AI tools are now in nearly every part of a tax practice. They summarize client emails, pull numbers from source documents, draft engagement letters, and answer research questions in seconds. Many preparers will use them more this filing season than ever before.
The speed is real. So is the risk. Federal law has controlled how preparers handle client tax information for decades, and those rules don’t make an exception for software that showed up last year. In June 2026, the IRS Office of Professional Responsibility (OPR) issued Alert 2026-19, “Introductory Guidelines for Responsible AI Use in Federal Tax Practice.” It does not create new rules. It makes clear that the rules you already follow apply in full when AI is involved.
This article covers the two Internal Revenue Code sections every preparer should think about before uploading client data to an AI tool, the penalties involved, and practical ways to get the benefits of AI without putting your clients or your practice at risk.
The Two Statutes That Matter Most
Section 7216: the criminal provision
Section 7216 applies to anyone in the business of preparing returns, anyone providing services in connection with return preparation, and anyone paid to prepare a return for someone else. It is a misdemeanor for that person to knowingly or recklessly:
disclose any information furnished to them for, or in connection with, preparing a return, or
use that information for any purpose other than preparing, or helping to prepare, the return.
The key word for AI users is “recklessly.” You don’t have to intend harm. Uploading a client’s documents to a public chatbot without understanding where the data goes, who can see it, or whether it will be used to train the model is exactly the kind of conduct a reckless standard is meant to reach.
Section 6713: the civil provision
Section 6713 covers the same people and the same conduct: disclosing return information, or using it for anything other than return preparation. The difference is that Section 6713 has no knowledge or recklessness requirement. The IRS’s own Section 7216 FAQ says plainly that the civil penalty does not require knowing or reckless conduct. An honest mistake can still trigger it.
Why AI fits squarely inside these rules
The Treasury regulations define the key terms broadly:
Tax return information means any information, including a name, address, or identifying number, furnished in any form or manner for or in connection with preparing a return. It includes information from third parties and information the preparer derives or generates from it.
Disclosure means making tax return information known to any person in any manner whatever. Sending it to an AI vendor’s servers is a strong candidate.
Use includes any time a preparer refers to or relies on return information as the basis to take or permit an action.
One detail surprises many practitioners. The regulations say tax return information includes statistical compilations “even in a form that cannot be associated with, or otherwise identify, directly or indirectly, a particular taxpayer.” Stripping names helps a great deal, but it does not by itself take the information outside the statute. There is a narrow exception for certain anonymous statistical compilations, but there is no general “de-identified data” safe harbor.
Exceptions don’t clearly cover public AI tools
Preparers sometimes assume the existing exceptions cover AI. Be careful with that assumption.
Preparer-to-preparer exception. You may share information with another U.S.-based preparer for return preparation or auxiliary services, but only if the services don’t involve substantive determinations or advice affecting the tax liability. Asking an AI tool to analyze a client’s facts and tell you how to treat an item is a substantive determination.
Software contractor exception. You may share information with a contractor for programming, maintenance, repair, testing, or procurement of software used for return preparation. It applies only to the extent necessary for those services, and only if every individual who will receive the information gets written notice of Sections 6713 and 7216. A free consumer chatbot account doesn’t come with that arrangement.
Offshore processing. Disclosing information to anyone outside the United States requires prior written consent. A U.S. preparer generally cannot get consent to send a Form 1040 filer’s Social Security number overseas at all; it must be redacted or masked. The regulations’ own example treats a foreign contractor employee viewing data on a U.S. server as a disclosure outside the United States. If you don’t know where your AI vendor processes data, you can’t know whether you’re complying.
Whether a particular AI product qualifies for an exception is still unsettled. That uncertainty is why the safest path, when real client data is involved, is written client consent plus a vetted enterprise tool.
