A General AI Reads “1919” as a Year. Your Credit Analyst Reads It as the Borrower Information Form

Three and a half hours. That is what a decent credit analyst spends on one 7(a) change-of-ownership file before she writes a word of the memo: three years of the seller's 1120-S with the K-1s, the Form 8825 behind the real estate, the depreciation on Form 4562, the buyer's personal 1040s with Schedule E page two, an interim balance sheet with an accounts receivable aging, and then the global cash flow that has to hold at 1.15 times or better with the new debt service in it.

Every lender who tried a general-purpose assistant on that job in 2024 got the same result: something that looked like a credit memo, read beautifully, and was wrong in ways only an SBA person would catch. It called the 10% equity injection a down payment. It treated the seller note as a standby letter of credit. It cited an SOP paragraph that does not exist. Your chief credit officer read four sentences and told you to stop.

Why the general model kept getting SBA files wrong

Not because it is stupid. Because this trade writes in a private language and never explains it.

Say "1919" to anyone in your shop and they hear the Borrower Information Form. Say "1920" and they hear the lender's application. A general assistant hears years. Say "504" and your team hears a CDC debenture structure with a bank first at 50%, a CDC second at 40% and 10% from the borrower; a general assistant hears a subsection number or an error code. "Full standby" means no payments of principal or interest for the life of the loan, evidenced on a standby agreement — not a letter of credit, not a deposit. "EPC/OC" is an eligible passive company holding the real estate and an operating company running the business, and every rent, lease and guarantee condition follows from it. "CAPLines" is a real program, not a typo. "1502" is the monthly report your servicing specialist files, not a form your borrower signs.

A vertical model is one that has been trained on a single trade's own documents — its forms, its shorthand, its rules and its finished work — so that it reads "full standby seller note on an EPC/OC change of ownership" the way your closer reads it, and tells you which page of which document it got each answer from. In 2026 that stopped being a science project. Tuned, trade-specific models became a distinct category precisely because general ones miss the vocabulary, the units and the edge cases a trade takes for granted.

The five mistakes that used to make it useless

Worth naming them, because they are the exact things you should test any tool against before it touches a live file:

• Cash flow spreading. A general model adds back depreciation and amortization and stops. It misses officer compensation adjustments tied to the buyer's actual post-closing salary, rent normalization when the operating company will now pay rent to the buyer's own holding entity, and one-time add-backs that need a seller's explanation letter in the file.
• Global versus business coverage. Business-only debt service coverage of 1.4 times looks fine until you fold in the guarantor's mortgage, the spouse's student loans and a rental property running at a loss on Schedule E.
• Program caps. The 7(a) maximum is $5 million. A 504 project uses a debenture capped at $5 million, or $5.5 million for qualifying manufacturers. A general model blends these into one number and your memo goes to committee wrong.
• Eligibility and affiliation. Ownership percentages, control, franchise agreements and prior government debt drive whether a deal exists at all. This is where a confident wrong answer costs six weeks.
• Dictated notes. Your BDO leaves a two-minute voice note after a client visit. General speech recognition writes "seven A" and "cap lines" and "eat ran." A tuned one writes 7(a), CAPLines and E-Tran. Speech recognition itself got roughly ten times faster in 2026 — the fix was never speed, it was vocabulary.

flowchart TD
  A["Tax returns, interim financials, 1919 packet"] --> B["General assistant"]
  A --> C["Model tuned on your credit files"]
  B --> D["Add-backs missed, standby misread"]
  C --> E["Spread matched to your template, sources cited"]
  D --> F["Analyst re-spreads from scratch"]
  E --> G["Analyst checks flagged items only"]
  F --> H["Memo to loan committee"]
  G --> H

What tuning actually fixes, in the analyst's chair

The tuning material is not exotic. It is your last two hundred approved credit memos, your spreading template with your own add-back conventions, your loan policy, your eligibility checklist, the SBA forms your closers use, and the current SOP. Trained on that, the assistant produces a first-draft spread in your format — not a generic one — with each figure traceable to the page it came from: "officer compensation, 2024, $214,000, page 3 of the 1120-S, line 7."

