UK specialist and commercial lending is under real margin pressure. Our own market analysis puts margins as having narrowed from the 5-6% many lenders were used to towards 3%, so businesses that used to compete on rate and risk appetite now also have to compete on cost and speed.
Development finance has become a bigger part of the picture too: it now accounts for £32bn of UK lenders’ outstanding loan books, 19% of total outstanding commercial real estate debt and 16% of all new lending, up from 15% the year before.
And the industry standard for a single drawdown – from request to funds released – still runs at 10 to 20 working days. Review time isn’t a footnote in that timeline. It’s often the single biggest variable in it.
That’s the backdrop against which a lot of the current excitement about agentic AI sits, and the scepticism too.
Financial institutions are being asked to balance speed and functionality against compliance and accuracy and many are finding it hard to get real value out of AI while deploying it safely.
The fear, reasonably, is that faster means looser.
Speed and compliance don’t have to be opposites though, as we found recently with a UK specialist bank working in property development finance.
TACKLING BANKING BOTTLENECKS
When a bank funds a construction project, every payout depends on a detailed Quantity Surveyor (QS) report.
These reports cover costs, timelines, financial viability and construction permits, and it’s the analysis of them that decides whether funding gets released. Given what’s riding on that decision, the review has to be careful. But being careful takes time.
We’ve seen banks spend up to an hour reviewing a single report. Multiply that across tens of reports a week and the effort adds up fast, along with the risk of something being missed through fatigue or inconsistency.
The knock-on effects are impactful. Delayed funding means materials that can’t be bought, subcontractors who don’t get paid, and projects that stall.
QS reports also arrive in wildly different formats from different surveyors, which makes the process harder to standardise and harder to evidence from a regulatory perspective.
TESTING WITH A DIFFERENT APPROACH
A UK specialist bank came to us with exactly this problem.
We built a proof of concept using a small number of purpose-built AI agents rather than one general model, each with a narrow job: one ingests the report, one extracts the relevant data, one checks it against the bank’s own rules and reconciles it against the underlying financial data.
A functional model was up and running within just five weeks. It cut review time from up to an hour down to five or six minutes, a reduction of up to 90%.
For borrowers and contractors, that’s the difference between funding on time and a project stalled waiting on a drawdown decision.
It also gave the bank something they didn’t have before: a consistent, structured output every time, regardless of how the original report was formatted.
That consistency is what makes the process auditable. You can show a regulator not just what decision was made, but why, and point to exactly where in the source document that decision came from.
Because the same checks are applied every time, the process also surfaces things a rushed manual pass might not: numbers that don’t reconcile across documents, a permit that’s missing, a timeline that doesn’t match the rest of the report. None of that replaces judgement, it just means less gets through by accident.
Built into the design from the start was a human checkpoint.
The system flags what it finds for a person to review, it doesn’t make the drawdown decision itself. That’s very deliberate.
The AI’s job is to do the reading, so the bank’s own experts can spend their time on judgement and exceptions rather than routine extraction.
When a decision genuinely matters, like whether a construction project gets its next stage of funding, a person needs to be the one accountable for it, not a model.
WHERE AI ACTUALLY HELPS
The lesson from this isn’t that AI should be doing more. It’s that there’s a specific, well-defined layer in lending operations, the document-heavy work between something arriving and a decision being made, where structured AI genuinely helps without taking the decision away from the people who own it.
In this example, applying more consistent, rule-based validation has ultimately created a stronger control environment for the client’s property development division.
Financial institutions are right to be cautious about AI – implemented poorly, it has the ability to increase risk, distrust, and damage compliance.
But done well, with solutions built specifically to target critical bottlenecks, it can quietly take the friction out of processes that have run the same manual way for years, freeing skilled people to spend their time where their judgement actually matters.
That’s probably the more useful lesson than the specific numbers.
QS report review just happened to be where we found it first.
Every bank we’ve spoken to since has pointed to a different process built the exact same way: paperwork arrives, it gets checked against policy, exceptions get flagged with the evidence attached, a person decides.
The bottleneck moves, the shape of the fix doesn’t.
