By Sagar Shankaran, Founder of CallSphere
Four hand entries of the same forty fields per loan. How AI working inside nCino, E-Tran and the core cuts closing hours and the monthly 1502 rebuild.
Key takeaways
You tried this in 2024. Somebody sold your bank an assistant, it lived in its own browser tab, and using it meant copying loan data out of your origination system and pasting the answer back in. Your loan closer, who already keys the same borrower facts into four places, correctly decided it was a fifth. It got quietly abandoned around the third quarter and nobody has mentioned it since.
Here is what is different in 2026: the assistant no longer lives in a tab. The Model Context Protocol — MCP, a common plug between AI assistants and business systems — became the standard way software vendors let an assistant read and write inside the system you already pay for. Oracle shipped AI Agent Studio for its Fusion applications. Microsoft shipped Sales Agent and Service Agent. Salesforce wired agents into Slack, which matters directly if your loan origination system runs on the Salesforce platform, as nCino does. The practical meaning for a lender: the AI works inside the loan file of record instead of asking your closer to be the connection between two screens.
Follow a single 7(a) from committee approval to funded, in a bank running a normal setup — an origination system such as nCino, Baker Hill NextGen or Abrigo, document preparation in LaserPro, the SBA's own E-Tran through CAFS, and the core at Jack Henry or Fiserv.
That is four to five separate entries of substantially the same forty fields, by hand, on every loan. It is where transposition errors are born, it is why the collateral description on the UCC-1 sometimes does not match the security agreement, and it is why your closer works late in the last two weeks of September when everybody wants a loan number before the SBA fiscal year ends on the 30th.
flowchart TD
A["Loan committee approves"] --> B["Agent reads approved terms in the origination system"]
B --> C["Fills the 1919 and 1920 data into E-Tran"]
C --> D["Builds the LaserPro document set"]
D --> E{"Closer reviews the differences report"}
E -->|Corrections needed| B
E -->|Clean| F["Loan boards to core, ticklers created"]
F --> G["Monthly 1502 file reconciled and filed"]
Committee approves Monday at 4. Tuesday morning the closer opens the loan in the origination system and asks for a closing package. The assistant reads the approved terms from the record itself, drafts the E-Tran entry, assembles the document request, and produces one screen she actually wants: a differences report. Ownership on the 1919 says 51/49; the operating agreement in the file says 60/40. The use of proceeds in the memo includes $95,000 of working capital that the settlement sheet does not show. The flood determination came back in a special flood hazard area and no flood insurance requirement was added to the conditions.
She resolves four items in twenty minutes, approves, and the entries are written back into the systems — not pasted by her. The tickler records are created from the actual conditions in the approval rather than a template someone copied in 2019. Closing moves from an all-day job to a morning, and the errors that used to surface at funding surface before documents go out.
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 same shape works on the servicing side, which is where most banks are quietly bleeding. Your servicing specialist spends the first two working days of every month assembling the 1502 report for the SBA's fiscal transfer agent: loan status, balances, interest paid, payments received, reconciled against the core. When the assistant can read the core and write the file, that becomes a review of exceptions — the six loans whose status changed, the two that paid off, the one where the payment posted to the wrong loan — instead of a full manual rebuild.
Insurance certificates expire. Borrowers do not send financial statements 120 days after year end. UCC filings lapse at five years, and a lapsed UCC on a $1.2 million equipment loan is the kind of finding that turns into a repair when the guaranty purchase package goes in.
Today this is a report your portfolio administrator runs, a spreadsheet, and a series of emails that get sent when there is time. With the assistant inside the servicing record, the chase becomes routine: it reads the tickler, drafts the borrower request naming the exact document, logs the outreach against the loan, and escalates the ones that have aged past your policy to the relationship manager with the loan number, the exception and the last three contacts attached. The write-back is the whole point. An email nobody logged never happened as far as your next examination is concerned.
