By Sagar Shankaran, Founder of CallSphere
Routine loan files go to a cheap AI model, hand-calculated income goes to the strong one. How a mortgage shop writes that routing rule, priced out per month.
Key takeaways
She gets in before the loan officers do, because that is the only hour of the day nobody calls her. Encompass is open on one screen, the condition report on the other: forty-one outstanding conditions spread across twenty-three files, and eleven of those files have purchase contracts with closing dates inside the next three weeks. Spring buying season does this every year. The contracts pile up in April and May, and the conditions pile up right behind them.
Most of that queue is not hard. A paystub to match against a written verification of employment, a bank statement with a $4,200 deposit that needs a letter of explanation, an insurance binder with the wrong mortgagee clause, a gift letter missing the donor's account statement. Any competent processor clears those in minutes. But six of the forty-one are genuinely hard: two self-employed borrowers with K-1s from two entities, one with rental income on Schedule E that includes a property bought mid-year, one with restricted stock that vests unevenly, and two where the Desktop Underwriter findings came back Approve/Ineligible and somebody has to work out which overlay is biting.
The hard six are what actually blows the closing date. And they are sitting in the same queue as the easy thirty-five, in whatever order they arrived. Suppose your lock extension costs about two basis points a day. On a $400,000 loan that is $80 a day, every day, and on a busy month across a shop doing 180 units, extensions are a line item somebody in accounting complains about every quarter.
Owners get pitched AI as if the only question is whether the machine is smart enough. Wrong question. A $1.2 million jumbo to a salaried physician with one W-2 employer, two years on the job and a clean Approve/Eligible is an easy file. A $210,000 FHA purchase where the borrower runs a landscaping business, files a Schedule C, took a big Section 179 deduction, and has a co-borrower drawing partial disability is a hard file. The loan amount tells you nothing.
Here is the sentence to keep: model routing means a fast, cheap AI reads every file first, and only the files that trip a rule your underwriting manager wrote get handed to the expensive model that thinks harder and costs more. The rule is the product. The models are commodities.
What trips the rule in this business is not mysterious. It is income that has to be calculated rather than read: anything that lands on Fannie Mae Form 1084 or Freddie Mac Form 91. It is any file where the automated findings are not a clean Approve/Eligible. It is non-QM, bank statement loans, asset depletion, anything with a foreign national, anything where the appraisal came in short and somebody is contemplating a reconsideration of value. Everything else — the W-2 borrower, one employer, paystubs and a Work Number verification that agree with each other — is routine, and routine is the majority of your volume.
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 mistake shops make is letting the vendor decide the line, or worse, letting it be decided by whoever configured the software last. The person who should own that page is the same person who owns your credit policy: the underwriting manager. They already know which file types produce your repurchase requests and your early payment defaults. That knowledge is the routing rule.
flowchart TD
A["1003 submitted through the borrower portal"] --> B{"Clean Approve/Eligible and one W-2 employer?"}
B -->|Yes| C["Fast model reads paystubs, W-2s and the VOE"]
B -->|No| D["Strong model reads K-1s, 1120S and Schedule E"]
C --> E["Processor spot-checks and clears conditions"]
D --> F["Underwriter recalculates income on Form 1084"]
E --> G["Conditions posted to the file the same day"]
F --> G
Notice what the chart does not say. It does not say the fast model approves anything. It reads documents, compares them to each other, and writes what it found into the conditions log. A human still clears the condition. The routing decides who does the reading and how much thinking the reading is worth — not who makes the credit decision.
Two years ago, splitting work between a cheap model and an expensive one was something a company with its own engineering team did. In 2026 it became ordinary practice. Cisco built model routing directly into the personal AI agent it is rolling out to roughly 90,000 employees, with an on-premises emphasis, precisely to keep cost in line with capability. When a company that size treats routing as table stakes, the tooling underneath it stops being exotic and starts showing up in the products smaller shops buy.
The second thing that changed is price. Frontier AI costs about a tenth of what it did in 2025, so reading a forty-page income packet is now cheap enough to run on every application instead of only the ones somebody flags. The third is spending control: Anthropic's enterprise update on 2 July 2026 added spend limits per team and per person, alerts at 75% and 90% of budget, and the ability to set which model is the default for which group. Your branch manager can set the default to the cheap model and require a reason to use the expensive one — the same way you already require a second signature to waive a condition.
