There’s a debate running through every industry right now: is AI actually worth the hype and the budget? You’ve seen the headlines questioning the ROI. In today’s issue, I want to skip the abstraction and talk about what we actually see working in commercial lending — specifically SBA lending — from our conversations with the teams underwriting these loans every day.
Here’s the misconception worth correcting. The most concrete use case our customers describe isn’t “AI for productivity.” It’s AI for risk detection. And in SBA lending, that distinction is becoming the difference between a clean guaranty and a repair.
This connects directly to the new show we are launching called Unglamorous; stay tuned to this channel for more information. We have to do it right. The first episode is deep dives into the risk hiding where lenders don’t look — tax debt that never shows up in a lien search. This is the same lesson one layer over: the compliance risk that never shows up inside any single document. The pattern of this whole show is that the thing that hurts you is the thing your standard process isn’t built to see.
Three ways we’re seeing SBA teams put that to work:
1. AI as a pre-audit tool. Like most regulated industries, SBA lending adopts new tools well after consumers do. AI is changing that completely, but not in all the ways yet; as a pre-audit screening tool, it is already happening. So what’s notable is how the sharpest teams are framing it: as a way to find inconsistencies before an examiner, a guaranty-purchase review, or an enforcement action does. SBA lending is getting more regulated, not less — the current administration’s “stop the bad guys” posture lands directly on bankers who just want to close loans. Early detection has never mattered more. Running your own file through AI before it moves down the assembly line is a low-risk way to surface problems while they’re still cheap to fix.
2. The hidden compliance stack is full of contradictions. A simple but powerful idea came up in one discussion: most institutions have policy, procedure, credit memo, and approval language that doesn’t line up cleanly — and nobody has time to read all of it side by side. AI surfaces those mismatches fast, saving underwriting, processing, and closing teams real hours on the back end of what’s become an increasingly document-heavy process.
3. Referral-partner oversight is now searchable. A leader at one of the largest SBA lenders in the country raised this one: using AI to review referral-source websites and public-facing materials for misleading SBA language. That’s a sharp break from the past, because it pushes compliance beyond the credit file and into third-party reputation risk. And it matters more every month — AI has made it easier than ever to spin up a polished, non-compliant referral website.
The thread connecting all three: most compliance failures aren’t visible inside one document. They only appear when you compare policy, procedure, loan structure, and outside marketing claims against each other. That cross-document view is exactly what humans miss and what AI is unusually good at.
One counterintuitive takeaway. The lenders benefiting most from this aren’t the ones with the most dialed-in systems or the fastest process — they’re the ones with the most fragmented documentation. AI is quietly displacing traditional OCR across industries, and while it still feels strange that reading a document well requires an expensive model, that’s the world we’re in; a birdy told us today that Claude’s new Fable model is the best at reading and creating PDF files. As one SBA leader put it: prompts are a premium. The more efficient you are at getting exactly what you need from these tools, the further ahead you’ll be.
📄 Pre-filled 8821s, one click. Generate an authorization with your taxpayer’s name and TIN and our designee info already filled in — per entity or individual on the request. Download, sign, done. No more blank forms.
📍 Real-time order status. Every request now shows exactly where it is — from authorization to the IRS to delivered in your inbox. No more wondering what’s pending.
👨👩👧 Bundle a whole client group in one order. Put the individual and their related entities — including K-1 business owners — on a single request. One order covers every party, with no re-keying the same client three times.
💳 New: prepaid credits with volume pricing.
The compliance contradictions above are one blind spot. The tax-debt blind spot from this week’s episode is another — and it’s the one we close. ModernTax pulls verified IRS transcript data at the source, not from a borrower’s PDF packet, so the one thing you never have to wonder about is whether the tax picture is real.
And for the high-volume, small-dollar, government-guaranteed lenders reading this: we’re building a fully autonomous, dedicated verifier purpose-built for SBA Express workflows — the kind of per-pull economics above, at machine speed. More on that soon — reply if you want early access.
OpenAI + Thrive build self-improving tax agents — Codex-powered Tax AI processed 7,000 returns at up to ~97% draft accuracy with ~31% less prep time. Worth watching for anyone in tax-adjacent infrastructure. (Aravind Srinivasan & Samay Shamdasani, Thrive Holdings / OpenAI)
Crete Professionals Alliance rebrands to “Current” — the same Thrive-backed accounting roll-up, leaning hard into AI-led tax work. (Accounting Today)
A New Era for SMB Tech (Shirlawn Capital, Substack)
Investing in Lassie (a16z)
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.