AI Support for Debt Rehabilitation: What to Test First

Generative AI could help a person organise financial information or prepare questions for a debt adviser. Whether it improves debt recovery or well-being is a separate question requiring evidence from the people and services concerned. A helpful conversation is not proof of a sustainable repayment plan.

What Finance Friend demonstrates

Toivonen, Salo-Lahti, Ranta and Haapio’s 2025 chapter on financial well-being and debt rehabilitation, in Generative AI, Contracts, Law and Design (Springer), describes Finance Friend, a GPT Builder prototype explored as a personalised coach. The public abstract presents possibilities and design concerns. It does not report a controlled trial establishing better repayment, reduced debt or improved mental health.

The prototype should therefore be treated as an exploratory design example. Claims about particular calculations or distress responses need the underlying test records; they cannot be inferred from the abstract.

Use behavioural science without promising an outcome

The original COM-B paper by Michie, van Stralen and West connects behaviour with capability, opportunity and motivation. It provides a way to organise intervention design. It is not evidence that an AI coach addressing those headings will work for over-indebted people.

Financial education can work. Kaiser and colleagues’ meta-analysis of randomised experiments found positive average effects on financial knowledge and behaviour. That finding does not establish lasting debt relief from a chatbot; interventions and outcomes need their own assessment.

The diagram is an illustrative arrangement proposed by Ada Studio. The components, their mapping to COM-B and the usefulness of several agents require testing. It is not Finance Friend’s verified architecture, an automated counselling service or a validated distress detector.

Start with bounded assistance

Possible tasks to evaluate include preparing a list of documents for an adviser, explaining a supplied term in accessible language and organising a user-confirmed budget. Keep the original document visible and check amounts, interest, dates and fees independently.

Do not let a draft explanation determine eligibility for insolvency procedures, the validity of a claim or a response deadline. Those questions depend on the jurisdiction, current law and individual circumstances. The legal routes in Switzerland, Germany and Finland cannot be treated as interchangeable.

Separate assistance from regulated decisions

Under Article 6(2) and Annex III, point 5(b), of the current EU AI Act, AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score are high-risk, except systems used to detect financial fraud. Following Regulation (EU) 2026/1744, the classification rules and the high-risk requirements and obligations (Chapter III, Sections 1 to 3) apply to Annex III systems from 2 December 2027 (Article 113). A budgeting explanation is not automatically that use. Assess intended purpose, deployment and applicable classification conditions, including Article 6(3), whose derogation never applies to an Annex III system that profiles natural persons; the name “coach” settles none of them.

Financial information can reveal highly private circumstances, but it is not automatically a special category under GDPR Article 9. Swiss law differs on one relevant point: data relating to social assistance measures are sensitive personal data under Article 5 letter c number 6 FADP. Identify the actual contents, the lawful basis or any justification required under Swiss law, recipients, retention and access permissions. The GDPR and, where applicable, Swiss data-protection requirements need separate assessment. Permission to access an account does not by itself establish permission for every downstream use.

Test with the people the service is meant to help

Ada Studio proposes an evaluation that checks comprehension, factual accuracy, accessibility, referral quality and whether suggestions are feasible for the person. Include missing information, irregular income and conflicting documents. Avoid measuring success only by engagement or how agreeable the conversation feels.