Yes, obtain a Section 7216 consent before sending a client's tax return information to an AI provider when the disclosure is not specifically permitted without consent. A preparer does not need a blanket AI consent simply to prepare a return. The answer depends on who receives the information, why, and whether a regulatory exception genuinely covers that workflow.
The critical mistake is treating the tool as an invisible extension of your office. An uploaded return, pasted figure or connected file may reach a separate company and its subcontractors. A signed engagement letter, a vendor's privacy notice and a client's casual agreement to “use AI” are not automatically a valid Section 7216 consent.
Trace the information before choosing the form
Start with the route a single field takes. Record the source document, the employee who enters it, the AI service that receives it, any connected application, the output, and where that output is stored. “We use AI” conceals the part that matters: a return extract may pass through an application provider, an AI model provider, a logging service and a human support queue. Ask the vendor to identify its legal entity and its subprocessors rather than guessing from its brand name.
Tax return information is broad. The regulations define it as any information furnished in any form or manner for, or in connection with, preparing the taxpayer's return, including the taxpayer's name, address or identifying number. That can include a client's name, income figure, occupation, uploaded worksheet, note about a deduction, or an apparently harmless question that reveals a distinctive transaction. Removing a tax identifier alone does not make a copied paragraph safe. The return redaction tutorial covers that narrower preparation task; here the question is whether a real client's information may leave your control at all.
For each workflow, write one sentence: “A preparer copies [specified fields] from [specified records] into [named service] to [specific task], and the output goes to [reviewer].” If you cannot finish that sentence, pause. The omission may hide a new purpose such as marketing analysis, a new recipient such as a plugin, or a whole-document upload where a short summary would suffice. Obtain a qualified tax or legal review of the route before drafting client-facing permission.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowFive routes that produce different answers
1. A synthetic training prompt. Suppose a practice asks an AI tool to explain how to word a generic records request using fictional amounts and no client details. There is no actual client's tax return information in the prompt. Section 7216 consent for that hypothetical disclosure is not the issue. Staff still need a rule against “helpful” later edits that paste a real client email or worksheet into the same conversation.
2. An internal analysis in the preparer's firm. A preparer may use authorised information to prepare the return, subject to the statute, regulations and professional duties. Do not assume every internal AI system is internal in the legal sense: a cloud tool may involve a distinct recipient. Check its contract, processing arrangement and the regulatory exception being relied on. A tool that runs under the firm's account is not automatically the firm's employee.
3. A third-party assistant for return work. Imagine an employee uploads a client's bank-interest worksheet to an outside AI provider to categorise entries for return preparation. This is a disclosure question even if the result never leaves the practice. An exception may apply to a particular auxiliary service or another preparer in the circumstances described in the regulations; do not stretch an exception to cover an ordinary general-purpose chatbot without analysis. Record the precise legal basis and the data minimisation step.
4. A separate advisory or marketing use. A practice wants an AI tool to scan completed returns and generate leads for a retirement-planning service. The purpose has changed from preparing returns. The regulations treat uses and disclosures beyond preparing the return as needing consent unless a specific exception applies, and a preparer generally may not make its services conditional on the client signing. A general engagement letter is especially weak here because the client may have agreed only to a return. Keep a refusal path that leaves the basic preparation service available where the rules require that freedom.
5. A preparer or processor outside the preparer's jurisdiction. Revenue Procedure 2013-14 prescribes extra consent wording when return information may be disclosed to a preparer located outside the United States, and even a disclosure to a colleague in the same firm needs consent when that colleague works abroad. If the Social Security number is included, both preparers must maintain an adequate data protection safeguard and the consent must say so. The exact rule turns on the recipient's role and location, not a marketing statement that the AI company has a domestic headquarters. Ask where people and systems that process the information are located and have specialist counsel examine the arrangement.
These are screening examples, not a claim that one vendor category always needs consent. A processor's legal relationship, the task, the information transferred and the relevant exception all matter. A six-column register makes the decision auditable: workflow, information, recipient, purpose, possible exception, and approval. If the exception column contains “probably because software”, it is not ready for client data.
What a valid consent has to communicate
For Form 1040-series returns, Revenue Procedure 2013-14 prescribes format and mandatory language. A separate written document is generally required for each disclosure or use, subject to the rule for multiple disclosures or multiple uses in one document. A disclosure consent and a use consent must be separate documents. The prescribed statements vary according to whether the disclosure is outside preparation or to another preparer assisting preparation. Do not copy a generic web form and add “AI” to its title.
