Machine learning can combine credit-bureau history with live cash-flow signals to find borrowers a three-digit score misses. The likely future is not scoreless lending, but a contest over which data, models and explanations deserve to shape access to credit.
For decades, consumer lending has been organised around an elegant compression exercise: turn a messy financial history into a standardised score, then use that score to help decide whether a borrower qualifies and at what price. Artificial intelligence is breaking open that compression. Instead of asking only what a borrower did with past credit, lenders can increasingly ask what money is flowing into an account now, how volatile balances are, whether bills are paid consistently, how income behaves through the month and how combinations of signals interact in ways a conventional scorecard may miss.
The temptation is to call this the end of the credit score. The evidence points to something subtler. In the United States, the Federal Reserve’s October 2025 review of alternative data estimates that roughly 32 million adults are “unscoreable” under traditional approaches—about 7 million credit-invisible adults and 25 million with thin files. At the same time, the US mortgage system is not abandoning scores: in April 2026, the Federal Housing Finance Agency announced a new phase of credit-score competition involving VantageScore 4.0 and FICO 10T alongside the transition away from decades of reliance on older models.
That is the central tension. AI can make underwriting more granular, dynamic and inclusive. But lending is not merely a prediction problem. It is also a legal decision, a consumer-protection decision, a balance-sheet decision and, for banks, a model-governance decision. A model that predicts default slightly better may still be unusable if the institution cannot explain it, validate it, control its data, monitor drift or defend the result to a regulator and a rejected applicant.
ANALYTICAL THESIS
AI is not ending the credit score so much as changing its job. The score is becoming a benchmark inside a broader decision system that combines bureau history, cash flow and lender-specific models. The competitive question is which institutions can add information without losing accountability.
The evidence in four numbers
| Signal | Current evidence | Why it matters |
| US adults unscoreable | ≈32 million | Alternative data can potentially evaluate people whom conventional score requirements exclude. |
| FinRegLab approval simulation | ≈4% higher | The strongest ML models approved more borrowers than comparable simpler models at mainstream risk thresholds while also reducing approvals among borrowers who later struggled. |
| US mortgage scoring shift | 2026 rollout | A major market is modernising and adding score competition rather than eliminating scores. |
| EU high-risk AI timetable | 2 Dec 2027 | AI systems used to evaluate natural persons’ creditworthiness are in Annex III; the high-risk obligations now apply later under the revised timeline. |
From a three-digit score to a decision stack
A traditional credit score is powerful because it is standardised. It compresses a set of bureau variables—payment history, debt use, age and mix of accounts, and recent credit activity—into a portable signal that lenders, investors and risk teams can understand. Its weakness is the same as its strength: compression throws information away, and it depends on a borrower having enough reportable credit history to be scored in the first place.
The emerging alternative is not simply “more data”. It is a layered decision stack. The Federal Reserve’s alternative-data analysis shows how cash-flow variables can map to familiar risk concepts: overdraft history can resemble delinquency information; regular deposits and average balances can provide information about liquidity; account tenure can resemble the length of a credit history. These are not exotic behavioural proxies. They are financial facts that may reveal repayment capacity more directly, particularly when a borrower has limited bureau history.
AI changes the economics of using those signals. A human underwriter could always read bank statements, but doing so for a small unsecured loan may cost too much relative to the revenue on the account. Machine learning can classify transactions, detect patterns, combine variables and make the process fast enough for high-volume lending. The technology therefore expands not only what can be measured, but what can be measured economically.
The most important gain may come from the model, not the exotic data
The strongest recent US evidence complicates the popular story that alternative data alone is the revolution. FinRegLab’s 2025 empirical study of machine learning and cash-flow underwriting built models using traditional credit-bureau data, cash-flow data and combinations of the two, then compared predictions with actual performance on new accounts. Machine-learning models outperformed traditional logistic-regression approaches across the data configurations tested. The model combining bureau and cash-flow information was the most predictive overall, while the machine-learning model using bureau data alone ranked second.
In simulations at risk cut-offs used by mainstream lenders, the two strongest machine-learning models increased approvals by about 4% compared with similar models using simpler analytics, while also substantially reducing approvals among consumers who later struggled with their loans. That is meaningful evidence that better analytics can move the frontier between access and loss.
But it should not be overgeneralised. The study is an empirical comparison on particular datasets and historical account performance, not proof that any AI model will widen access or improve fairness. Model design, data coverage, economic conditions, product type and the lender’s chosen risk threshold all matter. AI can sharpen a signal; it can also sharpen a mistake.
ANALYTICAL INFERENCE
The likely winner is “score-plus” underwriting: a standardised bureau score or score-like benchmark remains useful for portability and governance, while AI adds cash-flow and other permissioned signals around it. In markets with weak bureau coverage, the balance can tilt further toward alternative data; in highly securitised markets, standard scores are harder to displace.
Why the credit score is unlikely to disappear
The simplest reason is institutional. Credit scores do more than rank borrowers. They create a common language across originators, insurers, investors, securitisation structures, pricing systems and risk committees. Replacing that common language with a proprietary model that produces a probability of default from hundreds or thousands of features can improve local prediction but make comparison harder across lenders and portfolios.
The US mortgage market illustrates the point. FHFA’s April 2026 announcement said the Federal Housing Administration would permit VantageScore 4.0 and FICO 10T for FHA-insured mortgage underwriting, while Fannie Mae and Freddie Mac were moving forward with the new models and immediately accepting Vantage-scored loans from approved lenders. The direction of travel is competition among more modern scoring models, not the disappearance of scoring.
There is also a governance advantage to a benchmark that can be independently reproduced and monitored. A lender can use an internal AI model as a challenger or overlay while still reporting a familiar score to credit committees and secondary-market counterparties. The score may lose some of its authority at the point of decision, but retain value as a reference point, a control variable and a market standard.
AI sees borrowers the score cannot—but every new signal creates a new governance problem
The inclusion case is strongest where a borrower is financially active but credit-file poor: a young adult, a recent migrant, a first-time borrower, a small entrepreneur or a household that uses debit rather than revolving credit. Cash-flow data can reveal stable income, regular rent, controlled spending and resilience through a temporary shock even when the bureau file is sparse.
The global opportunity is even larger in markets where formal credit reporting covers only part of the population. The World Bank’s digital-credit guidance describes how algorithmic scoring can use alternative data in automated lending, while also warning that consumer harms can emerge when digital credit grows faster than governance. The same data richness that helps include a thin-file borrower can become intrusive if lenders reach into behavioural, device or platform data that have only a weak connection to repayment capacity.
That distinction matters. Permissioned financial data—income, balances, recurring obligations, overdrafts—has an intuitive relationship to ability to repay. Non-financial data such as device characteristics, location patterns or online behaviour may be predictive in a statistical sense but harder to justify to customers, regulators and boards. It can also encode socioeconomic proxies that are difficult to detect. The strategic question for lenders is therefore not “How much data can we ingest?” but “Which data can we defend?”
Explainability is becoming a hard design constraint
In the United States, the current Regulation B notification rule requires a creditor taking adverse action to provide specific principal reasons. It expressly says that stating that an applicant failed to achieve a qualifying score is insufficient, and the official interpretation requires disclosed reasons to relate to factors actually considered or scored. That makes explanation architecture part of underwriting architecture: if the lender cannot reliably translate the model’s decision into the actual principal reasons, the model creates a compliance problem regardless of its predictive power.
The policy backdrop has moved. The CFPB withdrew its 2022 and 2023 circulars on complex-algorithm adverse-action notices in May 2025, while the underlying Regulation B reason-giving requirement remains in force. The Bureau also amended Regulation B in April 2026 to state that ECOA does not recognise disparate-impact liability. For multinational lenders, that does not settle the fairness question: intentional discrimination prohibitions, data and reporting rules, state laws, prudential governance and non-US regimes still create a broader control environment.
Europe is moving in the opposite direction on AI-specific classification. The EU AI Act’s Annex III classifies systems intended to evaluate the creditworthiness of natural persons or establish their credit score as high-risk, subject to the Act’s exceptions. The timetable itself changed: under the current EU implementation timeline following the Digital Omnibus changes, Annex III high-risk obligations are scheduled to apply from 2 December 2027 rather than August 2026. That delay gives firms more time, but not a reason to postpone model inventories, documentation, human-oversight design and data-governance work.
Supervisors are preparing to audit the model, not just the outcome
Banks have long managed credit models, but AI makes traditional model-risk disciplines more demanding. Validation must cover not only predictive performance but stability, drift, data lineage, feature behaviour, segmentation, overrides and the effect of updates. Third-party models add another layer because the bank may depend on a vendor’s intellectual property, data transformations or infrastructure without seeing every internal detail.
The US prudential agencies’ revised 2026 model-risk management guidance emphasises a risk-based approach covering model development and use, validation and monitoring, governance and controls, and vendor or third-party products. Notably, the guidance says generative AI and agentic AI models are too novel and rapidly evolving to be within its scope, and the agencies plan further work that specifically considers AI. That is a useful reminder that regulation is still catching up to the technology stack.
Supervisors are also building their own technical capacity. BIS Innovation Hub’s Project Noor is developing tools that use explainable-AI techniques to help supervisors interpret models, assess fairness and test robustness. The implication for banks is significant: model governance is shifting from “show us your policy” toward “show us how the model behaves.”
The hype is ahead of production lending
Generative AI dominates the public conversation, but much of the AI actually used in credit decisions is older and more structured. The Bank of England’s 2025 assessment of AI in the financial system said use of AI in credit-risk management was still in its infancy in aggregate, although some firms were using established techniques such as gradient-boosted decision trees in pre-screening, application scoring, pricing and provisioning. That matters because the immediate transformation is less about a chatbot deciding whether to grant a mortgage and more about machine-learning models quietly improving—or complicating—the existing underwriting engine.
