Episode 37

Three Actions to Prepare for AI Underwriting

Every loan is a prediction, and I argue that AI is where lending shines because models can use far more data and compute to separate risk better than a score alone. I explain that this lets lenders automate clear approvals while keeping humans for borderline and “not yet” cases. We must build the exam file first with plain-English decline reasons, independent bias testing, vintage tracking, and a written second-look policy. I warn that legal duties and fair lending risk remain with the lender, so vendor diligence and explainability cannot be outsourced. I also advise deciding how we will treat member agents—block them, tolerate them, or build for them—because members and their agents will expect speed. I urge credit unions to start now and to decide where freed lending time will go before piloting any model.

Edward Walters distilled a singular thesis that I both recount and interrogate: every loan is, at its core, a probabilistic forecast. I observe that the canonical underwriting heuristics—credit score, debt-to-income, loan-to-value, and declared income—originated in an era of scarce data and constrained computation, and that the current abundance of information and processing power compels a reconfiguration of risk assessment. I describe how contemporary models can partition risk far more granularly than the blunt instrument of a FICO band, and how Upstart’s reported experience—where a preponderance of personal loans proceed from request to funding without human intervention—illustrates both the operational efficiencies and the strategic dislocations that follow. I assert that the prudent operational posture is to “automate the yes and staff the not yet,” thereby preserving human judgment for borderline files while realizing scale benefits for clear approvals. I articulate pragmatic implementation priorities that transcend vendor claims: construct the exam file before deployment, ensure decline reasons are articulated in plain English, commission independent validation and bias testing, and codify a second-look policy to govern human review. I emphasize vintage tracking of model performance so that institutions can observe how widening the approval envelope impacts delinquencies and losses over time. Finally, I urge leaders to decide where liberated underwriting capacity will be directed prior to pilot launch, because that strategic allocation determines whether automation deepens member service or merely expedites throughput without enhancing portfolio stewardship.

Takeaways:

  • I argue that lending is fundamentally prediction and that we should deploy AI to separate risk better than score alone.
  • I advise we automate clear approvals but assign human staff to review the borderline and not yet cases.
  • I insist we prepare plain English decline reasons, perform independent bias testing, maintain vintage performance tracking over time, and codify a written second-look policy before deployment.
  • I warn that legal duties and fair lending concerns remain; we must own decline notices and conduct thorough vendor diligence.
  • I recommend we decide now whether to block, tolerate, or build for member agents and make systems simple for authorized agents.
Transcript
Speaker A:

Edward Walters of Upstart opened his keynote at America's Credit Union's Lending Conference in Las Vegas with a line he credits to Andre welcome to the slowest day of the rest of your career.

Speaker A:

He's right.

Speaker A:

And his write up why lending is where AI shines is worth your time.

Speaker A:

The core argument Every loan is a prediction.

Speaker A:

And the factors we still lean on Credit score dti, ltv income were picked when data and computing were scarce.

Speaker A:

That scarcity is gone.

Speaker A:

This is banking on disruption reported from the trenches, not the keynote stage.

Speaker A:

I'm Fred Cadena.

Speaker A:

This is a Dispatch.

Speaker A:

Just me one idea.

Speaker A:

About 10 minutes.

Speaker A:

Upstart's numbers are Upstart's numbers and and they come footnoted.

Speaker A:

That's what vendor content does.

Speaker A:

The direction is what matters.

Speaker A:

Models that separate risk far better than a score alone and personal loans that go from rate request to funding with no human involved.

Speaker A:

:

Speaker A:

The keynote opens three questions it doesn't settle.

Speaker A:

Here's where I'd take them.

Speaker A:

Automate the yes staff the not yet no human involvement is a great outcome for the clear approvals.

Speaker A:

It shouldn't mean no humans.

Speaker A:

If a model clears most files on its own, the underwriters and loan officers who used to work them get time back.

Speaker A:

Where that time goes is a strategy choice and it's where a credit union can look.

Speaker A:

Nothing like a fintech put it on the borderline a second look.

Speaker A:

A person reviews files that scored near the line.