Illustrative numbers for a bank closing 40 SBA loans and 120 conventional commercial loans a year with two closers.
| Assumption | Today | With write-back |
| Hours per SBA closing | 6.5 | 3.0 |
| Hours per conventional closing | 3.0 | 1.75 |
| Annual closing hours | 620 | 330 |
| Monthly 1502 preparation | 16 hrs/month = 192 | 5 hrs/month = 60 |
| Total hours | 812 | 390 |
| Capacity released | 422 hours, roughly one quarter of a full-time closer | |
Four hundred and twenty-two hours will not let you cut a position, and you should not want to. What it does is absorb a 30% increase in volume without adding the third closer at $58,000 plus benefits, and it takes the September crunch off your operations manager's back. Prove it with two numbers you can already pull: average calendar days from approval to funding, and the count of documentation exceptions found after closing. Both should fall inside a quarter. If they do not, the assistant is not really writing into your systems and somebody is still copying and pasting.
Your vendors have to support it. Ask your origination system provider, your document preparation provider and your core provider one direct question: can an approved assistant read and write in your system under our own user permissions, with a log of what it changed? If the answer is a slide deck, you have your answer. Adoption of AI among small businesses reached 66% this year and roughly 70% of owners say their people need more training — vendor readiness is now the bottleneck, not appetite.
Permissions have to be real. The assistant should hold the same rights as the person it works for, not a shared administrator login. Every write needs to be attributable in your audit trail. Your information security officer will ask this before your credit officer does, and rightly.
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.
And somebody has to own the exception list. The value of a differences report collapses the moment people start clicking through it. Name the person who reviews it and make it a step in the closing checklist, the same way the flood determination is.
Do not let anything submit into E-Tran without a human approving the screen. Your eligibility position, the use of proceeds, and the certifications on the 1919 and 1920 are yours as a Preferred Lender, and delegated authority means the SBA holds you to what was submitted under your name. Draft it automatically, submit it deliberately.
Keep the exception decisions human. Whether to waive a landlord waiver, accept a lower life insurance amount, or fund with an appraisal reconsideration pending is credit judgment and belongs to your chief credit officer, documented in the file. And keep the difficult borrower calls human — the one where the year-end financials show a 40% revenue drop is a conversation, not a tickler.
Partly, but not entirely. The work that pays first is often on the servicing side — reading your core reports, drafting exception outreach, assembling the 1502 file for review — and that can start with the reports you already produce. Meanwhile put the question in writing to your vendor and put it in your next renewal discussion. Roadmaps move when renewals depend on them.
Examiners object to unattributable changes and undocumented processes, not to software. Keep a record of who approved what and when, write the procedure into your loan policy, and keep the human approval step at the points where you make a certification to the SBA. Note also that state AI statutes now in force — Texas TRAIGA and California SB 53 both effective 1 January 2026, with several other states following — mean this belongs in your compliance review, not just in operations.
The differences report on closings. One loan type, one checklist, no writing into any system at first — just have it read the memo, the forms and the file and tell your closer what does not agree. If it catches things she would have caught at funding, then turn on the write-back.
The same disconnection shows up on your phone line: a borrower calls about a payoff or a missing insurance certificate, the message sits on a desk, and nothing is logged against the loan. CallSphere builds AI voice and chat agents that answer calls and web chat around the clock, capture what the caller needs, and book time with the right person on your team so the request lands somewhere your servicing group will actually see it. It is the same principle as everything above — the work should end up in the record, not in somebody's inbox.
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.
How AI that writes inside ServiceTitan instead of living in a separate tab removes about 29 hours a month of supply ticket matching and equipment re-keying.
Independent lots key each auction VIN into six systems. In 2026 assistants write inside the dealer management system itself. What it saves on a 60-car lot.
A homebuilder's variance purchase order is keyed four times before it hits job cost. MCP lets AI write in the system of record and stops the margin leakage.
Why roofing shops key the same roof three times, and how 2026's MCP connectors let an assistant write squares and material orders into the system of record.
In 2026 the first thirty minutes of an SBA referral happens without you. Publish the credit box facts an assistant needs, or get dropped from the shortlist.
One load of corn gets typed in four times between the field and the checkbook. What the 2026 common plug between AI and farm software actually removes.
© 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