Build Your CPA Firm a Second Brain Before Busy Season

It's the end of October, so a CPA firm has about two quieter months left to fix whatever drove everyone crazy last spring. Here's a good candidate: the interruption.

Picture a February afternoon. You're halfway through a messy S corp return when someone from the front desk leans in. A caller wants to know whether the firm handles multi-state returns, what an extension costs, and what a new rental-property client should bring. None of that takes your expertise. The answers already exist, scattered across an engagement letter template, a fee sheet, a checklist and an email you wrote two years ago. You're just the only one who knows where they are.

That's the job for a Second Brain: an AI knowledge system that reads your firm's own documents and answers your staff's questions from them. Your people ask, and it answers and shows where the answer came from.

Why the build has to be careful

On June 24, the IRS Office of Professional Responsibility issued OPR Alert 2026-19, its introductory guidelines for AI use in federal tax practice. It doesn't ban AI. It says existing Circular 230 duties still apply. Practitioners answer for AI-assisted work, need to verify facts, citations and calculations, and must handle client data "using only secure, enterprise-approved AI." The people overseeing a firm's tax practice are responsible for adequate procedures under Section 10.36, and that extends to AI.

The Tax Court supplied the cautionary tale in February. In Clinco v. Commissioner, three of the four cases cited in a taxpayer's brief appeared not to exist. Judge Holmes imposed no sanctions, and the court didn't establish that AI was used, but he wrote that submitting fictitious caselaw "is a recipe for sanctions." That was an attorney's brief, not a CPA's work, but the lesson travels: if you can't trace an answer, you can't rely on it.

Firms are already moving this way. In a Virginia Society of CPAs survey this year, 47% of respondents said they use AI for internal knowledge management or research. Only 144 members answered, so it's a small sample, but it shows where some of your peers are looking.

What to build

Three pieces, in this order:

• The spec. One page covering who can ask, what the system may answer and what it must refuse. "What does our firm do?" is in scope. "What does the tax code say?" goes to your research tools and a credentialed human.
• The ingestion pipeline. Engagement letter templates, fee schedules, organizer checklists by client type, document request templates, your procedures manual and approved service descriptions. Every file gets an owner and a review date, and last year's version comes out when the new one goes in. Skip that step and the system will quote an old fee with total confidence.
• Grounded chat. Every answer cites the document and section it came from. If your documents don't cover the question, it says so instead of guessing.

Leave client returns out of version one. Section 7216 limits how preparers can use and disclose return information, so bringing that data in is a separate decision for the partners and your counsel. Your firm's operating knowledge is useful enough on its own.

How to know it works

Don't trust the demo. Before launch, have your most experienced person write 30 real questions from last season, with the correct answer and source document for each. Include a few the system shouldn't be able to answer. Then score it: right, wrong, and whether it declined when it should have. Wrong answers usually trace back to a stale or contradictory document, which was worth finding anyway.

Once it's live, have someone open the cited source on a weekly sample of answers and keep a log. That catches drift, and it gives you something concrete to show when someone asks how the firm supervises its AI use.

The payoff is modest and real. The front desk can answer that caller from a document you approved. A prospect gets a straight answer about your services on the first call. Document requests go out with the right checklist. And you stay in the return.

A Lima firm can start small: a dozen core documents in November, the 30-question test by mid-December. Paglow Automations builds this for Allen County businesses, from the spec through the grounded chat.