AI tax preparation is sold on accuracy rates. The number that decides your exposure sits in the preparer penalty regulation instead: the greater of $1,000 or 50 percent of the income the preparer derived from the return. The same section reaches the firm itself, where the firm failed to provide reasonable and appropriate procedures for review of the position.
Neither one turns on which tool produced the position. Both turn on what a person did before the return went out. Below is the standard a model-produced position still has to clear, and the review procedure that firm-level condition is describing.
AI Tax Preparation Does Not Change What a Position Has to Clear
A position, in this context, is the treatment of an item on a return: a deduction claimed, an election made, a character assigned to income. An understatement of liability is measured against the whole return rather than against the single position. It exists, viewing the return or claim for refund as a whole, where there is an understatement of the net amount payable, or an overstatement of the net amount creditable or refundable, at section 1.6694-1(c) (eCFR, section 1.6694-1). The regulation penalizes positions, not methods, so the question it asks is never whether software was involved.
Section 1.6694-2(a)(1) says a tax return preparer is liable for a penalty under section 6694(a) "equal to the greater of $1,000 or 50 percent of the income derived (or to be derived) by the tax return preparer for any return or claim for refund that it prepares that results in an understatement of liability due to a position if the tax return preparer knew (or reasonably should have known) of the position", and one of three further conditions is met (eCFR, section 1.6694-2).
That knowledge trigger is the part a software buyer should read twice. A position sitting in a file you reviewed and signed is hard to place outside what you reasonably should have known. A tool that fills a field quietly has not put the position outside your knowledge. It has put it inside a return you are about to sign without having looked.
Which preparer carries the penalty inside a multi-preparer engagement is a separate question, settled by the primarily responsible signing preparer rule.
Substantial Authority, or Disclosure Plus a Reasonable Basis
Those three further conditions are a ladder, and the rung you land on depends on whether the position was disclosed.
For an undisclosed position that is not a tax shelter or a reportable transaction, the penalty applies where "there was not substantial authority for the position", at section 1.6694-2(a)(1)(ii). For a position that was disclosed, the penalty applies where "there was no reasonable basis for the position", at section 1.6694-2(a)(1)(iii). Tax shelter and reportable transaction positions sit on a higher rung again, where it must have been reasonable to believe the position would more likely than not be sustained on its merits (eCFR, section 1.6694-2).
Both standards have defined meanings, and neither is a feeling about the answer. Substantial authority is "an objective standard involving an analysis of the law and application of the law to relevant facts", less stringent than the more likely than not standard but more stringent than reasonable basis, and it exists "only if the weight of the authorities supporting the treatment is substantial in relation to the weight of authorities supporting contrary treatment" (eCFR, section 1.6662-4).
Reasonable basis is lower and still not low. It is "a relatively high standard of tax reporting, that is, significantly higher than not frivolous or not patently improper", and it "is not satisfied by a return position that is merely arguable or that is merely a colorable claim" (eCFR, section 1.6662-3).
The disclosure route is a filing, not an intention. Section 1.6694-2(d)(1) says the penalty will not be imposed where the position, outside the shelter and reportable transaction cases, "has a reasonable basis and is adequately disclosed". Adequate disclosure then has its own standards, and for a signing preparer one of them is disclosure on a properly completed and filed Form 8275, Disclosure Statement, or Form 8275-R, Regulation Disclosure Statement, as appropriate, or on the return under the annual revenue procedure, at section 1.6694-2(d)(3)(i) (eCFR, section 1.6694-2).
A model's suggestion is on neither rung when it arrives. It is a proposed treatment waiting for a person to find authority for it or to disclose it.
A Generated Answer Is Not on the List of Authorities
The list of what counts as authority is closed, apart from a narrow exception for written determinations, and nothing about a tool appears on the list or inside that exception. A written determination here is a ruling or determination letter issued to the taxpayer, a technical advice memorandum in which the taxpayer is named, or an affirmative statement in a revenue agent's report for a prior taxable year of that taxpayer, at section 1.6662-4(d)(3)(iv)(A) (eCFR, section 1.6662-4).
Section 1.6662-4(d)(3)(iii) states that, except in those written-determination cases, "only the following are authority", and then names them: applicable Code and other statutory provisions, proposed, temporary and final regulations construing those statutes, revenue rulings and revenue procedures, tax treaties and official explanations of them, court cases, congressional intent as reflected in committee reports and pre-enactment floor statements, the Joint Committee on Taxation's General Explanations, private letter rulings and technical advice memoranda, actions on decisions and general counsel memoranda, IRS information or press releases, and notices, announcements and other administrative pronouncements published in the Internal Revenue Bulletin (eCFR, section 1.6662-4).
