Lending has an unusual cost structure: the lines that matter most move against each other. Cut underwriting effort and credit losses rise. Tighten the credit box and volume falls. Funding cost follows the market, not management. So most efficiency proposals a lender receives are really proposals to move cost from one line to another, and experienced finance teams read them that way.
That is the test I’d apply to any vendor’s business case, ours included. Routine underwriting work is one of the few places where I think the case for automation passes it.
The cost that has no budget line
Credit losses and funding are usually the largest items on a lender’s P&L, and they get the attention. Operating cost comes next, and staff is normally the largest part of it. That is where routine underwriting sits: experienced people reading bank statements on applications whose outcome was never in doubt.
It has no budget line of its own. It sits inside salaries, overtime and, above all, the hires approved each year because volume went up. Those requests are usually well argued: applications up 15%, the team stretched, service levels slipping. They get approved because the alternative is a backlog. Nobody in that discussion prices the share of the new hire’s time that will go on files a model could have cleared.
What the research shows
A randomised experiment in auto lending, published in Management Science (Jansen, Nguyen and Shams, 2025), randomly assigned around two million applications to either human underwriters or an algorithm. The algorithm came out ahead: 10.2% higher loan profits and 6.8% lower default rates.
The breakdown matters more than the headline. On low-risk, straightforward loans, people and the model performed about the same. The algorithm’s advantage came on riskier, more complex cases, where human performance fell away.
From a finance seat, that says something specific. On the simple part of a book, automating review takes cost out without moving the loss line. No credit risk is quietly absorbed to fund the saving. A lender should still want that proven on its own portfolio before building a case on it, and we expect to be asked. But it is a better starting point than most efficiency proposals get.
What 51.7% actually means
One of our clients automated 51.7% of its applications with no change in decision quality. The instinct on a finance desk is to turn that into a salary saving. That undersells it.
If half of incoming volume no longer touches a person, the link between applications and headcount is broken for that half. The team can stay the same size while originations grow. That is a change in how operating cost scales, and it compounds in a way a one-off saving does not. It is also harder to show a board, because much of the return is hires that never happened. Someone has to build that counterfactual explicitly, or it disappears in the next budget round.
On metrics: cost per decision is the number everyone leads with, vendors included. Treat it with care. It falls whenever throughput rises, whatever the decisions were worth. Track loss rate, approval rate and applications per underwriter together. Any one of them on its own can be made to look good.
Two costs to include from day one
Model monitoring. A decisioning model reflects the conditions it was built in, so recalibration is a running cost with a named owner at the lender. Auditors will ask about it. It is cheaper to have the answer in the business case than to add it later.
Data. The US open banking rule is being rewritten, and the revised proposal reportedly lets banks charge for data access, which the original rule banned. If the model depends on bank data, the case should show the margin impact if that data starts carrying a price.
The return depends on the lender, not the software
Automation removes the work. What happens to the freed capacity is a decision inside the lender, and that decision determines whether the business case was right.
Holding headcount flat while volume grows is the simple version, and the one most finance teams will favour. Moving experienced underwriters onto exceptions, model oversight and early shifts in borrower behaviour is harder. But it puts the best judgement on exactly the cases where the research says it matters.
Either works. Leaving the freed time where it is does not, and that is what happens by default when nobody owns the decision.




