Neil Catton
AI in UK mortgages and housing finance: a briefing

The regulator’s model ranks borrowers better. Its own case study was declined by a filter.

74%higher arrears hazard above 90% loan-to-value

The FCA’s own research, June 2026

0.751how well the machine-learning model tells higher-risk households apart (AUC).0.669the conventional model, on around 3.1 million first-time-buyer households.74%higher arrears hazard above 90 per cent loan-to-value.

The human moment

A borrower who can afford the loan is declined automatically.

  1. Separated after prolonged economic abuse
  2. Stable work, able to repay
  3. An impaired credit history
  4. Automatically declined

The better model and the automatic decline are the same technology. What separates them is whether a person can still see, and change, what happens to the individual.

Why it matters now

From 1 January 2027 a valuation must be refreshed on an estimated 10 per cent fall.

And at least every five years, or three in some cases. The industry told the PRA that automated valuation models are routinely used.

£1.76tnmarket, growing again, with arrears falling.
5 Feb 2026a solely automated decline became lawful, with safeguards.
4 Nov 2024support for borrowers in difficulty had to be tailored to the individual.
The argument

Nobody has drawn the line between the two.

Both models are equally accurate on average. One is better at telling higher-risk households apart.

A borrower able to repay was still declined automatically.

If this is yours to answer for

Find out where a person can still see, and change, the decision.

  1. ReadSixteen pages, including seven things that can be done now, none needing new law.Download the PDF
  2. TalkOne conversation about your own lending, and whether the question applies to it.Book a meeting
  3. TestSix to eight weeks: a readiness assessment, use case triage and a regulatory and legal position paper. It fits before the Basel 3.1 valuation rules apply.How it works

Every statistic about AI-generated payslips and synthetic mortgage applicants traces to US trade press or vendor blogs, so all are excluded on the record, as are vendor accuracy claims for automated valuation. The FCA, PRA, Financial Ombudsman, Ministry of Justice, HM Land Registry and UK Finance carry the argument. Verified at 24 September 2026.

If you want the detail

What’s in the briefing

What’s inside the sixteen pages

Written for an executive who knows the sector well and AI not at all:

  • A note on the numbers: what is included, what was excluded and why
  • A £1.76 trillion market, growing again, with arrears falling
  • The pressure point: fewer possession cases, each one taking longer
  • The FCA’s model comparison and its automatic-decline case, read together
  • Four risks specific to mortgages, none of them about the technology working
  • Where AI actually is, and what the largest lenders do not say
  • The threat side: the title, the money at completion, and AI as the lock
  • The UK legal position, and the Basel 3.1 valuation change nobody has framed
  • Beyond the UK, stated as a comparator rather than an obligation
  • Seven things that can be done now, none needing new law
  • Three levels defined by where accountability sits, and three horizons
  • An eight-workstream way in

Questions people ask before reading

Is this vendor material?

No. Every figure traces to the FCA, the PRA, the Financial Ombudsman Service, the Ministry of Justice, HM Land Registry, UK Finance’s own data collection, the statute book, or a lender or professional body speaking on its own record. No vendor source appears anywhere in it.

Is it free to download?

Yes. It downloads directly, with no form and no email address required.

Does the PRA say lenders can use automated valuation models?

The briefing states only what the primary record supports: from 1 January 2027 a valuation must be refreshed on an estimated 10 per cent fall and at least every five years; the PRA has accepted desktop valuations since 2020; and the industry told the PRA that automated valuation models are routinely used and asked for them to count.

Does it cover Scotland and Northern Ireland?

Conduct regulation is UK-wide and the briefing covers it. The court possession statistics it uses are for England and Wales, and it says so, because Scotland and Northern Ireland run separate repossession law.

How current is the research?

Verified against primary sources as at 24 September 2026, including the FCA’s mortgage lending statistics for Q2 2026, UK Finance and Ministry of Justice figures published on 13 August 2026, and the FCA’s research note and CP26/18 of June 2026.

What if I want to take this further?

The briefing ends with an eight-workstream action plan. The first three, readiness assessment, use case triage and a regulatory and legal position paper, form a natural first engagement of six to eight weeks, which fits before the Basel 3.1 valuation rules apply on 1 January 2027.