By Sagar Shankaran, Founder of CallSphere
Sampling 25 of 400 loan files was a payroll decision. With 2026 AI prices a credit union can scrub every file for HMDA, Reg B and rate errors before funding.
Key takeaways
Ask your loan quality control person how many of last month's files she actually opened. At most credit unions the answer is somewhere between fifteen and thirty, out of three or four hundred decisions. Then ask her how the sample size was chosen. She will tell you the truth: it is how many files one person can review in the days she has left after everything else.
The sample size was never a risk decision. It was a payroll decision, and everyone in the building knows it. As of 2026 the payroll constraint is gone, and that changes what your loan operations supervisor should be doing on the first Monday of the month.
Three things go wrong in consumer loan files at a credit union, and none of them are exotic.
The first is data quality on your HMDA reportable loans. Fields get keyed wrong — the wrong action taken code, a missing rate spread, a census tract that does not match the property address, an ethnicity and race field recorded as "not provided" when the application clearly shows it collected. Those errors sit quietly on the loan application register all year and surface as a resubmission demand, which means your mortgage staff spends February re-scrubbing a full year of records instead of taking applications.
The second is adverse action. Regulation B gives you thirty days to notify the applicant, and the notice has to state the actual principal reason for the denial. The failure mode is rarely a missed deadline. It is a notice that says "insufficient income" on a file that was denied for derogatory credit, because the loan officer picked the first plausible reason from the drop-down in the origination system. That is the exact pattern a fair lending review looks for, and a sample of twenty-five will not find it reliably.
The third is the funding package on indirect auto — your largest channel by volume and your thinnest by margin. A missing GAP disclosure, a mismatch between the contract rate and the approved rate sheet, a stipulation marked satisfied with nothing behind it, a title application that never went to the state. Every one of those is caught eventually. "Eventually" means after funding, when your only remedy is a phone call to a dealer who has already moved on.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for financial services in your browser — 60 seconds, no signup.
The cost of running capable AI fell roughly tenfold from 2025 into 2026. Strong models now cost on the order of $2 to read a million words of text, and running the work on your own equipment is roughly 90 percent cheaper again for high-volume jobs. The practical meaning for a lender is simple: a check that costs a fraction of a cent per document can be run on every file, so sampling becomes a choice rather than a budget.
In 2025 the same idea existed and the math did not work — reading four hundred full loan files every month cost real money and someone had to justify it against a fraud loss number nobody could forecast. Now it is a rounding error next to the shredding contract.
flowchart TD
A["Loan decision closes in the origination system"] --> B["Overnight job pulls file, notes and disclosures"]
B --> C["Check 1: HMDA fields against the application and credit file"]
B --> D["Check 2: Denial reason matches the actual decision"]
B --> E["Check 3: Rate, term and fees match the approved sheet"]
C --> F{"Anything mismatched?"}
D --> F
E --> F
F -->|No| G["File logged clean, no human time spent"]
F -->|Yes| H["Exception queue for the loan ops supervisor by 8am"]
Your loan operations supervisor arrives to an exception list rather than a cart of folders. Nine items out of last night's 400. Two are HMDA field mismatches on purchase-money mortgages — one census tract, one action taken code recorded as withdrawn where the file shows a denial. Four are indirect auto packages missing proof of insurance past the stipulation window. Two are adverse action notices where the stated reason does not line up with the decision notes. One is a rate on a signed retail installment contract that is 40 basis points above the approved tier.
That last one is the whole month's return in a single line. It is a call to the dealer's finance manager before the contract is purchased, not a chargeback conversation in September. And the other 391 files got read — genuinely read, every page — by something that costs less than the break-room coffee.
Two changes that make this work in practice. First, the exception has to arrive with the evidence attached — the page, the field, the sentence in the notes — or your supervisor is doing the review twice. Second, every exception needs a disposition code, because in six months your compliance officer will be asked how many exceptions were raised, how many were real, and what you did about the pattern.