The next phase could be more radical. An AI system could retrieve permissioned data, verify documents, challenge anomalies, select a model, propose terms and generate the adverse-action explanation as one workflow. But autonomy raises a new control question: if several models and agents contribute to the decision, which component owns the reason? Banks will need traceability across the entire decision chain, not merely an explanation of the final score.
Human judgment will not vanish; it will move to the edges
AI is strongest where information is hard, codifiable and repeatable. Human lenders remain valuable where information is soft, contextual and exceptional: a temporary disruption in cash flow, a family business restructuring, a new industry without long performance history, or a customer whose financial statements do not capture the economics of a relationship.
A 2025 BIS working paper on AI and relationship lending found evidence from Italian banks that AI-based screening and monitoring can coexist with relationship lending and improve some normal-times outcomes, but during the Covid shock AI did not provide additional credit or interest-rate protection. The lesson is not that AI fails in crises; it is that model performance is regime-dependent, and the soft information embedded in relationships can remain valuable when the future stops looking like the training data.
That is also the governance principle articulated by the Reserve Bank of India’s deputy governor in an April 2026 speech on AI in finance: AI can improve credit delivery, especially for borrowers poorly served by standard templates, but accountability remains with the institution and explanations must make sense to customers, managers and supervisors. The safest operating model is therefore not “human versus machine”. It is a deliberate allocation of tasks, with humans accountable for the system and empowered to investigate exceptions rather than rubber-stamp automated outputs.
What changes for banks, fintechs, regulators and investors
For banks, the competitive advantage will come from the quality of the decision system rather than from owning the most exotic algorithm. Clean consented data, reliable reason codes, disciplined model monitoring, fast manual-review paths and the ability to compare AI decisions with conventional benchmarks may matter more than a marginal lift in model accuracy. Banks with rich deposit relationships have a natural advantage because they already observe cash flow—provided customers permit its use and governance is sound.
For fintech lenders, AI can remain an entry strategy because it reduces the disadvantage of having less historical relationship data. But the bar is rising. As banks adopt similar techniques and open-finance infrastructure improves, a proprietary model is less defensible if competitors can access comparable data. The durable moat may shift toward distribution, data permissions, servicing performance and the institution’s ability to validate models across cycles.
For regulators, the problem is no longer whether to allow algorithms into credit. They are already there. The harder task is to create rules that preserve explainability and accountability without freezing the market around yesterday’s model forms. Supervisors will need technical tools, common terminology and proportionate expectations that distinguish a transparent cash-flow model from an opaque system built on weakly related behavioural proxies.
For investors, better underwriting could support growth by approving more genuinely creditworthy borrowers without a corresponding rise in defaults. But model risk can also become concentration risk: if many lenders buy similar models, use similar data aggregators or optimise to similar historical patterns, errors can correlate. The institution with the highest apparent approval lift may not have the best risk-adjusted economics if the model is fragile, costly to explain or vulnerable to a shift in borrower behaviour.
Conclusion: the score becomes a benchmark, not a verdict
The credit score is unlikely to disappear because finance values standardisation almost as much as prediction. What AI is ending is the idea that one three-digit number should carry the full burden of a lending decision. Bureau scores will increasingly sit beside cash-flow signals, transaction patterns, document intelligence and lender-specific models that can distinguish borrowers who look similar under older systems.
That can widen access to credit and improve risk selection. It can also make decisions harder to audit, challenge and explain. The institutions that gain most from AI underwriting will therefore be those that treat accuracy, inclusion, explainability and governance as one engineering problem rather than four competing objectives. The future of lending may be more intelligent than the credit score. It will also have to be more accountable.
References
1. Federal Reserve Board — Alternative Data: Expanding Access to Credit (Oct 2025)
2. FinRegLab — Machine Learning & Cash Flow Data in Consumer Underwriting (Jul 2025)
3. FHFA — Homebuying Advances into New Era of Credit Score Competition (Apr 2026)
4. CFPB — Regulation B § 1002.9 Notifications (current rule and interpretation)
5. CFPB — Withdrawn Guidance, including AI adverse-action circulars (May 2025)
6. European Commission — AI Act Annex III: High-Risk AI Systems
7. European Commission — EU AI Act Implementation Timeline (current)
8. OCC — Model Risk Management: Revised Guidance, Bulletin 2026-13
9. Bank of England — Artificial Intelligence in the Financial System (Apr 2025)
10. BIS Innovation Hub — Project Noor: Explaining AI Models for Supervision (Aug 2025)
11. BIS — Artificial Intelligence and Relationship Lending, Working Paper 1244 (Feb 2025)
12. Reserve Bank of India / BIS — AI in Finance: What Can Change, What Must Never Change (Apr 2026)
13. World Bank — Digital Finance Inclusion: Screening and Approving the Customer