The consent must let the client understand what information is involved, who receives it and why. Identify the recipient accurately enough that the client knows the destination. Describe the specific data category and purpose, rather than “information as needed to improve our services”. If you propose to disclose an entire return, the procedure calls for a way for the client to request a more limited disclosure. Narrowing the feed before requesting consent is often simpler and more respectful than seeking permission for everything.
State duration in the approved form. The revenue procedure says that when a client does not specify a duration, a consent is valid for one year from signature; a shorter defined period may make more sense for a single task. A valid consent must be affirmative and signed. An unticked opt-in that the client actively selects is different from a box preselected by software. Paper and electronic presentations have separate technical rules, such as 12-point type on paper and an approved electronic signature method. For a PIN signature, for example, the client must enter a PIN of at least five characters personally; software that pre-fills it or accepts a single button click does not meet the rule.
The client must see the prescribed warnings and complaint information, with the sequence required by the relevant consent type. There are special statements for disclosure to an overseas preparer, and extra safeguard requirements when certain identifiers are transferred. The exact mandatory wording belongs in a form reviewed against the current revenue procedure. This tutorial deliberately gives AI-specific insert wording, not a complete signable legal form: omitting the mandatory statutory statements would mislead a preparer into deploying an invalid document.
AI-specific wording to insert into a reviewed disclosure form
Use the following as an illustrative information block inside the correct, legally reviewed disclosure consent. Replace every bracket with a confirmed fact, or remove the workflow. Do not leave blanks for staff to fill after signing. The business names and figures below are examples, not assertions about any product's data handling.
For the purpose of [categorising the expense entries in your current return workpapers], [preparer's legal name] requests your permission to disclose [the expense description and amount fields from the current-year worksheet, excluding your name, contact details and tax identifier] to [AI provider's exact legal name]. The provider will process that information to return a draft classification for review by our preparer. We will not use this permission to send your full return, to generate marketing offers, or to make an automated final tax decision. The requested consent lasts until [specific date].
The wording identifies the action but does not assert that the provider truly excludes training, retention, support access or onward disclosure. Verify those matters in the contract and record the vendor's answer. If a vendor retains prompts for quality review, say so where material. If subcontractors receive the data, identify the disclosure chain with counsel. A client cannot knowingly agree to a pathway the practice has not mapped.
A second illustrative block concerns a different purpose: “We request permission to use the income and expense totals from the return you supplied for [specified advisory analysis], to produce a draft cash-flow discussion for you. This permission does not authorise disclosure to [AI provider] unless a separate disclosure consent covers that recipient.” This belongs, if appropriate, in a use consent reviewed for its own mandatory language. Do not merge it with the disclosure block merely because one AI workflow performs both actions.
A third block tests how specific the recipient needs to be: “Provider: [exact company name]; processing service: [named product]; subprocessors relevant to this workflow: [reviewed list or referenced schedule]; permitted data: [two fields]; purpose: [one task].” If procurement cannot fill the first three items, the consent drafting process has revealed a vendor due-diligence gap. The AI vendor questions tutorial gives a broader buying checklist.
A 14-client pilot, calculated from the data path
Consider an illustrative four-person tax practice preparing 14 returns during a pilot. It wants AI to suggest categories for unusually worded business expenses. Each client's worksheet contains 120 expense rows. The team first proposes uploading every full worksheet: 14 × 120 = 1,680 rows, with payee names, notes and identifiers mixed in. That volume is not a reason to obtain a broad consent; it is a reason to redesign the input.
The manager samples one worksheet and finds only 18 ambiguous descriptions that actually need help. A staff member copies those descriptions into a controlled review sheet and replaces payee names with generic labels. The proposed transfer becomes 14 × 18 = 252 descriptions, and only if those fields are truly free of client information. The reviewer checks a sample for a distinctive description such as “settlement with former partner” that might still reveal a client fact. That entry stays out of the tool, even without a name.
Suppose three clients have return information in the remaining descriptions, and counsel concludes that a disclosure consent is required for this particular provider and processing arrangement. The practice seeks a properly drafted, signed consent from those three before sending any of their data. One declines. The staff categorise that client's entries manually. Two consent. For the remaining 11, the practice uses only genuinely synthetic or sufficiently de-identified material after a documented check. The pilot does not treat consent as a substitute for minimisation.
The expected benefit is modest and measurable. If manual classification takes six minutes for each of 18 ambiguous entries, 14 worksheets represent 1,512 minutes of initial review. AI may suggest labels quickly, but the practice still has to verify every category against documents and tax treatment. It records preparation time, correction count and any rework. If the tool saves 12 minutes per consented worksheet but adds 20 minutes of vendor checks, consent handling and review, the pilot may not pay off. The numbers are illustrative; the method avoids a false promise of automatic savings.
