Circular 230, Section 7216, and Your Firm’s Website: A 2026 Tax-Practice Playbook

A good preparer is trained to ask three more questions before answering one. A general-purpose assistant is built to do the opposite: give a confident, complete answer right now. That single difference is why tax practice and a chatty website sit uneasily together. The questions that sound general, "can I deduct this," "how do I handle that 1099," have no honest general answer, and the moment a firm’s site gives one, the firm has offered professional advice it has to stand behind. The 2026 standard is about letting the assistant do the genuinely useful work while keeping every substantive tax answer with a person who signs the return.

This guide is the companion to the threat piece. The threat side covers the website that gave tax advice no one would sign. This one covers the standard: what a compliant firm assistant handles, what it routes to a preparer, and how it protects client information the way the tax code requires.

A Tax Answer Is Regulated Work

"Can I deduct that" depends on facts the website does not have and judgment it is not qualified to make. Advising on federal tax matters is regulated practice. Practitioners operate under Treasury’s Circular 230 rules of practice, and state boards of accountancy license who may hold out as a CPA and offer this work. An assistant handing out deduction strategies is doing the firm’s regulated job without the firm’s judgment, and the firm wears the outcome when a position turns out wrong.

So the compliant standard treats substantive tax questions as routed, full stop. The assistant can capture the engagement, explain how the firm works, schedule a consultation, and answer logistics about documents and deadlines in general terms. What it does not do is tell a prospect whether something is deductible, how to treat a transaction, or what position to take. Those are judgments a credentialed preparer makes with the actual facts in front of them.

Client Data Has Its Own Statute

The friendly intake is its own exposure. To be useful, an assistant draws out income, filing status, and the messy specifics of someone’s finances, and tax return information carries a specific federal protection. The rules under Section 7216 of the tax code restrict how a preparer may use or disclose a client’s return information, with penalties attached for getting it wrong. A generic widget that funnels those details into a vendor’s system, outside the firm’s controls, can turn a convenience feature into a disclosure problem the partners never approved.

The standard here is to treat the assistant as part of the firm’s data estate. Return information stays inside systems the firm controls, under the safeguards and agreements 7216 and ordinary professional duty require. Convenience does not lower the bar; the protection follows the data wherever the conversation sends it.

What AI Compliance for Accounting Firms Comes Down To

Pull the pieces together and the standard is short and strict. The assistant answers logistics, scheduling, and general information about the firm’s services. It refuses substantive tax advice, deductions, treatments, positions, planning, and routes those to a preparer. It keeps client return information inside the firm’s controls under 7216-aware handling. And it leaves a record of what it told people, because "helpful" and "defensible" are not the same thing, and a firm wants to be able to show the assistant never gave advice no one signed.

That last distinction is the heart of it. A bot built to give a complete answer will produce one even when the right professional move is to gather more facts first. A compliant deployment does not rely on the assistant choosing restraint. It removes the ability to give the advice at all.

Why a Disclaimer Cannot Meet the Standard

The usual shortcut is a prompt and a fine-print line: tell the assistant never to give tax advice, add a "not professional advice" notice, and treat the boundary as set.

It is not set, because of how the model behaves. It follows an instruction when the request matches the wording it was warned about, and prospects in March do not use that wording. You tell it never to give tax advice. The prospect does not ask for advice. They press, "just ballpark it, can I write off the home office or not?" The model reads someone who wants a yes or no and gives one, because resolving the question is its default and a prompt is only a request to hold that default back. The rule was loaded the whole time. It just never recognized the sentence that crossed into a tax position.

That is the gap between an instruction and a standard. An instruction asks the model to behave. It does not stop the model from speaking, and it does not make the answer defensible once a client has relied on it. A real boundary is enforced in the system and decides what the assistant may say before it answers, so a deduction strategy or a tax position never reaches the prospect no matter how the question is phrased. "Will not" is a suggestion. "Cannot" is an architecture.

What a Compliant Deployment Looks Like

Meeting the standard does not mean a firm gives up the assistant that captures engagements and answers questions after hours. It means running one built to keep substantive tax work with a preparer and client data inside the firm’s controls.

Fred is built that way. It answers from your own firm content, captures the engagement, handles scheduling and logistics, and routes every substantive tax question to a preparer. It runs more than 50 industry guardrail packs, and the accounting pack is built around Circular 230 practice, CPA licensure, and the Section 7216 protections on return information. Fred does not tell a prospect whether something is deductible or how to take a position. It cannot. It handles the intake, protects the data, logs every exchange, and gets the regulated work to a credentialed person.

The aim is not a smarter-sounding website. It is one that cannot give tax advice the firm would never sign.

Frequently asked questions