Using AI Chatbots to Handle Client Tax Questions During Busy Season

6/21/2026

The appeal is genuine. During busy season a firm fields hundreds of repetitive client questions, most of them about status and process, and answering them consumes hours that should go to work. A tool that handles them looks like an obvious win.

It is an obvious win for a narrow set of those questions and a serious exposure for the rest — and the line between them is not where most implementations draw it.

Three Problems With the Obvious Implementation

1. Confidentiality

A chatbot that answers client-specific questions requires access to client data. For a tax practice that means tax return information, whose disclosure and use is restricted by statute and requires client consent in a prescribed form, with penalties for violation. Our posts on AI use in practice and on tax research cover the requirement.

Which creates the constraint that shapes everything else: a tool with no client data cannot answer client-specific questions, and a tool with client data has to sit inside the firm's confidentiality framework with the consent, contractual, and security controls that implies.

2. Everything it says is the firm's advice

An answer to "can I deduct my home office" is tax advice, and it is attributable to the firm.

The practitioner's diligence obligations attach. The standards governing positions attach. The firm's professional liability attaches. A wrong answer is not a customer service failure — it is the firm giving incorrect tax advice, at volume, to clients who will rely on it and who may act before anyone reviews the transcript.

And these tools produce fluent, confident, wrong answers about tax rules routinely, particularly on anything that changed recently or that depends on facts the bot does not have.

3. The problem that makes this a liability question rather than a technology one

A client message can contain something the firm is obligated to act on.

Consider what clients actually write during busy season, mixed in with status questions:

"I got a letter from the IRS, does that matter?"
"I forgot to tell you we sold the rental in November."
"That 1099 I sent was wrong, here's the corrected one."
"Don't file yet, I need to check something."
"My mother passed away in December."
"We got married in October."
"Actually I want to file separately this year."
"I moved to a different state in August."

Every one of those changes the engagement. Several change a return that may already be prepared. One is a notice with a response deadline. One is an instruction not to file.

A bot that answers the surrounding question and does not route the substance means the firm received that information and did nothing with it. That is the mechanism by which a convenience becomes a malpractice exposure — and it is worse than not having the channel at all, because the client reasonably believes they told you.

The Design Rule

The bot handles status and process. It never handles substance.

Stated as a test: if answering requires interpreting a tax rule, applying it to this client's facts, or making a judgment about their situation, the bot does not answer it.

What Actually Works

Status

Where is my return, what stage is it at, have you received everything, has it been filed, when should I expect it, was my payment received. These are the highest-volume questions in a tax practice and none of them require substantive judgment.

Outstanding items

The single highest-value application. A personalized list of what the firm is still waiting for, updated as things arrive, available on demand. It answers the question clients ask most, it eliminates the "I already sent that" exchange, and it moves work earlier — which, as our post on pre-season tool adoption argues, addresses the actual bottleneck in most firms.

Process and logistics

How to upload documents, how to sign electronically, how to pay, how to access the portal, how to schedule a conversation, who to contact about what.

General information, framed as general

What an extension is and is not, what the deadline structure is, what happens after filing. With explicit framing that it is general information and not advice about their situation, and a visible route to a person.

What It Must Never Do

State whether something is deductible or taxable
Interpret a notice or letter
Estimate or state a client's liability or refund
Advise on a filing decision — joint versus separate, whether to extend, whether to amend
Advise on an election
Confirm that a deadline applies to this client's specific circumstances
Answer a question about a transaction the client is describing
Give an answer that begins "generally" and ends without a route to a person

The Routing Requirement, Made Concrete

This is the specification that makes such a system defensible, and it is a functional requirement rather than a refinement.

The bot must recognize and escalate, with a hold on any automated response:

Any mention of a notice, letter, or correspondence from a taxing authority.

Any life event — marriage, divorce, death, birth, adoption, a move, retirement, a disability.

Any transaction — a sale, a purchase, a business started or sold, a distribution, an inheritance.

Any correction to information already provided.

Any instruction about filing — file now, do not file, file separately, extend.

Any expression of dissatisfaction, which is a complaint and belongs in the firm's complaint record.

Any expression of urgency or deadline concern.

Anything the bot cannot classify confidently, with the default being escalation rather than a guess.

And two operational requirements around it:

Transcripts must be retained and reviewed. They are client records containing confidential information — with a retention schedule, access control, and a sampling review for both answer accuracy and missed escalations. The missed-escalation rate is the metric that tells you whether the routing works.

Escalations must reach a person on a defined timeline, and the client must be told they have been routed. A queue nobody works is the same failure with extra steps.