The practice keeps a row-level record: which client's data were used, which signed document authorised the disclosure, what was uploaded, when it was uploaded, which employee approved it, and where the output went. If a client withdraws future permission or a vendor changes its terms, the manager can locate affected work. This is a more useful control than a folder named “AI consents” with no link to actual uploads.
Five small situations that expose a bad consent
An email-writing request. A preparer pastes “Please send [first name] a reminder for a missing charitable contribution receipt” into AI. The first name and missing receipt may reveal a client matter. Replace the name with a placeholder and state only the generic request. Then use a human to merge the actual client details in the firm's approved communication channel. That can avoid a disclosure while still using AI for phrasing.
A spreadsheet connector. A workbook contains 40 clients, but the user asks AI to classify only five rows from one return. A connector may expose more than the visible selection. Test its permission scope with dummy data and check whether it can read the whole workbook or folder. If the scope is wider than the task, split out an approved extract or do the job manually. “I selected five cells” is not evidence that only five cells were disclosed.
A staff training example. An employee changes the client's name but retains exact figures, a rare asset description and a distinctive transaction date. The training prompt is still recognisable to colleagues and may be tax return information. Substitute invented figures, dates and circumstances throughout. A risk register for AI workflows can record the chance of re-identification and the control owner.
A new model inside an old product. The practice has a consent naming one recipient and task. The vendor adds a third-party assistant for document extraction. Do not infer that the old form follows the new route. Recheck the contract, recipient, purpose and technical access. If the authorised disclosure has changed, seek specialist advice about a new consent before enabling the feature.
An overseas support queue. Sales material describes local data hosting, but support staff elsewhere can inspect failed prompts. Hosting location and human access are different questions. Ask for a written answer about support access, escalation copies and subprocessors. The specific overseas-preparer rules apply to the circumstances set out in the revenue procedure; a broad assertion that all offshore processing has identical wording would be unsafe.
Build a consent decision sheet staff can actually use
Put a short decision sheet beside the upload control. The fields should be completed before a live case is used, not after a staff member has already obtained an attractive AI answer. Keep the legal basis field distinct from the business reason. “Faster” is a business reason. A documented regulatory exception or a valid signed consent is a legal basis. A firm may also have professional confidentiality duties and client contract limits beyond Section 7216.
| Question | Illustrative completed entry | Evidence to retain |
|---|---|---|
| What is the task? | Suggest category for 18 ambiguous expense descriptions. | Approved workflow description and sample input. |
| What leaves the firm? | Description and amount only; no return or identifier. | Export specification and a checked sample. |
| Who receives it? | Named AI provider and reviewed processors. | Contract, processing schedule and access answer. |
| Why may it leave? | Consent for this disclosure, if required after review. | Signed, dated form linked to the case. |
| Who checks the result? | Assigned preparer checks each category. | Review log and corrections. |
| What happens on refusal? | Manual categorisation. | Staff instruction, with no client penalty where required. |
Make the consent state match the technical state. A “consent pending” case must not be available in the AI upload queue. A signed consent should permit only the approved information and purpose, not every file in the client folder. If a form has expired, an upload rule should block that client's case until the status is reviewed. A spreadsheet with a manual status column is acceptable for a small pilot if an owner checks it; elaborate automation is unnecessary before the decision process works.
Review the vendor and the human process together
Ask how long prompts, files and outputs are kept; whether they train models; who can inspect them; what subcontractors receive them; whether support logs contain the content; how deletion works; and how an incident is reported. Get answers in contractual or documented form. A sales demonstration is not a reliable description of the production tenant. Check the actual configuration that staff use, including sharing defaults and connected apps. The AI privacy terms tutorial helps read those claims, but the Section 7216 decision remains specific to return information and the proposed use.
Restrict access to the tool and keep approved prompts in one place. Staff should know that the AI output is a draft, not a tax conclusion. The Office of Professional Responsibility's June 2026 alert on responsible AI use in tax practice says practitioners must independently authenticate factual and legal information, tells them to handle client data only in secure, enterprise-approved AI, and reminds them that preparer penalties apply to unauthorised use or disclosure of return information. A convincing answer is not proof that its reasoning, citations or calculations are sound. Pair the permission decision with a review protocol for the resulting work.
When a client says no, record only the refusal and the alternative method needed to serve them; do not ask AI to explain the refusal using their return details. If several clients decline, examine whether the tool still has a worthwhile use on synthetic examples, public guidance or administrative copy. Consent should be a meaningful choice, not a campaign to persuade reluctant clients. The team can use the same decision sheet for each new workflow rather than assuming permission follows the vendor account.
