a credit line or blocking account functions require human control mechanisms without exception.
• Rights of data subjects: under Art. 22(3) GDPR, institutions have to make sure that customers can obtain the intervention of a person, express their own point of view and contest the decision.

Exceptions apply under Art. 22(2) GDPR only where the automated decision is necessary for entering into or performing a contract, is authorised by law, or is based on explicit consent. Even then, accompanying safeguards such as the right to manual review remain legally required.

Fairness and bias. What BaFin requires.

Beyond data protection, the financial supervisor sets strict requirements for operational governance. The Federal Financial Supervisory Authority (BaFin) stresses that the general organisational and risk-management requirements under § 25a of the Kreditwesengesetz (KWG) and § 23 of the Versicherungsaufsichtsgesetz (VAG) apply without restriction to algorithms and AI systems². Institutions have to be able to demonstrate at any time how automated systems arrive at their results.

A central risk lies in unconscious distortions (bias) within the training and context data. Where historical data is taken over uncritically, indirect or direct discrimination against protected groups under the Allgemeines Gleichbehandlungsgesetz (AGG) becomes a real danger, for example on grounds of age, gender or ethnic origin. If an AI system systematically treats requests differently depending on language patterns or place of residence, the supervisor intervenes.

• Risk classification: the European AI Act classes AI systems for assessing creditworthiness or for risk assessment in life and health insurance as high-risk systems.
• Data governance: training and knowledge data have to be representative, free of errors and free of impermissible discriminatory features.
• Transparency duties: clear labelling duties apply to support chatbots, among them those that Article 50 of the AI Act sets out for systems that interact with people.

For customer support, a clear rule follows from this: an AI agent may not offer dynamic terms in chat on the basis of opaque customer scoring. It has to deliver identical, traceable facts to every customer, drawn from the same verified documents.

Complaints and suspicion. When the human takes over.

There are situations in day-to-day finance where automation is not merely unsuitable but negligent. That applies above all to security incidents, suspected fraud and formal regulatory complaints. In those moments, a few minutes of response time often decide between financial damage and a supervisory reprimand.

When a customer reports a phishing incident, unauthorised debits or the loss of access credentials, no AI should be running time-consuming dialogue loops. Emergency protocols take over: blocking cards and access, and immediate routing to specialised fraud teams. The same applies to suspicions under the Geldwäschegesetz (GwG). If a support agent spots suspicious transaction patterns, reporting duties under § 43 GwG can be triggered. An AI may neither assess such processes nor archive them on its own.

Language models also reach a natural limit with emotionally charged complaints. Goodwill decisions call for human judgement and negotiating skill that no machine can reproduce with legal certainty. A structured transition to human specialists is the only reliable route here. How to set that process up cleanly is described in the guide to the handover to a human.

Support is not investment advice.

One of the sharpest dividing lines in the financial sector runs between general support and investment advice, which requires authorisation. Under § 2(8) sentence 1 no. 10 of the Wertpapierhandelsgesetz (WpHG) and the European MiFID II directives, investment advice exists as soon as a personal recommendation on transactions in specific financial instruments is given that is based on the personal circumstances of the customer.

For generative AI this carries a considerable risk. If a user asks in the support chat about a suitable way to invest their savings and the language model formulates an apparently helpful recommendation for a particular ETF or equity fund, the institution is legally already in the territory of investment advice. Where neither a suitability assessment nor the regulatory information duties were observed, serious supervisory sanctions can follow.

• Permitted in support: explaining key figures, providing key information documents (KIDs), naming order fees and execution venues.
• Prohibited for support AI: statements about future performance, recommendations to buy or sell particular securities, and questions about individual risk appetite in order to find a product.

Support agents have to be instructed with clear system boundaries so that they intervene the moment the wording comes close to advice. Instead of offering an assessment, the system refers factually to the official product documents or offers an appointment with certified advisers.

Reliable knowledge. No AI hallucinations.

False statements from a customer service bot are irritating in e-commerce; in the financial sector they are a compliance breach. So-called hallucinations, where language models generate plausible but factually wrong answers from freely associated training data, can only be prevented by a strictly deterministic knowledge architecture.

The standard for financial institutions is retrieval-augmented generation (RAG). The language model does not act as a free writer but purely as an intelligent intermediary: before every answer, the system searches a curated internal knowledge base, extracts the relevant passages and formulates a precise piece of information from them. If the stored documents hold no unambiguous answer, the system refuses to guess and starts the handover to the team.

One indispensable element is complete source transparency. Every statement the agent makes has to be linked transparently to the underlying document, for the customer and in the audit log, for instance to the relevant section of the terms or the price list. How teams secure this control mechanism technically is explored in the article on how to prevent AI hallucinations in support.

A seamless handover. From agent to team.

A reliable AI system in finance is not one that answers every question, but one that knows its own limits. The goal is not full automation at any price but a hybrid architecture: the agent filters out routine requests, structures incoming data in advance and passes complex or regulatorily sensitive cases to the support team without a break in the medium.

When the hand-off happens, customers must not be pushed into a separate queue or made to describe their matter again. The entire conversation so far, including the intent the bot identified and the relevant context data, has to be available directly in the ticket system or in the support team's shared inbox.

• GDPR-compliant routing: processing and storage of conversations on servers within the European Union.
• No model training: contractual exclusion of any use of customer data to train external foundation models.
• Complete context transfer: handover of the chat history to human agents for seamless further handling.

Platforms such as ComLayer show how such an architecture works in practice: the system runs entirely in European data centres and answers customer questions exclusively from verified knowledge sources, with an exact source citation. Where a question cannot be evidenced beyond doubt from the documents, or where it touches sensitive topics such as fraud or advice, the AI hands the case over immediately, with a pre-drafted reply, to the team's shared inbox. Compliance, data protection and service quality stay controllable throughout.

Frequently asked questions