Assumptions, illustrative: 400 consumer loan decisions a month (240 indirect auto, 110 direct consumer, 50 credit card), 46 denials, and an average file of about twelve pages or 6,000 words of text once you include the application, credit report summary, notes and disclosures. Reading cost of roughly $2 per million words handled, with a modest allowance for re-runs.
| Item | Today: 25-file sample | All 400 files |
|---|---|---|
| Words read per month | 150,000 | 2,400,000 |
| Direct cost to read | Staff time only | About $5 — call it $9 with re-runs |
| QC analyst hours | 18 hrs | 6 hrs on 9 exceptions |
| Coverage of denials | About 3 of 46 | 46 of 46 |
| Rate or fee errors caught pre-funding | Roughly 1 in 16 | Effectively all of them |
Take the single most defensible saving: a HMDA resubmission. If a full-year re-scrub costs you 120 hours of mortgage staff time at a loaded $54 an hour, that is $6,480 in labor, ignoring the examination write-up entirely. Catching field errors monthly at $9 a month — $108 a year — does not require a sophisticated return-on-investment argument. It requires someone to set it up once.
It is reliable on comparisons: this field against that document, this rate against that sheet, this reason against those notes. Those are mechanical, and mechanical is where these tools are strongest.
Still reading? Stop comparing — try CallSphere live.
See the financial services AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
It is unreliable on judgment. It cannot tell you whether the denial itself was correct — whether that debt-to-income ratio should have been an exception given eleven years of membership and a spotless share draft history. It cannot read a handwritten dealer notation reliably every time, and indirect packages are full of handwriting. And it will not catch a document that is missing entirely unless you have told it exactly which documents that loan type requires, which means somebody on your side has to write that checklist down — a good exercise that most credit unions have never done for anything but mortgages.
Every exception is a flag, not a finding. A human decides. Keep it that way in writing, because the moment a machine-generated exception becomes an automatic action on a member's application, you have created a decisioning system that Regulation B, your fair lending policy, and — if you have any operations touching the EU — the transparency obligations that took effect on 2 August 2026 all have opinions about.
Keep humans firmly in charge of member-facing consequences: any denial, any repossession decision, any hardship or skip-a-pay refusal. Use the tool to make sure the paperwork behind the human's decision is right. That is a narrow claim, it is defensible in an examination, and it is where the money actually is.
Do not attempt all 400 in month one. Take last month's 46 denials, and check one thing: does the principal reason on the adverse action notice match the reason in the decision notes? It is a single comparison, the documents are already in your origination system, and you will know within an afternoon whether your process has a problem. In most credit unions that first run finds between three and eight mismatches, and that result is what gets the rest of the program funded.
It validates format — that a field is filled in and legal. It cannot tell you the census tract is wrong for the address on page four, or that the action taken code contradicts the notes. Format validation and truth are different problems, and the resubmission demands come from the second one.
That you expanded quality control coverage from a sample to the full population using an automated document review, that every exception is dispositioned by a named employee, and that no member-facing decision is made by the tool. Bring the exception log. Examiners respond well to more coverage, provided the human accountability is unambiguous.
The cost side is trivially yes — it would be about $1.50 a month. The question is whether the setup effort is worth it at that volume. If you already have someone comfortable with your origination system's reporting, yes. If not, start with denials only and stop there; even at 60 loans a month, adverse action reason mismatches are the finding most likely to hurt you.
They push back on speed, not on accuracy. If your exception arrives before funding and includes the specific page and the specific number, most finance managers fix it the same day, because a returned contract costs them more than it costs you. The pushback comes if your exceptions are vague or arrive a week late.
A closing note on the phones this creates. Full-file review means more stipulation calls and more clarification calls, and those land on the same member service line that is already busiest between 11:30 and 1:30. CallSphere builds AI voice and chat agents that answer the member line and web chat around the clock, handle the routine questions, capture what the member said, and book a callback with the loan officer who actually owns the file. It does not review loan documents; it keeps the follow-up from falling on the floor.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Open models closed the gap in 2026. A worked buy-versus-rent comparison for a 58-employee credit union, including the staffing cost nobody budgets for.
Vendor acknowledgments quietly restate price, date, quantity and freight terms. Why checking all 1,200 a month became economic in 2026, with worked arithmetic.
A credit union's core agreement plus nine amendments now fits in one question. Find the notice window, the escalators and the exit fees before renewal.
SAR confidentiality rules out cloud chatbots. Here is what an on-site model does for a credit union BSA officer, the hours math, and the examiner questions.
Copied DARs, clustered tour hits and injury language sit unread in 8,400 reports a month. Here is the 2026 arithmetic that makes reading all of them economic.
Held-away 401(k)s, annuities and non-traded alts have no feed into Orion. How browser-driving AI agents cut five days out of the quarter-end reporting run.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI