Tax Practice, and Why a Model's Answer Is Not Authority

Tax professional describes at least four distinct activities, and lumping them together is why generic advice about AI in tax is unhelpful.

Compliance. Preparing and filing returns. High volume, deadline-driven, heavily software-mediated already, and the part most people picture. The work is largely data assembly, classification and form population.

Advisory. Answering what the treatment of a transaction is, or structuring one before it happens. Lower volume, higher value, and it turns on research and judgement rather than data handling.

Controversy. Dealing with an examination, an assessment or an appeal. Adversarial, procedural, and driven by evidence and argument.

Planning. Working out, across years and entities, how a client's affairs should be arranged. The most valuable and the least structured.

Where the tooling lands is uneven. Compliance was automated long before generative models and gains mainly at the edges: extracting figures from source documents, flagging inconsistencies, drafting explanations of a return to a client.

Advisory is where generation genuinely changes the day, because research and memo production are its core activities and both are text work.

Controversy gains in correspondence and document review. Planning gains least, because the input is a client's whole situation and the output is judgement.

So this cursus concentrates on advisory, which is where the change is real and where the professional standards bite hardest.

The feature that distinguishes tax from almost every other professional domain is that it has a formal, enumerated hierarchy of what counts as a source, and getting a position wrong has a defined penalty consequence.

At the top sits the statute itself, in the United States the Internal Revenue Code, together with other statutory provisions. Below it, Treasury regulations construing those statutes, in proposed, temporary and final forms. Then administrative guidance: revenue rulings and revenue procedures. Then court decisions. Then tax treaties and official explanations of them.

There is also a body of material that carries weight with important limits: legislative history including committee reports and floor statements by a bill's managers, the Joint Committee on Taxation's General Explanations known as the Blue Book, private letter rulings and technical advice memoranda issued after 31 October 1976, actions on decisions, and general counsel memoranda.

This is not a convention among practitioners. It is set out in Treasury Regulation section 1.6662-4(d)(3)(iii), which enumerates what may be treated as authority for the purpose of determining whether there is substantial authority for a position.

The practical consequence is that a tax answer is not a fact to be looked up. It is a position supported by identified authorities, weighed against the authorities that cut the other way, since the regulation requires authorities both for and against to be taken into account.

That structure is what the next step turns on.

The same regulation that lists what counts as authority also says what does not, and the exclusion is unusually direct.

Under Treasury Regulation section 1.6662-4(d)(3)(iii), conclusions reached in treatises, legal periodicals, legal opinions or opinions rendered by tax professionals are not authority. The regulation adds that the authorities underlying such expressions of opinion may provide a reasonable basis for a return position.

Read that carefully, because it settles the question that most discussion of AI in professional services leaves open.

The considered written opinion of an experienced tax professional is already not authority. Neither is a leading treatise. What has weight is what those opinions cite.

So a generated answer is not authority, and this is not a special rule about AI. It falls on the wrong side of a line that was drawn decades earlier and that already excluded the most respected sources in the field. Nothing about model quality changes it, because the exclusion is not about reliability. It is about the difference between a source of law and a commentary on it.

Which yields the operative discipline for the whole cursus. A model's output is a lead, not a conclusion. Its value is pointing you toward the code section, the regulation, the ruling or the case, which you then read and cite.

An advisor who cites the model has cited nothing. An advisor who used the model to find the revenue ruling, read it, and cited it has done ordinary tax research faster.

The research path, with the model placed where it belongs.

A client question arrives. The model helps at the first step: framing the issue, identifying which code sections and doctrines are likely in play, and suggesting search terms. This is genuine value, because knowing where to look is a real skill and the model has broad coverage.

The output of that step is a set of leads, and the diagram marks it explicitly as not authority.

Each lead then goes to the primary source. Read the code section. Read the regulation. Read the ruling or the case in full, not a summary of it.

What survives that reading becomes the authority set, and the regulation requires weighing the authorities against the position as well as for it.

That weighted set supports a position at a stated confidence level, which determines whether disclosure is required, and the memo records the reasoning.

The shape to notice is that the model sits entirely on the left of the diagram, before any authority is touched. It accelerates finding and it never appears in the citation.

Tax has something most professions lack: a graduated vocabulary for how sure you are, with each level attached to a consequence.