Illustration, not a promise. Assume a shop that works 260 applications a month, of which about 62% are salaried borrowers with a single employer and clean automated findings.
| Line | Assumption | Result |
|---|---|---|
| Applications worked per month | Stated | 260 |
| Routine (one W-2 employer, clean findings) | 62% | 161 files |
| Hand-calculated income files | 38% | 99 files |
| Processor minutes per routine income review today | 35 min | — |
| Minutes after the fast model does the first pass | 9 min to verify | 26 min saved |
| Monthly hours returned | 161 × 26 min | 69.8 hours |
| Fully loaded processor cost | $38/hour | $2,652 |
| Fast model, routine files | $0.30/file | $48 |
| Strong model, hard files | $1.90/file | $188 |
| Net monthly difference | — | about $2,400 |
Now the honest part. Running all 260 through the expensive model would cost $494 instead of $236, and that $258 is not why you route. You route because the two-entity K-1 file stops sitting behind thirty W-2 files, and the hard six reach your underwriter Tuesday morning instead of Thursday afternoon. Measure the win in lock extensions avoided and closings that held their date.
Keep it away from anything that produces a denial. Adverse action under Regulation B is not a documentation task, it is a legal one, and the reason stated has to be the actual reason. Keep it away from your HMDA data. A tidy-looking guess on an application register field is worse than a blank you go back and fix, and your compliance officer is the one who signs that submission.
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.
Keep it away from the reconsideration of value file, from any borrower who has disclosed a hardship, and from the judgment call on whether a twelve-month employment gap is explainable. And no model gets the final word on self-employed income. The strong one can lay the K-1s, the 1120S, the 1065 and the Schedule E side by side and show its arithmetic — real time saved. Your underwriter still signs the 1084, because if the quality control sample comes back with an income variance, the person who answers for it has a name.
One limit shops discover the hard way: a fast model reading a scanned paystub from a small employer's in-house payroll will sometimes take year-to-date gross where you wanted the pay period figure. That is a document-quality problem, not a smart-versus-dumb problem, and it is exactly why the processor still spends nine minutes on the file instead of zero.
You can, and you will regret it. The vendor's default was written for an average shop and your credit box is not average. Have your underwriting manager write the rule from your own last twelve months: which file types produced missed conditions, repurchase demands, early payment defaults. One page, reviewed the way you review overlays.
Both exist, and the separate window is the one that fails. If a processor has to copy documents out of the loan file into another tool, she will stop during the last week of the month, which is exactly when you needed it. Make any vendor show you the finding written back into the conditions log on the file itself, linked to the page it came from.
Build the rule so it fails toward the expensive lane. Anything ambiguous goes to the strong model — the cost of over-routing is a couple of dollars, and the cost of under-routing is a missed closing date. Then sample: pull ten routed-easy files a week and have the underwriting manager check whether they belonged there. Adjust the rule monthly, not daily.
The dollars scale down, the calendar pressure does not. A 40-unit shop usually runs one processor and one underwriter, so a single hard file stalls three easy ones. Triage order is worth more to you than to the 180-unit shop, not less.
Do not start with income. Start with homeowner's insurance binders. Every file has one, the failure is boring and visible — wrong mortgagee clause, wrong effective date, coverage below the loan amount — and no credit decision rides on it. Run the fast model on every binder from last month and compare its findings to what your closer actually caught. If it holds up, move to paystub-versus-VOE matching. Then, and only then, talk about income calculation.
One last thing, because it is the part owners forget. Every condition you clear generates a phone call — the borrower asking whether you got the document, the realtor asking whether the closing date still holds, the listing agent calling the branch line at 6pm. CallSphere builds AI voice and chat agents that answer those calls around the clock, book the follow-up with the loan officer, and capture the lead when it is a new borrower calling about a rate. It does not underwrite your file. It keeps the phone from eating the hour your processor just got back.
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.
Sixty-one pages of addendum land three days before a DOT letting. How model routing gets an estimator the quantity changes that actually move the bid.
Model routing sends routine product requests to a cheap model and hard ones to a strong one. Here is who draws the line in a B2B software company, and how.
Cheap model for potholes, strong model for water quality, a pager for sewage in a basement. How a public works superintendent writes the triage table.
How water systems use a cheap-model first pass and a strong-model escalation so nobody drives out at 2 a.m. for a SCADA alarm that already cleared itself.
Sort courier calls before answering: cheap model for status and reschedules, strong model for damage claims and missed medical runs. Costs, rules and limits.
An AI agent in a lending shop should read widely, write to the conditions log, and send nothing with a routing number. The permissions to remove this Monday.
© 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