AI Underwriting Is Cutting Loan Approvals from Days to Seconds
Machine learning models trained on alternative data are reshaping credit decisioning — and the big banks are scrambling to keep up.
The traditional loan application process — submit documents, wait three to five business days, receive a decision — is being quietly dismantled. Across the lending industry, AI-powered underwriting systems are compressing approval timelines to under 30 seconds, while simultaneously improving accuracy and reducing default rates.
Companies like Upstart, Zest AI, and a wave of European challengers have moved beyond simple credit score lookups. Their models ingest hundreds of data points: bank transaction patterns, cash flow volatility, professional history, even the stability of a borrower's address history.
The data advantage
The most valuable signal is not how much someone earns — it is how consistently they manage cash flow month to month. A borrower who earns £2,400/month and spends £2,395 represents a fundamentally different risk profile to one earning the same with £400 consistently left over.
Traditional credit bureaus cannot capture this granularity. Open banking APIs can. In the UK, where the number of consumers and small businesses regularly using open banking passed ten million in 2024, a lender can read twelve months of categorised transaction history — with the applicant's consent — in the time it takes the application form to load. That is where the competitive moat for AI lenders is actually being built: not in the cleverness of the model, but in the freshness and depth of the data it is permitted to see.
How the models actually work
Most production underwriting models are not the deep neural networks of popular imagination. Lenders overwhelmingly favour gradient-boosted decision trees — XGBoost and its relatives — because they handle messy tabular data well and, crucially, because their outputs can be explained. Explainability is not a nice-to-have: in most jurisdictions a lender is legally required to tell a declined applicant why — in the US, Regulation B requires the specific principal reasons. Attribution techniques such as SHAP values let a lender translate a model score into the adverse-action reasons regulators expect — "income volatility too high relative to the requested amount", not "the algorithm said no".
The second unglamorous truth is that the algorithm is rarely the differentiator. Feature engineering — turning raw bank transactions into signals like days-in-overdraft per quarter, gambling-spend ratio, or income-source concentration — is where the proprietary value sits. Two lenders running the same algorithm over differently engineered features will write materially different loan books.
What the evidence shows so far
Upstart's published results claim its model approves materially more borrowers at lower APRs than a traditional scorecard, but the more interesting effect shows up at the margins of the credit spectrum: thin-file applicants — younger borrowers, recent arrivals, the self-employed — who are nearly invisible to bureau-based scoring yet perfectly legible to cash-flow underwriting. The expansion of access is real, which is a large part of why regulators have engaged with the technology rather than simply resisting it.
It is not a free lunch. Models trained on benign economic conditions can degrade sharply when conditions turn, and several AI lenders learned during the 2022–23 rate cycle that "the model has never seen a downturn" is a risk disclosure, not a footnote. Continuous monitoring for drift, with challenger models running in parallel, is now table stakes for anyone serious.
| Dimension | Traditional scorecard | AI cash-flow underwriting |
|---|---|---|
| Primary data | Bureau score, declared income | Bureau data plus open-banking transaction history |
| Decision time | Days, often with manual review | Seconds for prime applicants |
| Thin-file applicants | Frequently declined for lack of history | Assessable from live cash-flow patterns |
| Model type | Fixed rules / logistic scorecards | Gradient-boosted trees with SHAP attribution |
| Main failure mode | Misses creditworthy thin files | Drift when conditions leave the training distribution |
The regulatory squeeze
Credit scoring is classified as a high-risk use case under the EU AI Act, bringing documentation, human-oversight and data-governance obligations that many fintech lenders are still operationalising. In the US, fair-lending law requires lenders to demonstrate their models do not produce disparate impact across protected groups — a genuinely difficult statistical problem when a model ingests hundreds of correlated features. In the UK, the FCA's Consumer Duty adds an outcomes lens on top. The lenders treating model governance as core infrastructure rather than compliance theatre are the ones winning bank partnerships.
Where it spreads next
Consumer credit was the proving ground; the same machinery is now moving up-market. SME lending is the obvious frontier — small-business underwriting has always been too expensive to do well manually, which is why banks historically either declined the segment or priced it punitively. Cash-flow models change that equation: a lender reading a business's live banking and accounting data can underwrite a £50,000 working-capital facility profitably at volumes no human credit team could touch. Embedded lenders inside accounting and commerce platforms are pressing exactly this advantage.
Insurance underwriting and mortgage decisioning are following the same path, a few years behind. Mortgages are the hardest case — larger sums, longer horizons, more regulation — and most "instant" mortgage decisions today are instant agreements in principle, with the full underwrite still partly manual. The direction of travel, though, is identical: every credit decision that can be made from verified data streams eventually will be.
What this means for incumbents
High street banks are in a difficult position. Their legacy core systems were not designed for real-time ML inference at scale. Modernising them is a multi-year, multi-hundred-million-pound project. In the meantime, challenger lenders are eating the most profitable slice of the market: prime borrowers who value speed over brand.
The pragmatic response, now visible across most tier-1 banks, is to buy rather than build: partner with or acquire the AI underwriting layer, run it in shadow mode alongside the incumbent scorecard, and migrate product by product as confidence grows. Where AI is landing across the rest of the banking stack — fraud, servicing, compliance — is mapped in our complete guide to AI in financial services in 2026.
The end state is not in much doubt: sub-minute decisions on the majority of consumer credit, with human underwriters reserved for genuinely complex cases. The open questions are who owns the customer relationship when the decision happens inside someone else's checkout flow — and whether incumbents can modernise before that question gets answered for them.
Frequently asked questions
What is AI underwriting?
AI underwriting is the use of machine learning models to assess creditworthiness, replacing or augmenting traditional rule-based scorecards. Models ingest hundreds of data points — including bank transaction patterns, cash flow volatility, and employment stability — and produce a decision in seconds rather than days.
Is AI-driven credit scoring more accurate than traditional methods?
Early data from lenders such as Upstart and Zest AI suggests AI models can reduce default rates while approving a broader pool of borrowers, particularly thin-file applicants. Accuracy depends heavily on training data quality and ongoing monitoring for bias and drift.
How quickly can AI underwriting approve a loan?
Many AI-driven lenders deliver decisions in under 30 seconds for prime borrowers, compared with the three-to-five day turnaround typical of legacy underwriting. Complex cases requiring manual review still take longer, but the baseline has shifted.
What data do AI underwriting models use?
Beyond traditional bureau data, models ingest open banking transaction history (income stability, spending patterns, overdraft usage), employment and income verification, and application metadata. Cash-flow features derived from bank transactions are consistently the strongest predictors — which is why open banking access matters more than algorithm choice.
Do AI lending decisions have to be explainable?
Yes. Most jurisdictions require lenders to give declined applicants specific reasons, and credit scoring is a high-risk use case under the EU AI Act. In practice lenders use interpretable model families (gradient-boosted trees) with attribution techniques such as SHAP to generate compliant adverse-action reasons.