Field notes · Consumer credit

The Death of the Loan Officer

Underwriting went from a gut to a model to a data trail. The box keeps moving. Watch who it leaves outside.

Topic · lending & underwriting synthesized from years of notes Read · ~6 min Stance · judge a lender by its exclusions, not its approvals

What I've come to believe

Lending has cycled through three ways of deciding who deserves credit: a person's gut, a rigid model, and now data-rich re-underwriting. Each one fixed the last one's worst bias and quietly introduced its own. The decision method is not the interesting part. Who keeps getting left outside the box is.

The industry tells underwriting as a story of progress: from biased humans, to objective machines, to inclusive data. I read it from the other end. I work the part of the lifecycle where someone who was already approved is now behind, and from there the whole arc looks less like a march toward fairness and more like a box that keeps changing shape. The people who fall outside it change too, but there are always people outside it.

01How it actually works

Credit decisions have passed through three regimes, and it helps to see them as one continuous thing rather than three separate inventions.

First, character. For most of banking history a human in a branch decided, on reputation and the sense of a person across a desk. It was relationship-rich and deeply biased. A loan officer carrying a grudge could let it color a verdict on a stranger who simply reminded them of someone. Then, the centralized model: banks replaced the human with a score to kill exactly that arbitrariness. It worked, and it narrowed the box. Anyone who did not resemble a standard, well-documented profile got a quiet decline. Now, re-expansion: newer lenders use alternative data and open banking to widen the box again, finding creditworthy people the rigid systems had written off.

Each regime did not remove the bias. It relocated it. From who you remind the officer of, to who you resemble in the data, to who leaves a data trail at all.

The same applicants, three eras. The box widens, pinches, and widens again. The bias does not leave, it changes who it catches.

02What holds across the eras

Every fix relocates the bias instead of removing it. Character lending was biased toward who you reminded someone of. Model lending was biased toward who you resembled in the training data. Data lending is biased toward who generates a usable trail in the first place. There is no neutral regime, only a different set of people who get missed.
The width of the box is a business choice, not a law of nature. How wide the credit box opens reflects appetite, funding cost, and economics, not some objective truth about who can repay. When it narrows, that is a decision someone made, and it falls hardest on thin-file and near-prime borrowers who are perfectly capable of paying.
Inclusion and affordability only reconcile at the back end. Widening the box to bring in more people looks like progress at origination. Whether it actually helped them depends entirely on what the lender does on their first bad month. Approve more people with no humane way to handle hardship and you have just widened the funnel of people the system will eventually fail.

03How I read it

My take

This whole arc gets told as a front-door story: who gets approved, and how cleverly. I work the back door, where the swipes have stopped and someone cannot pay. From there, the celebrated "re-expansion" reads differently. Bringing thin-file and near-prime borrowers into the box is genuinely good, but the approval is not the test of an inclusive lender. The first missed payment is.

So my contrarian read is this: the most important underwriting decision is not made at origination at all. It is made afterward, in hardship and re-aging, when a real person hits a real shortfall and the system decides whether to work with them or write them off. That decision is where inclusion is either honored or exposed as a marketing line, and it is the part almost nobody builds for. The fancy data goes into deciding who to let in. The bad month, where it would do the most good, still mostly gets a call center and a script.

The plumbing here is the score. The point is the person on the other side of it, three payments into a year that went wrong.

04Where this is going

Data re-underwriting will keep widening the box, and that is fine. The differentiating skill is quietly moving downstream. The lenders that win the next decade will be the ones that treat the bad month as a product surface to design, not a loss to be minimized, and that bring the same data sophistication to forbearance that the industry spent twenty years pouring into acquisition. The loan officer is dead. The judgment they used to apply, person by person, on a hard month is the thing worth rebuilding.

Threads worth pulling: the "Decoding: Banks" series (11:FS), episodes on lending and the move from character-based to automated underwriting; plus anything on alternative-data credit and affordability rules post-2008. Figures and history here are paraphrased from those notes and are illustrative.