Accountable automation. The next mortgage operating model

The fastest mortgage journeys have changed what customers and brokers expect. In November 2025, Nationwide reported that more applicants were receiving mortgage offers within 20 minutes of starting their application, supported by automated income verification and automated valuations. Nationwide did not disclose the number or percentage of applicants achieving that turnaround. [16] The same lender statement identifies income verification as one contributor. It does not quantify how much of the reported speed comes from borrower data rather than property evidence. [16]

That is one lender’s reported experience for eligible cases. It is not a market average and it does not show that every mortgage, property or customer can follow the same route. The useful question for other lenders is where their own delay sits.

Automated valuation is part of the picture. LSL reports that automated valuation models account for about 25% of UK residential valuations, typically for lower loan-to-value lending. [12] The figure is a provider estimate, and choosing which cases suit an automated route remains a lender policy decision.

The market is large enough that operating-model choices matter. The FCA and Bank of England report £77.4 billion of gross residential advances in Q2 2026 alone. [1]

Speed for some cases raises expectations for all of them. When a customer or broker sees an offer in minutes for one application, a two-week wait for another feels like a failure even when the case is genuinely complex. The operating model must therefore explain why a case is taking longer, not only make easy cases faster.

The volume of repeat decisions is rising. UK Finance forecasts £300 billion of gross lending in 2026, including £77 billion of external remortgaging, alongside a separate £261 billion of product transfers excluded from gross lending. [27] These are forecasts rather than realised lending. Clean refinancing cases are a candidate starting point for controlled testing, not proof that most cases can be automated.

Where the time goes is the first analytical task. The government roadmap reports around 120 days from accepted offer to completion for purchases. [13] Only part of that sits with the lender, so an honest operating model separates lender-controlled time from chain, legal and valuation time before claiming any improvement.

Early AI in lending meant single tools, such as document extraction or a chat assistant. The next step is agent-like workflows that collect permitted information, compare sources, draft requests and prepare a case for review. That brings more value and more risk, because an agent can act across several steps before a person looks.

Agents should be given bounded work rather than authority.

Professional standards are moving the same way. The RICS global standard on responsible use of AI in surveying practice took effect on 9 March 2026, covering governance, professional judgement, transparency and client communication. [5] It places automated outputs inside a governed process with an accountable professional.

These are proposed operating boundaries for discussion, not a statement of what the law permits every lender to automate.

Adoption across financial services is already broad. The Bank of England and FCA report that 75% of surveyed firms were using AI in 2024, with a further 10% planning to within three years. 55% of use cases involved some automated decision-making but only 2% were fully autonomous. [20] Most firms are therefore already operating somewhere between assistance and autonomy, which is exactly where accountability is hardest to design.

The same survey found that only 34% of firms reported complete understanding of the AI they use, against 46% with partial understanding, largely because of third-party models. [20] For mortgage decisions that gap matters, because a reviewer cannot challenge what nobody can explain.

Automation changes the work people do rather than removing the need for them. Four roles become more important.

Brokers gain a clearer role too. One consolidated request and a realistic next-update date replace a sequence of piecemeal queries. Customers benefit through the lender and broker workflow rather than through a separate system they must learn.

Accenture reports that only one in ten UK organisations had scaled AI or embedded it into core operations, and 46% of executives reported little profit-and-loss impact so far. [10] Redesigning roles and decisions, rather than adding tools to an unchanged process, is one plausible reason some firms move further than others.

The roles need authority as well as titles. An evidence steward who cannot block an unpermitted source, or a model owner who cannot pause an agent, is a reporting line rather than a control. Each role should have a defined decision it can make, a defined escalation path and a record of when it acted.

The Bank of England and FCA survey reports that 84% of firms had a person accountable for AI. [20] The next step is accountability at the level of each decision pathway, so that a specific person owns how an agent is used in, for example, remortgage valuation routing.

Accountable automation does not need a new rulebook to begin. Several existing frameworks already describe what good looks like, although each has its own scope and none should be assumed to apply to every firm or case.

Taken together these point to the same design. Know which models and agents are in use, assign an owner, validate before reliance, monitor in use, explain outcomes and give people a route to human review. An operating model that records each of these steps as it happens will find assurance easier than one that reconstructs them later.

