How OakNorth Rebuilt Underwriting Around AI

In underwriting, every extra loan adds analyst time and, past a point, thins credit quality, so a lender that grows its book usually grows its cost base with it. OakNorth rebuilt the one workflow that sets those economics, the credit underwriting decision, and has grown its book while cutting the cost of running it and holding cumulative losses to a fraction of a percent.

For a lender, the margin lives in credit quality and cost

A lender's return does not come from cheap funding. It comes from the quality of the credit it writes and the cost of producing that credit, and both are set inside the underwriting decision.

OakNorth's 2025 results show what that looks like when it goes right. Gross originations rose 33% to £2.8bn, pre-tax profit reached £223m, and total facilities grew 18% to £7.2bn. [1] Over the same year the efficiency ratio, the cost of running the bank against its income, fell from 29% to 26%, a 10% improvement in operating leverage that OakNorth attributes partly to a focus on AI. [1] Secondary reporting on the same results describes the bank expanding its lending without increasing headcount. [2]

The timing is worth noticing. Gross bank lending to smaller UK businesses rose 9% to £68bn in 2025, and challenger and specialist banks accounted for 60% of it, up from 39% in 2012. [3] The growth in SME credit is coming from lenders built differently, not from the largest incumbents. The question for any of them is which decision to rebuild first.

Underwriting an SME is a thin-data judgement

Mid-market credit is hard because the information base is structurally thin. A mid-sized borrower does not come with the standardised disclosure, continuous market pricing, and rating-agency coverage that make large-corporate credit legible.

The risk is idiosyncratic and subsector-specific. A generic scorecard flattens exactly the variation that decides whether a loan is repaid, because a haulage firm, a care home, and a software business fail for different reasons and on different signals.

That has a direct operating consequence. Improve the judgement structurally and the whole book improves, because every loan is written against a better view. Leave the judgement manual and every new loan adds analyst cost that scales with the book, so growth and efficiency fight each other. The lever is the analysis step inside the decision, which is exactly where OakNorth went to work.

This is single-workflow redesign, not a tool

OakNorth did not spread an AI tool thinly across its teams. It rebuilt one high-stakes decision around a machine learning model, a statistical system trained on data rather than programmed with rules, and left the authority where it was.

The engine is ON Credit Intelligence (ONCI), built first to steer OakNorth's own loan book and now licensed to other banks. [4] It produces a forward-looking, subsector-specific, scenario-based view of a borrower, underpinned by 262 sub-sector models, and calculates default probabilities and forward-looking ratings at both borrower and portfolio level. [4][2] The analysis is decision support. The bankers and the credit committee decide whether to lend.

This is the Intelligent Ops argument in practice: redesign the operating model, do not add a tool on top of the old one. [5] Workflow redesign has the largest measurable relationship to EBIT impact from generative AI, larger than the AI tools themselves, yet only 21% of organisations say they have fundamentally redesigned even some workflows. [5] Adding a copilot to an existing process is a technology decision; changing the workflow around the decision is an operating-model decision. Rebuild the single decision that sets the economics and the capability compounds across every loan written after it.

Caption: OakNorth put AI on the analysis and kept the banker on the decision; ONCI produces the forward-looking view, the credit committee keeps the authority to lend, and the monitoring loop feeds the same engine back into the portfolio.

The number it moved

Across the book, the pattern every underwriting business wants shows up together: the book grew, the cost of producing it fell, and losses stayed low. OakNorth reports cumulative principal losses of £6.9m on £15.1bn of facilities granted since inception in 2015, about 0.045% over ten years. [1]

Two things keep that figure honest. It is cumulative, not annual, and its denominator is everything OakNorth has ever lent since 2015, not the current £7.2bn book. [1] It is also OakNorth's own characterisation of its credit performance, not an independently audited split that isolates AI from underwriting discipline, secured lower-mid-market lending, book vintage, and a mostly benign macro over the period.

Even with those qualifications, the shape holds. A book growing 18% in a year, an efficiency ratio falling three points, and a near-nil cumulative loss rate is the combination a lender is meant to find hard to hit at once. Where those three move together, the underwriting decision is doing the work.

How to run the same play without being a purpose-built fintech

The proof that this is a workflow and not a founder's advantage is that OakNorth sells the same engine to banks that predate it. PNC, one of the largest US commercial banks, uses ONCI for predictive forecasting, early-warning indicators, and integrated workflows that surface borrower deterioration earlier. [4] Earlier reporting named Capital One and Fifth Third as customers as well. [6]

An incumbent can run the same play in four moves. Isolate one high-stakes decision workflow. Rebuild its analysis step around AI. Keep the accountable human on the decision. Measure the credit and efficiency outcome on that one workflow before scaling it anywhere else.

One control line comes before the first workflow ships. The credit committee keeps the authority to lend, so a model-driven view never becomes a model-made decision. That is the operating principle OakNorth's design reflects, and it is what keeps "the model flagged it" from quietly becoming "the model decided."

Key Takeaways

• Pick one high-stakes decision workflow and rebuild its analysis step around AI, rather than spreading a tool thinly. The compounding comes from redesigning a single decision, as OakNorth did with credit underwriting, not from making everyone faster.
• Keep the accountable human on the decision and put AI on the analysis. OakNorth's credit committee retains the authority to lend, so accountability for the lend stays with a named person rather than the model.
• Measure the credit and efficiency outcome on the one workflow before scaling. The proof it transfers is that OakNorth runs the same engine inside incumbent banks; the full design method is in the Intelligent Ops paper.

Sources

[1] OakNorth. "OakNorth drives 33% increase in gross originations in 2025, delivering £223M in pre-tax profits." Press release, 17 March 2026. https://oaknorth.co.uk/press/oaknorth-2025-annual-results/

[2] Seoul Economic Daily. "UK's OakNorth Cuts Costs, Expands Lending with AI Credit Model." Seoul Economic Daily, 10 May 2026. https://en.sedaily.com/finance/2026/05/10/uks-oaknorth-cuts-cir-despite-loan-growth-through-ai

[3] British Business Bank. "Small Business Finance Markets 2026." British Business Bank, March 2026. https://www.british-business-bank.co.uk/about/research-and-publications/small-business-finance-markets-report-2026

[4] OakNorth Credit Intelligence (ONCI). Product and customer pages, accessed 9 July 2026. https://www.onci.com/

[5] Christopher Hughes. "Intelligent Ops: Why AI Agents Are Rewiring the Way Businesses Run." cgh.dev, 12 May 2026. https://cgh.dev/thinking/intelligent-ops/ (citing McKinsey & Company, "The State of AI: How Organizations Are Rewiring to Capture Value," March 2025)

[6] The Financial Brand (reposted by ONCI). "Inside the Bank Selling Its Own Fintech Software." The Financial Brand, June 2022. https://www.onci.com/financial-brand-june-2022