Speaker A:

With context a model can't weigh like the story behind a past credit event.

Speaker A:

A better fit product not approved for the consolidation loan.

Speaker A:

Maybe a smaller amount a share secured loan or a credit builder A not yet plan.

Speaker A:

Show the member what would change the answer and invite them back when it does.

Speaker A:

Walters tells the story of a nurse who consolidates four high interest cards and watches her credit improve.

Speaker A:

That's the approved member.

Speaker A:

The member declined today is still a member and likely a future borrower if someone takes the time.

Speaker A:

Machines are good at the yes people are still better at the not yet build the exam file before the pilot credit score.

Speaker A:

Dti, LTV and Income have one big advantage every examiner already understands them.

Speaker A:

A model that splits the 700 to 720 score band into seven risk tiers predicts better and is harder to explain.

Speaker A:

The law still asks you to explain it.

Speaker A:

Where things stand in:

Speaker A:

Buying the model doesn't move that duty.

Speaker A:

The lender owns the notice even when a partner owns the algorithm.

Speaker A:

,:

Speaker A:

That isn't the end of it.

Speaker A:

Disparate impact still applies under the Fair Housing act, and states are active.

Speaker A:

million in:

Speaker A:

A loan book outlives a policy cycle, so test for bias as if the old standards still applied.

Speaker A:

NCUA no AI specific rules.

Speaker A:

The:

Speaker A:

Expect a model that widens the approval box to be examined through that lens.

Speaker A:

NCUA can't examine third party vendors directly, so the vendor diligence is yours to bring.

Speaker A:

Banks SR26 2 replaced SR11 7 in April.

Speaker A:

It's lighter and more tailored and the agencies plan a request for information on AI Specifically what I'd have ready a plain English reason for any decline independent validation and bias testing Someone on staff who can challenge the model Vintage performance as loan season and a written second look policy.

Speaker A:

That last one is the human review above.

Speaker A:

It serves the member and the examiner at the same time.

Speaker A:

The agent on the other side of the table.

Speaker A:

Walter's AI sits on the lender's side of the table.

Speaker A:

The bigger shift may be on the member's side.

Speaker A:

Members will hand more of their financial chores to AI agents Shopping a rate, filling out the application, uploading documents, comparing offers.

Speaker A:

When the applicant is software acting for a member, the questions change.

Speaker A:

Can an agent apply on a member's behalf with consent?

Speaker A:

You can verify.

Speaker A:

Do your rates and terms read cleanly to a machine?

Speaker A:

Does your fraudstack block every agent by reflex?

Speaker A:

Many banks first instinct is to block.

Speaker A:

I'd build for it instead.

Speaker A:

Set the rules for how a member's agent identifies itself and proves consent.

Speaker A:

Then make it easy.

Speaker A:

The institution that's simple for an agent to work with gets the application.

Speaker A:

The one that throws a captcha at it doesn't.

Speaker A:

Walters makes the point that members expect speed.

Speaker A:

An agent won't wait at all.

Speaker A:

Where I'd start.

Speaker A:

Closes by telling leaders to start now.

Speaker A:

I agree.

Speaker A:

Three things for the list 1.

Speaker A:

Automate the yes staff the not yet Decide where freed up lending time goes before the pilot, not after.

Speaker A:

2.

Speaker A:

Build the exam file first plain English decline reasons independent bias testing, vintage tracking a second look policy.

Speaker A:

3.

Speaker A:

Pick your stance on members, agents.

Speaker A:

Block them, tolerate them or build for them.

Speaker A:

Not deciding.

Speaker A:

Banking is a decision.

Speaker A:

The slowest day of the rest of your career is a good day to start.

Speaker A:

This has been banking on disruption.

Speaker A:

Reported from the trenches.

Speaker A:

See you next time.

About the Podcast

Show artwork for Banking on Disruption
Banking on Disruption
A biweekly podcast featuring lively discussions of Financial Services trends, innovation & disruption, and business transformation on the Salesforce platform.

About your host

Profile picture for Fred Cadena

Fred Cadena