Third-party dependence is a growing part of the picture. The Bank of England and FCA report that a third of AI use cases were third-party implementations in 2024, up from 17% in 2022. [20] Lenders using external models or agents remain accountable for the outcome and should require version, validation and incident information from suppliers.

Human oversight can easily become a rubber stamp. Signs that control is real include the following.

Customer trust depends on this discipline. Deloitte reports that 63% of UK consumers had used a generative AI application by May 2026, yet 27% of AI users did not recognise the risk of factual inaccuracies. [6] EY reports that 74% of UK respondents had used AI recently but only 14% were comfortable relying on fully autonomous, agent-led systems. [9] These are surveys across AI contexts rather than mortgage-specific findings, but they point the same way. Our design hypothesis is that visible human accountability could improve trust. These surveys do not establish that effect in mortgage lending.

Explainability should work for three audiences. A customer needs a plain reason and a next step. An auditor needs to trace evidence, rules and reviewer decisions. A regulator needs to see that controls operated as designed. The record should answer all three without reconstruction.

Override data is one signal of oversight, not proof by itself. Low disagreement may reflect accurate outputs or a low-risk case mix, as well as weak review. Firms should interpret amendment rates alongside independent sampling, missed conflicts and case complexity rather than target a fixed rate.

If the controls hold, the shape of a lender could change over the rest of the decade. These are possibilities to debate, not predictions.

Each possibility carries a matching risk. Triage can entrench bias if poorly recorded properties always fall into slower routes. Fewer hand-offs can hide problems if the consolidated request omits a material conflict. Continuous monitoring can create noise that reviewers learn to ignore. The operating model should be tested against these failure modes as deliberately as against its benefits.

Policy direction supports earlier and more structured property information. The June 2026 government roadmap points towards upfront information and digital property logbooks, and OPDA reports strong consultation support for both. [13] [2] Lenders that design their operating model around governed evidence now would be better placed to use that information as it arrives, without assuming reform will deliver on a fixed timetable.

Public infrastructure is automating too. HM Land Registry plans to automate some high-volume applications, starting with residential mortgages. [31] If registration steps become faster, lender-side evidence and review will become a larger share of the remaining delay.

Supervised experimentation will shape how quickly this happens. The FCA reports that its second AI Live Testing cohort includes agentic AI use cases and firms such as Barclays, Experian and Lloyds Banking Group. [24] Lenders that can show controlled, well-documented testing are likely to find adoption easier than those that move from pilot to production without evidence.

Speed only counts as progress if customers are better served. Consumer Duty, where it applies, frames outcomes around products and services, price and value, consumer understanding and consumer support. [8] Its applicability depends on the product, activity and customer and should be assessed rather than assumed.

Questions worth testing for each journey.

Faster declines can be a good outcome when they are fair and explained, because they let customers seek other options without wasted cost. A faster pathway must also avoid excluding people whose properties or circumstances are unusual.

Customer trust in automation cannot be assumed. EY reports that only 14% of UK respondents were comfortable relying on fully autonomous agent-led systems. [9] A model in which people remain visibly accountable, and customers can ask for human review, is more likely to match what customers say they want.

A synthetic remortgage case shows how the pieces fit. The illustration uses invented facts and does not describe a real lender, customer or result.

The intended benefit is a clearer request and less avoidable waiting than a sequence of separate queries. No timing comparison has been measured for this invented case. Every step is attributable to a source, a rule or a named person, which is what makes the speed defensible.

An operating model should earn the right to scale. Agree measures and stop conditions before any change is tested.

Define the clock before the intervention, keep decision in principle, valuation turnaround and completion separate and report purchase, remortgage and product-transfer cohorts on their own. Use a staged design with independent quality review rather than a single headline figure.

The opportunity is significant and the path is practical. Start with one journey and synthetic cases, prove the controls, then test with authorised data. Property Spine currently demonstrates this approach with synthetic data only. Real benefits remain to be measured.

A simple scorecard keeps the evaluation honest.

A change that improves speed while worsening quality or outcomes should not scale, however attractive the headline.