What valid consent looks like
If you plan to disclose return information to an AI vendor, or use it in a way not covered by an exception, you need written consent before the disclosure or use. The consent must:
name the preparer and the taxpayer;
describe the specific information to be disclosed or used;
state the purpose and, generally, the specific recipient;
be signed and dated by the taxpayer (pen-and-ink or electronic signatures are allowed);
include the mandatory language in Rev. Proc. 2013-14.
A few rules catch people off guard:
Consent to disclose and consent to use must be in separate documents.
Consent must be knowing and voluntary. With a narrow exception, you can’t make return preparation conditional on the client agreeing to the AI disclosure. If a client says no, you need a non-AI workflow for that client.
Retroactive consent is not allowed.
If no expiration is stated, the consent lasts one year from signing.
You must give the client a copy of the signed consent.
The Penalties
IRC Section 6713(a): civil penalty
Who: Preparers and people providing preparation-related services
Standard: No knowledge or intent required
Penalty: $250 per disclosure or use, capped at $10,000 per calendar year
IRC Section 6713(b): civil penalty, identity theft
Who: Same, when the disclosure or use is connected to a crime involving misappropriation of a taxpayer’s identity
Penalty: $1,000 per disclosure or use, capped at $50,000 per year, with this cap applied separately from the regular cap
Note: Deficiency procedures don’t apply to Section 6713 penalties, so they can be assessed without the usual prepayment Tax Court route
IRC Section 7216(a): criminal penalty
Who: Same
Standard: Knowingly or recklessly
Penalty: Misdemeanor. Fine up to $1,000 ($100,000 if the identity theft provision applies), up to 1 year in prison, or both, plus costs of prosecution
Circular 230 Sections 10.51(a)(15) and 10.50: professional discipline
Who: Practitioners and their firms
Standard: Willful unauthorized disclosure or use
Penalty: Censure (a public reprimand), suspension, or disbarment from practice before the IRS; a monetary penalty up to the gross income from the conduct; and a separate penalty on the firm if it knew or reasonably should have known
Circular 230 Section 10.36: supervisory responsibility
Who: Anyone with principal authority over a firm’s tax practice
Standard: Willfulness, recklessness, or gross incompetence in failing to maintain and enforce adequate procedures
Penalty: Discipline for failing to supervise, which OPR now expressly ties to AI policies
FTC Safeguards Rule: breach reporting
Who: Tax preparation firms, which the FTC treats as financial institutions
Trigger: A breach involving unencrypted customer information of 500 or more consumers
Requirement: Report to the FTC as soon as possible and no later than 30 days after discovery, plus any state breach-notice duties
The per-disclosure math adds up quickly. A preparer who uploads source documents for 40 clients to an unapproved tool has potentially made 40 separate disclosures. At $250 each, that reaches the $10,000 annual cap, before any criminal exposure, Circular 230 discipline, breach response costs, malpractice claims, or damage to the practice’s reputation.
What OPR Expects From You Now
OPR Alert 2026-19 ties AI use to six existing Circular 230 obligations. The parts most relevant to data handling are direct:
“Practitioners must strictly handle all client data using only secure, enterprise-approved AI.”
“Never upload sensitive data to unsecured sites.”
“Willful mishandling of taxpayer information through AI may lead to disciplinary actions under Circular 230.”
The alert also warns about a risk unique to generative AI: client information can be “repurposed by the program to respond to an inquiry concerning another client,” or “spilled over into an algorithm and combined with a related tax issue involving a different client.”
Beyond confidentiality, OPR reminds practitioners that:
Due diligence (Section 10.22) means verifying every fact, citation, and calculation an AI tool produces before it goes to a client or the IRS.
Written advice (Section 10.37) can’t rest on AI output you haven’t verified. If the system’s logic is opaque, relying on it may be unreasonable.
Fees (Section 10.27) should reflect efficiencies from AI. Billing for time not actually spent may be a problem.
Firm procedures (Section 10.36) must cover AI training, secure data handling, accuracy monitoring, and vetting of third-party AI tools, and the firm must document all of it.
Your Written Information Security Plan (WISP), which federal law already requires, is the natural place to record your AI policy.
