Last month I made the case that AI’s most valuable job in SBA lending isn’t productivity — it’s catching what the examiner will. Point a model at your own policies, procedures, and credit memos, and it surfaces the contradictions before a guaranty-purchase review does.
That still holds. This week the IRS added the other half of the story that is relevant to us at ModernTax — and it’s the half that changes how we think about AI more broadly.
The IRS published its first formal guidance on AI in tax practice. It doesn’t ban anything. It does one thing: it makes you responsible for what your AI produces. Under Circular 230, you have to independently verify every fact, citation, and calculation a model generates. “The software said so” is not a defense — and courts are already sanctioning professionals who learned that the hard way.
Here’s why that’s a real build on last month, not a repeat:
Catching contradictions is about whether your documents agree with each other. The IRS is asking a harder question — whether your documents agree with reality.
Those are not the same test. And AI is only good at one of them.
A model can read your file and confirm the tax return, the spread, and the credit memo all tell the same story. What it cannot do is confirm the story is true.
Two documents can be perfectly consistent and both wrong.
Watch how careful lenders actually verify income, and you see the gap in real time: two screens open side by side, the borrower’s tax return next to the IRS transcript, checked line by line — the first 20 lines, Schedule C, Schedule K. Why the paranoia? Because the overwhelming majority of problems trace back to amended returns and version mismatches. The borrower’s PDF can be internally flawless, fully self-consistent, beautifully formatted by AI — and still not be the number the IRS has on file.
An AI-generated credit presentation doesn’t close that gap. It widens it. A more fluent, more confident memo built on a number nobody reconciled against the source is just a better-looking wrong answer.
This is exactly why, for all the hype, income verification is the one place lenders won’t let AI run unsupervised. As one commercial-lending platform put it to a room of bankers: “I heard fintechs say you can upload tax returns and generate a full credit presentation in five minutes. No bank in this country is going to go with that. They need to look at the financials. A lot of them want borrower character — do you trust this person?”
The word that keeps coming up in these conversations is “black box.” Lenders will let AI organize, format, and cross-check all day. They will not outsource the judgment about whether the underlying facts are real — because when the facts are wrong, the model can’t tell you, and the guaranty is still yours.
And the risk isn’t hypothetical. Feed a system unverified third-party data, and it will confidently build on garbage — we’ve heard lenders describe revenue figures off by an order of magnitude and scoring models that didn’t drive a single decision. Good AI on bad inputs isn’t insight. It’s a faster path to a confident mistake.
Last month’s takeaway was that the messiest files benefit most from AI’s cross-document view. Here’s the sharper version now that the IRS has weighed in:
The better AI gets at drafting, the more dangerous your unverified inputs become.
Fluency is a risk multiplier. When memos were clunky, a shaky number looked shaky. Now the output is polished, consistent, and authoritative-looking by default — which means the one weak link left, the unverified fact, is better hidden than ever. The value migrates to the input you can actually stand behind.
AI can make your file agree with itself. Our mission is to make sure it agrees with the IRS. ModernTax pulls verified transcript data straight from the source — not the borrower’s PDF packet — so the fluent, AI-drafted memo on top is built on a number that’s actually real. As the drafting layer commoditizes, the verified input is the part that keeps appreciating.
And for the high-volume, small-dollar, government-guaranteed lenders reading this: we’re building a fully autonomous verifier purpose-built for SBA Express workflows — machine speed on the input AI can’t fake. Reply if you want early access.
The last few weeks have been our biggest shipping stretch yet. Here’s what’s live for you now.
✍️ Enter taxpayer info once — get a signature-ready 8821 instantly. No more retyping the same name, TIN, and address into a blank form. On every new request — one at a time or a whole CSV/Excel batch — the system now auto-generates a fully populated Form 8821 for each entity and emails it straight to you. Collect the signature with your own tools, upload the signed copy, and we get to work. It scales to bulk authorizations without the busywork.
⚡ Faster results — we reuse records you’ve already pulled. When we already have an entity’s transcripts on file, new requests for that same entity now fulfill instantly instead of waiting in the IRS queue. A self-healing background sweep catches these automatically, so you get the records the same day whenever they already exist.
🔗 Self-service links for your clients — no login required. Secure, no-login pages your clients can open from a single link: complete their filing intake, ask questions about their estimates in a private thread, and review + prepay for work. Everything’s token-gated, so their information stays protected.
📊 Preliminary tax estimates from IRS records — Our new prep engine reads a taxpayer’s official IRS wage & income records and builds a first-pass estimate across the open years — a plain-English snapshot of what’s owed and the path forward, ready for an expert to finalize.
📠 Send signed forms straight to the IRS — from your dashboard (rolling out). Experts can now fax a signed Form 8821 to the IRS from within the portal, entering the exact number an agent gives them during the call, with live delivery confirmation on screen — no external fax tool required.
🛠️ Under the hood — Big scanned-document uploads now go through reliably, invoices are cleaner and land on time, and ongoing account monitoring got smarter authorization controls — so the whole pipeline is faster and more dependable.
AI adoption among tax firms nearly doubled in a year — 60% now use AI for tax research at least weekly, up from 33% in 2025, per a Blue J/CPA.com survey. The profession is moving fast, which makes the verification question above more urgent, not less.
When does the AI build-out actually pay for itself? Tomasz Tunguz models the industry reaching breakeven around 2029.
Someone, eventually, has to blink: the WSJ on an AI spending war where capex keeps outrunning revenue.
A very 2026 milestone: a startup that builds AI agents used one of its own to help raise $100 million.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.