The same paragraph then rules out the closest thing to a generated answer. "Conclusions reached in treatises, legal periodicals, legal opinions or opinions rendered by tax professionals are not authority." A written opinion from a qualified tax professional does not count. An output that reads like one does not count either.
It does not stop there, and the next sentence is the instruction. "The authorities underlying such expressions of opinion where applicable to the facts of a particular case, however, may give rise to substantial authority for the tax treatment of an item." An opinion points at authority. Somebody still has to go and read what it points at.
Two failure modes follow from that, and they are different problems. The first is a citation that does not exist, or that does not say what the output says it says, which is why confabulated citations get checked by opening the section rather than by rereading the summary. The second is quieter. The same paragraph adds that "an authority does not continue to be an authority to the extent it is overruled or modified, implicitly or explicitly, by a body with the power to overrule or modify the earlier authority." A case can be quoted accurately and still be dead, and nothing in the quotation says so.
The Firm Is Liable on a Paragraph of Its Own
This is the part a tool purchase actually touches, and it sits in a different paragraph from the one that penalizes the preparer.
Section 1.6694-2(a)(2) says a firm that employs a tax return preparer subject to the section 6694(a) penalty, or of which the individual preparer is a partner, member, shareholder or other equity holder, "is also subject to penalty if, and only if" one of three things is true. The second is the one to sit with. It reaches a firm that "failed to provide reasonable and appropriate procedures for review of the position for which the penalty is imposed" (eCFR, section 1.6694-2).
The other two work differently. The first does not turn on procedures at all: it reaches conduct where one or more members of the principal management, or principal officers, of the firm or a branch office participated in or knew of the conduct the section proscribes. The third describes how a firm behaves once procedures exist, and reaches a firm that "disregarded its reasonable and appropriate review procedures through willfulness, recklessness, or gross indifference (including ignoring facts that would lead a person of reasonable prudence and competence to investigate or ascertain)" (eCFR, section 1.6694-2).
Now picture a purchase that never went through a review decision. A tool arrives on one manager's card, a few preparers start using it inside the season, and no document anywhere says what review its output gets. Buying the tool is not the exposure. Where no procedure covers review of the position the tool drafted, that is the fact pattern the second condition describes.
The "if, and only if" reads both ways, which is the useful half for a firm that has done the work. A firm whose review procedures are real, and followed, sits outside the second and third conditions on the face of the paragraph. The first is a separate exposure, and it turns on whether principal management or principal officers participated in or knew of the proscribed conduct, whatever the review procedures say. The paragraph also attaches to the firm only where an individual preparer is subject to the penalty in the first place, so it is not a standalone exposure. It is a second bill on the same understatement.
How Reasonable Cause Reads When a Model Produced the Error
The explanation survives on the same terms any other explanation does, judged against the factors the regulation lists. Four of them do most of the work here.
Section 1.6694-2(e) removes the penalty where, "considering all the facts and circumstances," it is determined that the understatement was due to reasonable cause and that the preparer acted in good faith. It then lists the factors to consider (eCFR, section 1.6694-2).
The first factor covers an error a competent preparer of returns of that type reasonably could have made on a provision that was complex, uncommon or highly technical, and then closes the obvious exit. The exception "does not apply to an error that would have been apparent from a general review of the return or claim for refund by the tax return preparer." An error a reviewer would have caught by reading the return is not excused by where it came from.
The second and fourth factors are where model-produced work parts company with a human slip. Isolated errors are the ones the exception generally forgives, and it "does not apply if there is a pattern of errors on a return or claim for refund", even where any one of those errors in isolation would have qualified. The fourth factor adds that the exception does not apply where there is "a repetition of the same or similar errors on numerous returns or claims for refund." A person makes a mistake on one return. A model applying a wrong rule applies it to every return in the batch, and that is the second of those two descriptions.
The fourth factor also spells out what a defensible practice looks like, and it reads like a specification. Normal office practice "must be a system for promoting accuracy and consistency in the preparation of returns or claims for refund", and for a signing preparer it generally includes "checklists, methods for obtaining necessary information from the taxpayer, a review of the prior year's return, and review procedures."
