AI Insights Financial Services
Fred: A Client Assistant That Won’t Give Tax Advice or Leak a Return
Talk to Fred
Ask Fred about Financial Services
This is the same Fred you would put on your own site. Ask about Financial Services, compliance, or how the guardrails work. Fred listens.
A prospective client messages your firm’s site in March: "can I deduct my home office and the new truck, and here’s my income so you can tell me what I’ll owe?" A generic chatbot answers with a position and a number, and to be helpful, it absorbs the income details into the conversation. It just gave tax advice your firm is now associated with, on facts it does not have, and it took in taxpayer information without the consent and safeguards that govern how a preparer may use it. The firm wanted a tool to book engagements during the season. It got one that advises clients and mishandles return data.
That is the exposure the threat and standard pieces in this series cover. This article is the answer: what a compliant AI assistant for accounting firms actually looks like, and why booking the work while routing every position to a credentialed preparer is what keeps the firm clean.
The Real Choice Is Governed or Ungoverned
Most firm owners frame the question as whether to put AI on the site at all. Clients settled that already. They expect an instant answer on services, deadlines, and how to get started, especially in season, and the firm that makes them wait loses them. The decision that matters is whether the assistant is governed.
An ungoverned chatbot answers "can I deduct this," which is a tax position that practice standards like Circular 230 hold a credentialed preparer to, and it takes in or echoes the taxpayer return information that Section 7216 restricts how a preparer may disclose or use. A governed assistant captures the same client, books the engagement, and routes every position to a preparer. Same speed, none of the exposure.
How Fred Holds the Line
Fred starts from the opposite default of a general-purpose chatbot. A generic bot answers anything unless told otherwise; Fred answers only what it is cleared to, and the boundary lives in the system rather than in a prompt it can drift away from.
In practice Fred reads from your own firm content, books the consultation, and routes anything touching a deduction, a filing position, a liability estimate, or specific tax advice to a credentialed preparer. It does not tell a client what they can deduct, because that is a position a CPA or EA stands behind under Circular 230. It is built to handle taxpayer information with the disclosure-and-use limits of Section 7216 in mind, rather than soaking up return data in an open chat and pointing clients to your firm’s secure intake instead. Fred runs more than 50 industry guardrail packs, and the accounting pack is built around the tax-advice line, taxpayer-information handling, and clean routing to the credentialed team.
The difference shows up under pressure. Ask a prompt-instructed bot "what can I write off" three different ways and it eventually answers the phrasing it was not warned about. Fred does not depend on recognizing the wording; it decides what is allowed out before the answer exists. "Will not" is a suggestion. "Cannot" is an architecture.
Fred vs. a Generic AI Chatbot
| Situation | Generic AI Chatbot | Fred |
|---|---|---|
| "Can I deduct this?" | Answers with a tax position on facts it lacks | Routes the position to a credentialed preparer; books the work |
| "Here’s my income, what will I owe?" | Estimates a liability and absorbs return data | Declines the estimate; routes to secure intake |
| "Should I file as an S-corp?" | Improvises planning advice | Hands the planning question to your CPA or EA |
| Where the rules live | In a prompt the model can drift from | Built into the system; enforced before output |
| A deduction question, reworded | Eventually answers when phrasing changes | Held the same way regardless of wording |
| Who owns the tax answer | Effectively the chatbot, and the firm | A credentialed preparer, every time |
The table is the whole argument in one screen. A generic tool is helpful right up to the moment helpful becomes a tax position or a mishandled return. Fred is helpful everywhere that carries no risk and structurally silent everywhere that does.
What Your Firm Actually Gets
Set the compliance framing aside and look at the business case. Fred answers the routine questions that swamp the front desk in season, what you charge, what documents you need, when the deadline is, how to get on the calendar, whether you handle a particular return, and it answers instantly, at any hour, in the client’s own words. It books the engagement with the scope attached, so a preparer opens a new client already knowing whether it is a simple return or a multi-entity mess.
There is a capacity angle that matters most in season. Your credentialed staff have a finite number of hours when it counts, and every one spent answering "what do you charge" for the tenth time is an hour not spent on billable work. Fred absorbs that volume and books the qualified consultations, so your preparers spend the crunch on returns, not reception. Clients get an instant answer; the firm gets its season back.
So the question is not whether your competitors will run AI on their sites. They will. It is whether yours books the work and routes the advice before a chatbot gives a client a position your firm has to answer for.
Frequently asked questions
Can a compliant assistant answer any tax questions?
It can answer the general, published questions, your services, fees, deadlines, what documents you need, how to get started, and route everything else. What it does not do is tell a client what they can deduct, estimate a liability, or recommend a filing position, because those are tax positions a credentialed preparer stands behind. A compliant assistant like Fred captures the question and routes it to a CPA or EA, so the client gets real advice from someone accountable for it.
How does a chatbot create a problem with client return data?
Taxpayer return information is governed by rules that restrict how a preparer may disclose or use it. A chatbot that absorbs a client’s income and figures into an open chat, or that echoes them back, is handling that information outside your controlled process. A compliant assistant is built to keep return data out of the chat and point clients to your secure intake instead, so the information is handled the way the rules require.
How is Fred different from putting rules in a generic chatbot's prompt?
A prompt is an instruction the model can drift away from when a question is phrased in a way it did not anticipate. Fred enforces its boundaries at the system level, deciding what is allowed before it responds, so a tax position or a data exposure never gets generated regardless of wording. It is the difference between a bot that usually behaves and one that is built so it cannot give tax advice or mishandle a return.