For lenders, the change is not about replacing the underwriter. It is about changing how underwriters spend their time, and which decisions still need their input.
“Automating underwriting” is often perceived as removing the underwriter. The lenders Shape Finance spoke to described something more nuanced. They break a credit decision into its different components: gathering documents, checking data, handling clear-cut cases and making judgement calls. Each part is being automated to a different degree.
The wider market reflects a similar picture to this. In the Bank of England and FCA’s 2024 survey of AI in UK financial services, around 55% of AI use cases involved some form of automated decision-making, but only 2% were actually fully autonomous.
Removing the administrative burden
The clearest gains are in the work that comes before a decision.
At one UK mortgage lender, the platform reads broker documents, pre-populates the application, checks the data and approves documents against set rules. As one of their executives explained, “the actual underwriting that’s left is more about checking what the system’s done and then making a human-based lending decision at the end.”
The time savings here can be substantial. At a US specialist lender we spoke to, the credit team now uses AI built into its everyday software for the first read of a deal, such as asking it to compare two versions of a policy “line by line”. An initial review that used to take about a day now takes around three hours. The decision still sits with a person; what has changed is how quickly they reach it.
One risk and analytics leader at a specialist business lender summed up the approach:
“You can’t use AI to make decisions, but you can use AI to build a process to help make a decision.”
His team uses AI “to aggregate and extract the right data to give to an underwriter.”
Decisions that can already be automated
Automation is not limited to administration. Some credit decisions are already made without an underwriter.
At the same business lender, “about 20, 25% of our deals get auto-declined”, saving around 30 minutes on each, “huge amounts of time,” he said. The firm could also automate approvals: “We’ve got automatic declines and then we can do automatic approvals.”
Nor is this entirely new. An SME lending lead at a UK bank pointed out that “we’ve been making computer-based decisions on credit applications for years”, arguing that “an old-fashioned decision engine is, in my view, AI.”
However, automated decisions do carry trade-offs: colleagues often push back, arguing that some declined applications could have been approved. “People always challenge… there’s some deals we could have approved,” the business lender said, adding “but how do you know that?”
That question helps explain why declines are often automated first. A wrongly declined application rarely shows up in the data: the loan that would have performed well simply never exists. Automated declines therefore carry an invisible cost, while automated approvals carry a visible risk. Lenders need to measure both.
Why judgement remains necessary
The same lenders were clear about where models fall short.
The business lender described an early client making a niche novelty product. On the numbers, sales were “growing and growing and growing.” But it was a novelty product, and the business lasted only a year or two, far shorter than a five- to six-year loan.
The product itself is sometimes key, and sometimes very difficult to model, he said, without “that human in the loop.”
The SME lending lead made a similar point about rare risks: “There will always be room for human underwriting, especially as you get to bigger ticket size lending.” A warning sign that appears in “0.01% of cases”, he explained, will not make it into most models.
Policy can be just as difficult to automate. For the US specialist lender, eligibility rules are “always kind of subjective”, so “at the end of the day you still need the human brain to make sure that the deal works.” The Credit Officer compared the role to a lawyer’s: “We have a book to follow, but we always come up with a story based on the customer.”
A Head of Credit Risk at a UK bank added that portfolio-level decisions on risk appetite are “always very dynamic… this dynamic is very complicated to automate.”
The risk also runs the other way. The CEO of a small specialist business lender is concerned that once underwriters become used to AI, “they will trust it completely and cease to think.”
What these examples share is a difference between looking back and looking forward. Models are good at reflecting what the data already shows. Underwriters add most value when the question is about what happens next, such as: how long a product will sell, what a policy is intended to achieve, or a warning sign too rare to appear in the data.
A different division of labour
Taken together, the evidence points to the underwriter’s role evolving rather than disappearing.
The SME lending lead framed it this way: the goal is less about “what manual processes can we eliminate and more about how can we effectively deploy people within the processes that we have.”
For the business lender, now growing from two underwriters to five or six, the challenge is consistency: “Can I give this deal to everyone and get the same decision from everyone?” His answer is models and guided approaches that set boundaries for underwriters to work within. The question then becomes how wide those boundaries should be. With automatic declines at one end and potential automatic approvals at the other, the underwriter’s skill sits in the middle, “but how narrow do you want to get that middle?”
Make it too narrow, and the role changes in ways underwriters do not welcome. Underwriters he has recruited from heavily automated lenders describe how “the skill of the human underwriter has been lost to a model which says, you just need to check this one thing.”
Regulation points in the same direction. Following the Data (Use and Access) Act 2025, the ICO’s draft guidance on automated decision-making indicates that human involvement must be active rather than just a token gesture, and that reviewers must be able to change the outcome. A human in the loop only counts if they are genuinely exercising judgement.
The question is where judgement adds value
The lenders in this research are not choosing between people and automation. They are deciding where each belongs. That raises practical questions every lender will need to answer when it comes to automation:
- Which tasks are preparation, and which are decisions? Automate the first freely; be deliberate about the second.
- How wide is the middle? Where do automatic declines and approvals stop, and what triggers a human review?
- Can reviewers change the outcome? Do they have the time, information and authority to do so?
- How will underwriting skills be maintained? If underwriters only see flagged cases, how do they build the judgement those cases require?
Automation can remove much of the preparation, handle straightforward cases and give underwriters better information. Deciding where human judgement still adds value is now a design choice, not an afterthought.
Continue the conversation
Where should human judgement sit in automated credit decisions?
It’s one of the questions we’ll be exploring in the panel “Redefining credit decisioning in the age of AI” at LendTech Summit London on 17 November, Shape Finance’s invitation-only forum for senior lending, credit and technology leaders.