Safe Ways to Use AI This Season
None of this means you should avoid AI. It means you should match the tool to the task and keep client data where it belongs.
1. Use AI for tax research, not for client files
General research is the lowest-risk, highest-value use of AI in a tax practice. Ask about the law, not about your client.
Good: “What are the requirements for the qualified business income deduction for a specified service trade or business, and where in the regulations are they?”
Risky: Uploading a client’s K-1 and asking whether they qualify.
Guardrails for research:
Frame questions as hypotheticals with generic facts. “A single-member LLC taxed as a sole proprietorship with $85,000 of net income” is fine. A client’s name, location, employer, and exact figures are not.
Verify every answer against primary authority: the Code, Treasury regulations, IRS publications, revenue rulings, and case law. AI tools still invent citations. OPR’s instruction is plain: “always check citations.”
Prefer research platforms that ground AI answers in licensed tax content and link to the underlying source.
Keep a short record of what you asked, what the tool said, and how you verified it. That record is your due diligence and supervision evidence.
2. Remove PII before anything leaves your environment
If you need AI help with a client-specific fact pattern and you have not obtained consent and vetted the tool, strip the data first. Treat this as risk reduction, not a legal safe harbor, because the regulations treat even de-identified compilations as tax return information.
Remove or replace:
Names of the taxpayer, spouse, dependents, employers, payers, and businesses
SSNs, ITINs, EINs, PTINs, and IP PINs
Street addresses, cities, ZIP codes, and property descriptions
Dates of birth, death, marriage, or other life events
Phone numbers, email addresses, and usernames
Bank, brokerage, and loan account numbers, plus routing numbers
Policy numbers, case numbers, and IRS notice numbers
Unusual exact amounts or dates that could re-identify someone in a small community
A practical workflow:
Never upload source documents. W-2s, 1099s, K-1s, prior returns, and IRS notices are packed with identifiers. Write a short summary of the issue instead.
Use placeholders such as “Taxpayer A,” “Business 1,” and “State X.”
Round numbers when precision doesn’t change the answer.
Generalize unique facts. “A rural Western South Dakota cattle operation with a family-owned feedlot” can identify a client. “An agricultural business” usually can’t.
Check file layers before sharing anything. Drawing a black box over text in a PDF often leaves the text underneath readable. Use true redaction tools, flatten the file, and remove metadata such as author, file path, and document properties.
Watch screenshots. Browser tabs, client portal headers, file names, and taskbar previews often reveal names.
Turn off chat history and model training in any tool you use, and read the provider’s data retention terms.
Have a second person spot-check redactions during busy season, when mistakes are most likely.
3. Run AI locally when you need to work with real client data
A local large language model runs entirely on hardware your firm controls. Nothing goes to an outside vendor, which avoids the third-party disclosure problem that makes cloud chatbots risky. Popular options include Ollama, LM Studio, and llama.cpp, which run open-weight models such as Llama, Qwen, Mistral, and Gemma on a workstation or in-office server.
Local doesn’t automatically mean compliant. To do it right:
Confirm it’s actually offline. Disable telemetry, cloud sync, automatic uploads, and any plugins or “web search” features before client data touches the tool. Test with the network disconnected.
Treat the machine like your tax software server. Use full-disk encryption, multifactor authentication, user access controls, backups, and patching, consistent with IRS Publication 4557 and your WISP.
Keep client data segregated. Don’t let one client’s documents sit in a shared context or database that another client’s session can draw from. This addresses the cross-client “spill over” risk OPR described.
Size the hardware realistically. Useful models generally need a machine with substantial memory and, ideally, a capable GPU. Smaller models are faster but make more mistakes.
Verify output just as carefully. Local models are often less capable than leading cloud models. Running locally solves the privacy issue, not the accuracy issue.
Document it. Add the tool, its configuration, who may use it, and how output is reviewed to your WISP and AI policy.