Check a completed form before the first live upload
Have a qualified tax adviser or lawyer compare the final document with the current regulation and Revenue Procedure 2013-14, including the exact mandatory statements, consent type, signature method and any overseas or identifier rule. Then conduct a dry run with fictional data. Can the client understand the recipient, information, purpose and duration? Can the client choose a narrower disclosure where required? Is the relevant box affirmative? Does refusal leave the permitted route to preparation open? Does the system stop an upload until a valid signed form is recorded?
Finally, inspect one actual upload against the signed scope. If the form says “expense descriptions only” and the file contains a complete worksheet, the consent and the workflow disagree. Fix the export or seek a legally reviewed, genuinely informed permission before proceeding. Recheck whenever the provider, connected app, task or data changes. The useful output of this exercise is a controlled path from client decision to reviewed work, with a clear manual route when permission is absent.
Keep a version history for the approved form and the workflow it describes. For example, label the first approval “expense descriptions, provider A, current-year cases” and record the date and reviewer. If the firm later adds document extraction, staff should see a different version awaiting review. This simple version link prevents an old signature being treated as permission for a substantially wider service.
Questions before you ask a client to sign
Does removing a client's name always remove the need for consent?
No. A supposedly anonymous prompt can still contain tax return information or details that identify the taxpayer in context. Assess the actual material and the recipient before uploading it. A reliable route is to replace the case with a genuinely synthetic scenario and verify that no client's information remains.
Can one checkbox cover every future AI product?
A broad checkbox is a poor fit for a knowing, specific consent. For a Form 1040-series return, the consent must identify the intended recipient, information and purpose and meet the prescribed form and signature rules. A new vendor or new purpose warrants a fresh review and may require a new consent.
Can a client refuse and still have a return prepared?
For an optional disclosure unrelated to preparation, the client must be free to decline without losing preparation services. A narrower exception applies where disclosure to another preparer is needed to assist preparation; the prescribed wording differs. Have a qualified adviser review the exact workflow before linking consent to service terms.
Further reads
- How Small Tax Practices Use AI Through the Busy Season — Choose preparation tasks that keep a human reviewer in control.
- ChatGPT Prompts for Tax Preparers: Client Emails and Checklists — Use prompts for client communication without exposing returns.
- Questions to Ask Before Buying AI That Touches Client Data — Interrogate a vendor's storage and access terms.
- 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Spot confidentiality failures beyond the consent form.
- How to Mention AI Use in Client Contracts and Proposals — Separate engagement terms from specific data permissions.
- Tax Preparation AI Software: TaxDome, Canopy and ChatGPT Compared — TaxDome, Canopy and ChatGPT sit in different product categories. Compare their AI, their cost models and the consent step before return data goes in.
- TaxDome Pricing: Plans, Seasonal Seats and What to Ask on the Demo — TaxDome's new client-based price against the old seats: four practices modelled, active-account rules, the AI add-on and the questions to ask on the demo.
- TaxDome vs Karbon for a Small Firm: AI, Cost and Fit — Compare TaxDome and Karbon on list price, costs the pages leave out and AI by job, then score both with weights that fit a tax-heavy or year-round firm.
- AI Tax Return Prep for Small Practices: Extract, Review, Sign Off — A three-stage workflow for AI-assisted return prep: extract with page references, tie out with a spreadsheet, then review and sign off in writing.
- Circular 230 and AI: Due Diligence When a Tool Drafts the Work — A per-return checklist and sign-off sheet that turns the IRS's June 2026 AI alert and Circular 230 diligence rules into a routine.
- Tax Services AI for a Small Practice: Intake, Chasing and Review Notes — A three-part workflow for a small tax practice: intake checklist, document-chasing emails and reviewer notes, with a named person owning each step.
- Tax Workpaper Prep With AI: Tie the Return Back to Source Documents — How AI can draft a source-to-return tie-out schedule for a tax return, and the reviewer checklist that proves each figure before sign-off.
- TaxDome Automations: What to Automate First in a Small Practice — Which TaxDome automations to set up first in a small practice: onboarding, document requests and reminders, in order, with what to leave manual.
- WISP for Tax Preparers: Add Your AI Tools Without Gaps — One added section for your existing WISP that covers AI tools: register, vendor due diligence, data rules, MFA and incident steps, with sample wording.
- Can AI Draft an IRS Notice Response? What to Check Before Sending — A practitioner's routine for AI-drafted IRS notice responses: triage by type and deadline, redact first, verify every code, amount and date, then sign.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: IRS Section 7216 Information Center and Section 7216 FAQs; Treasury Regulations 301.7216-1 to 301.7216-3; Revenue Procedure 2013-14 (Internal Revenue Bulletin 2013-3); IRS Office of Professional Responsibility Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice (24 June 2026). Checked 4 November 2026.