The analyst's job changes from typing to checking. She opens the draft with the returns beside it, confirms the eight or nine figures that matter, resolves the flags — a $61,000 gain on sale of equipment that needs to come out, a related-party rent that has to be normalized to the new lease — and writes the credit narrative, which is the part that was always worth her salary. Three and a half hours becomes about seventy minutes, and the memos start looking alike, which is what your loan review consultant has been asking for since 2023.

A number your credit committee will care about: the repair exposure

The savings here are real but small. The exposure avoided is the reason to do it. When a 7(a) loan defaults and you send in a guaranty purchase package, the SBA reviews whether you followed the program's requirements. Missing verification of financial information, a defective standby agreement, an unsupported use of proceeds — these produce a repair or a denial, and the loss lands on you.

Illustrative arithmetic, using your own portfolio shape:

Cut the documentation-driven half of that repair exposure and you have found roughly $36,000 a year, against a tool cost that is now a rounding error — running frontier models costs about a tenth of what it did in 2025. But do not sell it to your board as savings. Sell it as file quality: every eligibility position stated, every figure sourced, every required form present at the point the loan number is assigned in E-Tran, not reconstructed from memory three years later by a special assets officer who was not there.

Where you keep a human, permanently

Eligibility calls stay with your SBA manager. Affiliation, control, prior loss to the government, citizenship and residency status of guarantors, franchise agreements — these turn on facts a document does not contain and a borrower may not volunteer. A tuned model is excellent at telling you which question to ask; it is not the one who signs the eligibility position.

Character stays human. The decline that saves you money is usually not a math decline. It is the buyer who cannot answer how he will replace the seller in the accounts, or the seller who is suddenly vague about why revenue jumped 22% in the twelve months before the sale. No amount of tuning reads a room.

And the SOP moves. Rules and procedures change, sometimes with effective dates that split your files in half. Assign one person — usually the SBA operations manager — to confirm what the assistant was trained on and when. A tuned model that is two revisions behind is more dangerous than no model at all, because it sounds exactly as confident as it did when it was right.

Start with the spread, not the memo

On Monday, take five closed files you already know cold. Have the tool spread them and compare, line by line, against what your analyst produced. Count the differences and sort them into three buckets: it was right, it was wrong, and it was a judgment call your shop makes differently. That third bucket is your tuning list. Do not put it near a live file until the first two buckets are boring.

Frequently asked questions

Will this send my borrowers' tax returns somewhere I cannot control?

That is a vendor and contract question, and it is the right one to ask first. Get in writing where the files sit, whether anything is retained for training, and how the arrangement fits your existing customer information program and vendor management review. Several banks run this work on their own hardware for exactly this reason, and the cost of doing so fell sharply in 2026.

Do I have to hand it my whole loan policy?

You get the benefit from the documents that encode judgment: approved memos, the spreading template, the eligibility checklist, and your standard conditions. Pricing grids and risk rating models can stay out. Start with two hundred memos; that is enough for it to sound like your bank instead of a generic bank.

Does this replace my analyst?

It replaces the typing, not the analyst. Shops that got this right kept the same team and pushed more files through, which matters most in the run-up to the SBA fiscal year end on 30 September when everything lands at once. If you cut the seat, you also cut the person who catches the wrong answer.

A note on where the calls fit

Faster memos only matter if the deal reached you. Most SBA inquiries arrive by phone, from brokers and buyers who call between other meetings and do not leave voicemail. CallSphere builds AI voice and chat agents that answer the line, ask the handful of questions that establish whether the deal fits your box, and book time with the right BDO. It does not spread a tax return or take an eligibility position — it makes sure the file gets to the analyst who can.