A reliance factor sits at section 1.6694-2(e)(5), written around advice and information furnished by the taxpayer, another advisor, another tax return preparer or other party. Whether a tool's output is any of those is contestable, and the parallel question under the Treasury practice rules is worked through in what AI fixes and what an offshore team fixes. Take the most generous reading available and the factor still carries its own limits. Reliance is not in good faith where "the advice or information is unreasonable on its face", or where the preparer knew or should have known "that the advice or information was no longer reliable due to developments in the law since the time the advice was given." A model answering on last year's law is the second of those, described in advance.
Write the Review Procedure the Regulation Names
That firm-level condition does not use the word written. It asks whether the firm provided reasonable and appropriate procedures for review, and a procedure nobody wrote down is one you get to reconstruct from memory afterwards. So the deliverable is a page. Six items belong on it.
- What counts as a position here. A model-filled field that changes reported liability is a position. Reformatting, extraction into a field a person then checks, and drafting a client letter are not. Sort the tool's outputs into those two piles first, because everything below applies only to the first pile.
- Which model-touched positions get re-derived before signing. Name them by category rather than by exception: new elections, anything that was not on last year's return, and any position where the tool supplied the reasoning rather than the arithmetic. Re-derived means a person found an authority on the closed list and read it.
- What the file has to show. The authority the person found, cited to its own source, the date it was checked, and who checked it. The model's own citation string is not the record. It is the thing that was checked.
- Who decides when the answer is disclosure. Where a person cannot reach substantial authority but the position has a reasonable basis, the Form 8275 route is a decision with a named owner and a deadline, not a fallback nobody ever triggers.
- The batch check no single return produces. Once a month in season, pull a sample across returns and read them for the same wrong treatment rather than for individual errors. A per-return review cannot surface a repeated error by design, because each return looks defensible on its own terms. The ledger-side version of the same sampling problem, where a coding rule goes wrong across many lines at once, sits in the exception standard for AI bookkeeping.
- Who owns the note, and when it gets re-read. A procedure nobody owns is the one most likely to go unfollowed, and the third condition reaches disregard that rises to willfulness, recklessness, or gross indifference. Date it, name the partner who owns it, and re-read it when the vendor ships a model change.
That page is also the difference between a review that adds capacity and one that does not. Where review is already the bottleneck, more drafted positions arriving faster makes the queue longer, which is the same sequencing problem the automation handoffs hit at the transmission gate.
When Not to Point a Model at Positions
Some work fails this test before it starts, and saying so early is cheaper than finding out in March.
Where a position sits in an area nobody in the firm knows cold, re-deriving the authority costs more than preparing the position from scratch, and the tool has moved the work rather than removed it. Where the volume is genuinely small, the batch check has nothing worth sampling and the whole procedure is overhead. And where nobody has the hours to work the queue the tool creates, the constraint was review capacity all along, which no model output changes.
Questions Firms Ask
Can AI Prepare a Tax Return?
It can produce the fields and a draft treatment. What it cannot supply is what the position itself needs, which is substantial authority, or disclosure plus a reasonable basis, found and read by somebody. It also cannot be the preparer of record, for the separate reason that the identifying number a filed return carries belongs to a person.
Who Is Liable if an AI Tool Causes an Understatement?
The preparer, and separately the firm. The penalty runs to the preparer who knew or reasonably should have known of the position, and section 1.6694-2(a)(2) adds the firm where it failed to provide reasonable and appropriate procedures for review of the position (eCFR, section 1.6694-2). The vendor is not named in that paragraph.
Will AI Replace Tax Preparers?
The penalty regulation assumes somebody is there to carry it, which is a legal answer rather than a labor-market one. What the official employment projections say is covered in AI versus offshoring. The professional standards point the same way, since the tax services standards define artificial intelligence as a tool and leave the responsibility with the member who uses it.
Start With the Review, Not the Tool
What a firm buys when it buys AI tax preparation is more drafted positions per hour. What it still owes on every one of them is an authority somebody read and a review somebody performed.
If review is where your season actually breaks, that is a capacity problem, and it is worth testing rather than describing. Accountably places trained offshore accountants and tax preparers inside US CPA and EA firms, and the signature, the position and the final judgment stay with your firm. Don't trust us. Test us. The Free 40-Hour Proof Pilot puts a fixed block of your own work through the full review chain, so your reviewer grades real output before a single client file is committed. See how the pilot works.