Remember that local use is still “use” of tax return information. It is permitted when it’s for preparing the return. Using the same data for marketing, cross-selling, or building products for other clients raises separate Section 7216 issues.
4. If you use a cloud AI vendor with client data, do it on paper first
Some firms will decide an enterprise cloud tool is worth it. If so:
Use a business or enterprise agreement, not a consumer account.
Get written commitments that client data won’t be used to train or fine-tune models, including by the vendor’s underlying model provider.
Confirm retention limits, encryption in transit and at rest, access logging, and U.S. data processing and storage.
Ask for independent security attestations such as a SOC 2 report.
Provide the written Section 6713 and 7216 notice the regulations require for contractors.
Obtain Section 7216 consent from each client that names the vendor, describes the information and purpose, uses the Rev. Proc. 2013-14 language, and is signed before any upload. Keep a non-AI process ready for clients who decline.
Check AI features already built into your tax, practice management, and document software. Many send data to outside model providers by default. “It’s inside our tax software” is not the same as “it’s covered.”
A Short Checklist for This Filing Season
Write or update a firm AI policy and add it to your WISP.
List every AI tool in use, including features embedded in existing software, and approve or block each one.
Ban uploads of client source documents to consumer AI tools.
Train staff on PII removal and verification before the season starts.
Update engagement letters and prepare separate Section 7216 consent forms if any approved tool receives client data.
Set up a local model for client-data tasks if the workload justifies it.
Require human review and citation checks on all AI-assisted work.
Review billing practices so they reflect AI efficiencies.
Confirm your breach response plan covers an AI-related exposure, including the FTC 30-day reporting rule.
The Bottom Line
AI can make a tax practice faster and more consistent. But your clients hand you their most sensitive financial information on the understanding that it goes to their return and nowhere else. Sections 6713 and 7216 put that understanding into law, and OPR has now confirmed it applies to every AI tool you use.
Use AI freely for research and general drafting. Keep client data out of tools you haven’t vetted. Strip identifiers when you must ask about a real situation, get written consent when data will leave your control, and consider running models locally for anything involving real client files. Above all, review everything. The signature on the return is still yours.
This article is general information for tax professionals and is not legal advice. Consult the statutes, regulations, and your own counsel regarding your specific practice.
Sources
· 26 U.S. Code § 6713, Disclosure or use of information by preparers of returns: https://www.law.cornell.edu/uscode/text/26/6713
· 26 U.S. Code § 7216, Disclosure or use of information by preparers of returns: https://www.law.cornell.edu/uscode/text/26/7216
· 26 CFR § 301.7216-1, Definitions: https://www.law.cornell.edu/cfr/text/26/301.7216-1
· 26 CFR § 301.7216-2, Permissible disclosures or uses without consent: https://www.law.cornell.edu/cfr/text/26/301.7216-2
· 26 CFR § 301.7216-3, Disclosure or use permitted only with consent: https://www.law.cornell.edu/cfr/text/26/301.7216-3
· IRS, Section 7216 Frequently Asked Questions: https://www.irs.gov/tax-professionals/section-7216-frequently-asked-questions
· IRS OPR Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026): https://www.irs.gov/pub/opr-taxpros/issue-number-2026-19-introductory-guidelines-for-responsible-ai-use-in-federal-tax-practice.pdf
· 31 CFR § 10.50, Circular 230 Sanctions: https://www.law.cornell.edu/cfr/text/31/10.50
· IRS Publication 4557, Safeguarding Taxpayer Data: https://www.irs.gov/pub/irs-pdf/p4557.pdf
· IRS, Written Information Security Plans are essential for tax pros: https://www.irs.gov/newsroom/written-information-security-plans-are-essential-for-tax-pros
· FTC, Safeguards Rule: What Your Business Needs to Know: https://www.ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-business-needs-know
· FTC, Safeguards Rule notification requirement now in effect: https://www.ftc.gov/business-guidance/blog/2024/05/safeguards-rule-notification-requirement-now-effect

