Ayesha's Saturday: What AI in Indirect Lending Actually Looks Like

Ayesha McFarlin has been processing indirect auto loans at America's First Federal Credit Union since 2011. That's fourteen years of looking at sales contracts, title applications, back-end product agreements, membership applications, and the eighteen or so other documents that show up in a single indirect package. She knows what a clean file looks like at first glance. She also knows every kind of dealer mistake before the file even makes it to her desk.

A few weeks ago, on a Saturday at the end of the month, she did something that would have been mathematically impossible six months earlier.

The team was close to its monthly goal and needed to push through a stack of contracts before close of business. AmFirst's indirect funding team is small. Just two full-time funders cover a volume that ranges from 1,000 to 1,500 loans per month. They don't usually work Saturdays, but Ayesha came in with one other processor, and the two of them funded 60 deals worth $2.2 million in roughly six hours.

For context: two weeks earlier, on a Monday, the team had funded 58 deals worth $2.2 million in a standard eight-hour day with three people. Roughly the same volume, but with one fewer person and two fewer hours. Per funder-hour, throughput roughly doubled.

What changed between those two days was that one of the two processors on Saturday had Saris running in her workflow, and the other did not.

What is actually happening inside a single loan file

To understand how the throughput doubled, it helps to look at what AmFirst's team actually does for every loan that comes through indirect.

An e-contract lands in MeridianLink from one of the roughly 150 dealers AmFirst works with. The package is sitting in the ready-for-fund queue. A funder pulls it up and starts reviewing. There are typically 18 to 20 documents in a package: the sales contract, the title application, back-end product agreements like maintenance or warranty contracts, insurance information, a membership application, a credit application, dealer disclosures, and assorted state-specific forms.

The funder's job is to verify that every piece of that package is correct, complete, and consistent with both the sales contract and the credit union's internal policies. That means a lot of stare-and-compare. The VIN on the title application has to match the VIN on the sales contract, which has to match the VIN on the insurance binder, which has to match the VIN on every other document where it appears. The borrower's date of birth has to match across every document that asks for it. The financed amount has to reconcile across the contract and the LOS. The rate has to match what was approved. The back-end products have to add up to under 20% of the total amount financed per AmFirst's policy. The LTV has to fall within the approver's authority limit.

A trained funder doing this manually moves through a package in 15 to 25 minutes when nothing is wrong, and longer when something is. Across 50 to 75 packages a day, the time adds up. Across a team of three funders, the time adds up faster.

The validations Saris is running underneath

When Saris is in the workflow, the review structure stays the same. The credit union's policy doesn't change. The systems don't change, the team doesn't change. What does change is who does the mechanical comparison work, and where the human's attention is spent.

Saris's agents run two categories of validation on every file that lands in the ready-for-fund queue.

The first is document-to-system validation. The agent reads the sales contract and compares the loan terms against what was entered in the LOS. If the term is 72 months in the LOS but 75 months on the signed contract, the agent flags the mismatch and surfaces both sources side by side for the funder to review.

The second is completeness and policy validation, and it covers a few things at once. The agent confirms every document the policy requires is actually in the packet. Where documents should agree, it checks that they do. If the sales contract shows a back-end product was elected, it verifies the corresponding agreement is present and the dollar figures reconcile. If a stipulation requires proof of income, it finds the pay stub or VOI, runs the income calculation, and compares it to the LOS. It checks the contracted rate against AmFirst's rate sheet, given the borrower's credit tier, term, and vehicle year, and surfaces any rate outside policy.

That same completeness and policy pass also runs the back-end product math: it sums every back-end item on the sales contract and divides by the total amount financed. If that figure exceeds AmFirst's 20% threshold. If a dealer has padded the back end above policy, the agent surfaces the calculation, not just the conclusion, so the funder can see exactly how it got there.

Across field-level checks, Saris agents are operating at 99.8% accuracy. At the workflow level, the stricter measurement that requires the full file to clear without intervention, the agents run at 94.2%. The remaining files surface as exceptions for human review, with the relevant context attached. This is the designed behavior of the system, not a residual failure to engineer toward zero.

From processing queue to exception queue

The reason Ayesha's Saturday looked different is that her workflow is no longer about reviewing every field in every document. It is about reviewing the items Saris has flagged as worth her attention. The agent already validated the VIN across the documents that contain it. It already ran the LTV against the approver's authority. It already confirmed the gap addendum is present and the back-end math is in policy.

What lands on Ayesha's screen is the file with the term mismatch, the rate that came in below tier policy, or the dealer that consistently sends the title application with a missing signature date. Her attention goes to the work that requires her judgment, not her keystrokes.

That shift pays off again further down the line. Post-funding audit and QC used to catch a steady trickle of errors that slipped past the funder's manual review and surfaced days later, bouncing back to whoever funded the deal. Now the same validations that clear a file in the funding queue catch those errors up front, before they ever reach QC. Ayesha even runs her teammates' funded files through Saris as a second check. The processors who used to dread an exception report from compliance are funding with a lot less anxiety.

That part of the AmFirst engagement is the part I keep coming back to. The throughput numbers are real and worth talking about. But the change in how a long-tenured lending team feels about its own work is what makes me think the rest of this category is going to look very different a year from now than it does today.

‍

‍

‍