What to Build First, What to Leave Alone

Seventy-five per cent of UK financial firms are already using AI in some form. That figure, from the Bank of England and FCA’s 2024 survey, is up from 58 per cent two years earlier. Inside most mortgage lenders, the conversation has moved on from whether to use AI to what to deploy next.

That second question is harder than it looks.

Internal AI, done well, releases real value inside today’s estate. Lloyds has compressed mortgage income verification from days to seconds using its Vertex AI platform. Nationwide has worked with Experian to automate employment and income checks, removing the need for applicants to submit payslips manually. Both are examples of AI doing useful work on the internal horizon, without waiting for Smart Data, digital verification services, or the wider ecosystem to mature.

But the internal horizon is also where most AI programmes quietly stall.

The AI use cases that consistently deliver return on investment inside a lender’s existing estate share a pattern. They operate on narrow, well-defined inputs. They produce narrow, well-defined outputs. They sit at points in the journey where the data is already structured enough to be trustworthy.

Income and employment verification, as Lloyds and Nationwide demonstrate, works because the inputs are standardised and the output is a simple yes or no that a human can check quickly.

Document classification, covering application packs, identity evidence and supporting documents, works because the documents themselves carry the structure. The AI is not being asked to reason about the mortgage journey. It is being asked to read what is on the page.

Case summarisation for underwriters works because the input is a case file that already exists, and the output is a summary that a human reviews before acting on it. The AI is an assistant, not a decision-maker.

None of this is transformation. It is applied productivity. But the real return on investment is not in any single deployment. It is in the AI maturity a lender builds while deploying it. Each use case teaches teams what AI can and cannot do in a live mortgage journey. It adds to the governance patterns harder work will eventually demand, and it produces the evidence that makes bigger AI bets defensible to a board. This is where maturity compounds. Lenders who build it now will be ready for what is coming next. AI-first lending is where this industry is heading, and the capability to operate that way is built one use case at a time.

The pattern of AI programmes that stall is equally consistent. They tend to be aimed at higher-value targets, such as predictive fall-through modelling, early affordability triage, or partial decisioning. These are use cases where the theoretical return is larger and the technical problem is harder.

These use cases need things the organisation is still building. They need consistent data definitions across teams, stable workflows, and a clear audit trail from input to decision. Most lenders have those ambitions on a roadmap. Few have them in production yet.

The model itself is rarely the issue. The issue is that a fall-through prediction model trained on inconsistent case-stage definitions produces unreliable outputs, and an affordability triage model running on ambiguous income categories makes decisions the organisation cannot defend.

The Bank of England and FCA survey notes that only 34 per cent of firms claim complete understanding of the AI they use. Forty-six per cent describe their understanding as partial. That gap widens quickly when the underlying data is ambiguous.

Most lenders now accept that AI has a real role inside their existing estate. The harder call is where to start, and what to hold back until the data and workflow work has caught up.

The investments that deliver ROI earliest are the ones that do not require the organisation to be better than it already is. They operate on clean inputs at specific points in the journey, and they release time and accuracy into processes that were already manual. That is where lenders build capability.

The investments that deliver ROI later, such as predictive modelling, decisioning support, or orchestration across multiple systems, depend on data consistency, workflow stability, and governance the organisation is still putting in place. Deploying them too early turns the model into a liability.

This is why tools like Pathfinder by Novus matter. They let lenders model the impact of a specific AI use case on speed to offer, speed to completion or fall-through before committing the investment. That removes the guesswork. Leadership teams can see which sequencing choices will release value quickly and which will sit on a shelf.

The external horizon is also coming into sharper focus. The Department for Business and Trade’s Smart Data 2035 strategy, published in March, positions smart data schemes as the foundation for AI-enabled services across the economy, with property named as a priority sector. We will look in the next edition at what that means for lenders making AI investment decisions today.

Internal AI earns its licence to operate by doing smaller work reliably. The lenders that get the order right build the AI capability they will need for AI-first lending when the ecosystem matures. The ones that skip ahead spend 2027 rebuilding what they deployed in 2026.