Creditworthiness Assessment

The EU AI Act classifies it as high-risk. The FCA expects it to be fair. Consumer groups demand it be transparent. And every lender needs it to be accurate. Creditworthiness assessment sits at the intersection of commercial necessity and regulatory obligation, and AI is reshaping both sides of that equation.

What is creditworthiness assessment?

Creditworthiness assessment is the process of evaluating whether a borrower can afford to repay a loan and is likely to do so. It goes beyond credit scoring, which estimates default probability, to include affordability analysis: does the borrower's income, after existing commitments, leave sufficient margin to service the proposed debt? In the UK, the FCA's responsible lending rules require lenders to assess both willingness and ability to repay.

The inputs are familiar: income verification, existing debt obligations, credit bureau data, employment history, and living expenses. What has changed is the depth and breadth of data available. Open banking provides real-time transaction data that reveals actual spending patterns, not just declared income and commitments. This gives lenders a more accurate picture but also creates new obligations around data handling, consent, and data quality.

The commercial tension is between speed and rigour. A borrower applying for a point-of-sale loan expects a decision in seconds. A mortgage application requires detailed affordability analysis that traditionally takes days. AI models that can perform rigorous affordability assessment at speed resolve this tension, but only if the assessment is genuinely rigorous, not just fast.

The landscape

The EU AI Act classifies AI systems used for creditworthiness assessment as high-risk. From August 2026, these systems must meet requirements for data quality, transparency, human oversight, bias testing, and documentation. Lenders must provide meaningful explanations to applicants who are declined. The explanations must be specific enough that the applicant understands what drove the decision and, where relevant, what they could change.

The FCA's Consumer Duty reinforces the expectation that creditworthiness assessments deliver good outcomes. A lender whose AI model systematically overestimates affordability for a particular demographic, leading to higher default rates in that group, is not meeting its Consumer Duty obligations. The model's aggregate accuracy is not sufficient. Performance must be fair across segments.

Open banking is expanding the data available for assessment. PSD2 in the EU and the UK's open banking framework give lenders, with consumer consent, access to transaction-level data from the borrower's bank accounts. This data reveals income regularity, spending patterns, gambling activity, and existing commitments that may not appear on a credit bureau file. The challenge is integrating this data responsibly: transaction data is noisy, context-dependent, and potentially discriminatory if used naively.

How AI changes this

Cash flow-based affordability models analyse transaction data to build a detailed picture of income and expenditure. Rather than relying on declared income and standard expense assumptions, the model observes actual spending patterns. Several UK challenger banks use cash-flow affordability as their primary assessment tool, reporting default rates comparable to or better than traditional approaches while serving borrowers who would be invisible to bureau-based assessment.

Real-time decisioning enables creditworthiness assessment at the point of need. An AI model that can process an application, verify income against open banking data, assess affordability, and return a decision in under five seconds enables embedded lending in e-commerce, automotive, and point-of-sale contexts. The technology exists. The governance framework, ensuring that a five-second decision meets the same standard as a five-day decision, is the remaining challenge.

Continuous affordability monitoring extends assessment beyond origination. Rather than assessing affordability once at the point of lending, AI models track borrower behaviour throughout the loan's life. Early warning signals, reduced income, increased gambling activity, growing reliance on overdrafts, trigger proactive outreach before the borrower misses a payment. This supports both the lender's credit risk management and the FCA's expectations on customer vulnerability detection.

What to know before you start

Explainability is a legal requirement, not an architectural nice-to-have. Under the EU AI Act, declined applicants must receive meaningful explanations. Under UK GDPR Article 22, individuals have rights around significant automated decisions. Design the model for explainability from the start. Intrinsically interpretable models (logistic regression, gradient-boosted trees with constrained depth) satisfy this requirement more naturally than deep learning architectures.

Transaction data is powerful but dangerous. A model that identifies gambling transactions and penalises applicants may be accurate in predicting default risk, but it raises discrimination concerns (gambling addiction is a health condition) and data protection questions (is this a special category of personal data?). Define clear policies on which transaction categories the model may use, document the rationale, and test for disparate impact.

Bias testing must cover protected characteristics and their proxies. Postcode, occupation, transaction patterns, and device type can all correlate with ethnicity, gender, or disability status. Test the model's decisions across these dimensions and define acceptable disparity thresholds. The Equality Act applies to algorithmic decisions as much as human ones.

Start with augmenting existing decisioning, not replacing it. Use AI-based affordability assessment to approve applicants who would otherwise be declined due to thin bureau files. This extends credit access, generates new lending volume, and builds a performance dataset to validate the model. The risk is contained because you are approving marginal cases, not changing the entire underwriting process.