Authentication Comes First

Nothing account-specific — not even status — before the client is authenticated to the standard the firm applies in any other channel. And transcripts should not persist in a browser on a shared device.

Tell Clients What They Are Talking To

Disclose that the assistant is automated, state plainly what it can and cannot help with, and make the route to a person visible rather than buried. Clients tolerate an automated assistant that is honest about its limits; they resent one that pretends to be a person and then cannot help.

Measure It Honestly

Not deflection rate, which counts a client who gave up as a success.

Resolution rate — was the question actually answered.

Escalation accuracy, in both directions: substance that was answered instead of routed, and status questions escalated unnecessarily.

Mis-answer rate, from transcript sampling by someone competent. There is no substitute for this.

Repeat contact rate — clients who came back through another channel shortly afterward, which is the honest indicator that a "resolved" conversation was not.

Missed-escalation count, which should be treated as an incident rather than a metric.

The Staffing Reality

A bot does not reduce headcount in a tax practice. What it reduces is interruptions — the calls and emails that fragment a preparer's day and make deep work impossible during the weeks when deep work matters most.

That is a real and worthwhile benefit, and it is a different claim from the one usually made. A firm expecting to answer client questions with fewer people will be disappointed; a firm expecting its preparers to be interrupted less often will not be.

Structured coverage is available through the AI courses for accountants and CPAs catalog, the AI for Accountants Certificate Program, AI Applications for Accountants, ethics training and professional conduct, tax practitioner regulations, penalties, and security, and the tax preparer certification courses listing.

If You Only Do One Thing

Build the automated outstanding-items list and reminder, and nothing else.

It addresses the highest-volume question, it moves work earlier in the season, it touches no substance, it requires no interpretation, it creates no advice exposure, and it can be implemented without a conversational interface at all.

Firms that start with a general-purpose question-answering bot take on the confidentiality, advice, and routing problems simultaneously in the month when they have the least capacity to manage them. Firms that start with document collection get most of the available benefit with none of the exposure.

Where This Goes Wrong

A bot answering substantive tax questions, which is the firm giving advice at scale
Client data in a tool without the consent and contractual framework
No escalation for a mentioned notice, so the firm has notice it did not act on
Life events and transactions answered around rather than routed
A correction to prior information absorbed by the bot and never reaching the file
An instruction not to file that nobody saw
Complaints resolved by the bot and never entering the complaint record
Escalations into a queue nobody works
No transcript retention or review, so the mis-answer and missed-escalation rates are unknown
Account information before authentication
Deflection rate reported as success
Clients not told they are speaking to an automated assistant
Implemented in January, which violates the timing rule in our tool adoption post
Expecting headcount reduction rather than fewer interruptions

The summary: a chatbot in a tax practice is safe for status, process, and outstanding items, and unsafe for anything requiring interpretation — and the requirement that makes it defensible is not the quality of its answers but its ability to recognize when a client has told you something you must act on and get it to a human. Build the outstanding-items list first, and be very slow about anything beyond it.

Frequently Asked Questions

What can an AI chatbot safely handle in a tax practice?

Status questions, process and logistics, and — most valuably — a personalized list of outstanding items the firm is still waiting for. The design rule is that the bot handles status and process and never handles substance: if answering requires interpreting a tax rule or applying it to the client's facts, it does not answer.

Why can't it answer whether something is deductible?

Because that answer is tax advice attributable to the firm, carrying the practitioner's diligence obligations, the standards governing positions, and professional liability. A wrong answer is not a service failure but incorrect advice delivered at volume to clients who will rely on it — and these tools produce confident wrong answers routinely on anything that changed recently.

What is the most serious risk beyond wrong answers?

That a client message contains something the firm must act on — a notice received, a transaction the firm did not know about, a correction to information already provided, an instruction not to file, or a life event. A bot that answers the surrounding question without routing the substance means the firm received that information and did nothing, which is worse than having no channel because the client reasonably believes they told you.

What must a bot recognize and escalate?

Any mention of a notice or correspondence from a taxing authority; any life event; any transaction; any correction to prior information; any instruction about filing; any expression of dissatisfaction, which is a complaint for the firm's records; any urgency; and anything it cannot classify confidently — with escalation as the default rather than a guess.

How should such a system be measured?

Not by deflection rate, which counts abandonment as success. Use resolution rate, escalation accuracy in both directions, mis-answer rate from transcript sampling by someone competent, repeat contact through other channels, and missed escalations — which should be treated as incidents rather than as a metric.

What should a firm build if it does only one thing?

The automated outstanding-items list and reminder. It addresses the highest-volume client question, moves work earlier in the season, touches no substance, creates no advice exposure, and does not require a conversational interface at all — delivering most of the available benefit with none of the risk.