The FICO score was designed in 1989 for a world where credit assessment happened once — at application — and the primary data sources were thin credit file records accumulated over years of formal credit usage. A score between 300 and 850, derived from payment history, credit utilisation, account age, credit mix, and recent enquiries, remains the dominant underwriting input for most consumer and small business lending in the US and, through its equivalents, in the UK. The question we have been asking since 2019 is not whether this model is imperfect — it manifestly is — but whether the data infrastructure to replace it has now matured to the point where AI-driven underwriting can outperform it consistently in commercially tractable cohorts.
The UK's Open Banking framework has created something genuinely novel: a standardised, consented mechanism for a lender to see twelve months of a borrower's transactional history in real time, including payroll inflows, rent payments, subscription commitments, and the texture of spending behaviour. This is categorically different from a bureau file. A traditional lender sees whether a mortgage was repaid on time; an Open Banking-enabled lender can see whether the borrower's account shows a consistent rent payment every fourth Friday, whether their business revenue is growing, whether they are drawing down their account balance to zero each month, and whether a recent credit card application was followed by elevated spending. These signals — individually weak, collectively powerful — are what machine learning models process well and rule-based scorecards cannot.
We are not claiming that AI underwriting eliminates credit risk or that bureau data becomes irrelevant. The bureau file captures credit history that predates Open Banking consents; it captures behaviour across accounts the borrower did not consent to share; and in the event of a dispute, it provides a documented audit trail that regulators expect. The most effective AI-native credit models we have evaluated use bureau data as a baseline and Open Banking transaction history as a real-time enrichment — with the ML model trained to weight each signal by its predictive contribution within a defined cohort. The AI is not replacing the bureau; it is solving the thin-file problem that the bureau cannot.
The commercially significant opportunity is in the segments the traditional scorecard structurally underserves. Self-employed individuals with variable income but strong underlying cash flow have been systematically excluded from prime lending — not because they are poor credit risks, but because their payslip-based income verification fails the standard affordability assessment. Recent graduates with no credit history are treated identically to individuals who have defaulted, simply because neither has a bureau score. Immigrants whose credit history did not transfer across borders start from scratch regardless of their financial standing at home. In all three cases, transactional data tells a materially different story than the bureau file, and AI models trained on consented Open Banking data can quantify the difference in credit-adjusted expected loss terms. That is the commercial case for AI credit decisioning — not a general claim about AI, but a specific answer to a structural gap in the existing underwriting architecture.
The regulatory dimension is material and should not be treated as a later-stage concern. The FCA's approach to algorithmic fairness in lending — embedded in the Consumer Duty requirements and the Equality Act 2010 — requires that AI underwriting models be explainable in terms the customer can understand, that they do not produce discriminatory outcomes across protected characteristics, and that they do not make decisions solely on automated grounds without a human review mechanism. Designing this in from the beginning is an architectural requirement, not a compliance afterthought. The lenders who treat explainability as a constraint will produce worse models; those who treat it as a design principle will produce models that are both fairer and more defensible in supervisory review. Our investment thesis on AI lending infrastructure is built on this distinction.