Working up from the bottom, as commonly understood in practice. Not frivolous is the floor. Reasonable basis sits above it and is generally understood as roughly a one in five chance of being sustained. Substantial authority is higher, commonly described as around a forty percent likelihood. More likely than not means greater than fifty percent. Should and will indicate progressively higher confidence.

These are practitioner conventions expressed as approximate probabilities rather than figures stated in the statute, and different authors put slightly different numbers on them. The ordering, though, is settled and load-bearing.

Why it matters operationally. Under section 6694 of the Internal Revenue Code, a return preparer faces a penalty for an understatement due to an unreasonable position where the position was not disclosed and there was no substantial authority for it, or where it was disclosed and lacked even a reasonable basis. For positions other than tax shelters and reportable transactions, that penalty is the greater of one thousand dollars or fifty percent of the income derived by the preparer.

So confidence is not a feeling. It determines whether a position can be taken without disclosure, whether disclosure on Form 8275 is required, and whether the preparer is exposed.

And this is precisely what a model cannot supply. It can present the authorities. It cannot tell you they amount to forty percent, because that judgement is what the profession exists to make.

Every professional field warns about invented citations. Tax has a structural feature that makes the problem sharper.

Tax authority is highly patterned. Code sections follow a numbering scheme. Revenue rulings are numbered by year and sequence, as in Rev. Rul. 2019-11. Regulations mirror the code sections they construe. Cases follow standard reporter formats.

A model that has learned those patterns can generate a citation that is perfectly well-formed and refers to nothing. Rev. Rul. 2017-42 looks exactly like a real revenue ruling. Section 1.263(a)-4(f)(5) looks exactly like a real regulation subsection.

And the reader cannot tell. In a domain where sources have distinctive titles, a fabricated one often looks slightly wrong. In tax, where the source is a number, there is nothing to look wrong.

A second failure is more insidious than fabrication. A real, correctly cited authority that says something other than what the summary claims, or that has been superseded. Tax authority changes constantly: rulings are obsoleted, regulations are amended, and legislation revises code sections while leaving the number intact. A model's recollection of a section reflects whatever it absorbed, and a section number that once meant one thing may now mean another.

The discipline is therefore not spot-checking. It is that every authority is pulled and read at source before it enters a memo, without exception, because the two failure modes are both invisible from the output.

Return preparation deserves its own treatment, because the honest answer is that generative tools change it less than the marketing suggests.

What was already automated. Data import from financial institutions, form population, arithmetic, diagnostics that flag inconsistencies, and electronic filing. Tax preparation software has done this for decades, and it is deterministic software rather than a model, which is the correct design for arithmetic.

What generation adds at the edges, genuinely.

Extracting figures from unstructured source documents. A client arrives with a folder of statements, letters and receipts in a dozen formats, and turning that into structured data was manual. Document extraction handles it, with the output checked against totals.

Explaining a return to the client. The single most common client question is why the number changed, and answering it well takes writing time that nobody bills for.

Drafting the responses to routine correspondence.

And first-pass review, in the specific sense of flagging things worth a human look: a figure that moved sharply year over year, a classification that looks inconsistent with prior years, a missing schedule.

What it should not do. Compute anything. The arithmetic belongs in the deterministic software that was built for it, and a language model producing a tax figure is the worst available arrangement, because the output is confident and unverifiable by inspection.

An order that follows the authority discipline rather than the apparent capability.

First, issue framing. Given a client fact pattern, what areas of tax law are in play, what doctrines might apply, what should be searched. This is where the model is strongest and where nothing it produces gets cited, so the risk is close to zero and the value is real: it surfaces the issue you had not considered.

Second, document handling. Extracting structured data from a client's unstructured pile, and summarising a long agreement to find the provisions with tax consequences. Verified against the document, which you then read for the parts that matter.

Third, client explanation. Turning a position you have already reached into something a client understands. Reviewed before it goes.

Fourth, memo drafting, once you have the authorities in hand and have read them. The model structures and writes; the analysis and the citations are yours.

And not at all: producing a conclusion you then rely on, generating citations you do not pull, computing anything a client will act on, or answering a client's question directly from model output.

The distinguishing principle throughout. Use it before you touch the authorities, to find them, and after you have read them, to write. Never in the middle, where the authority is weighed, because that middle is the profession.