Can AI write a tax research memo?

It can draft the prose, frame the issue, and assemble the analysis. What it should not do is supply the authorities from its own memory, because that is where fabricated section numbers come from. Keep the model as the writer and use a retrieval step against primary source for every citation.

What structure should a tax research memo follow?

Facts, issue, analysis against authority, and conclusion. The facts and the conclusion stay with you. The model can draft the issue framing and the analysis, as long as the authorities it works from are retrieved rather than generated.

How do I stop AI from inventing citations in a memo?

Give the model the retrieved authorities as fixed inputs and instruct it to use only those, introducing no citation that is not in the list. Then open every cite against primary source before the memo leaves your desk. A citation you did not personally open is one you are taking on faith.

Is it safe to rely on an AI-drafted memo?

Only after a verification pass: open every link, match each quote to the source text, confirm the authority governs the tax year in the facts, and ask which provision cuts against your conclusion. The model drafts toward the answer you implied, so the provision that hurts is the one it is most likely to omit.

How to Draft a Tax Research Memo with AI (With Real Citations)

The fastest way to ruin a good research memo is to let the tool that writes the prose also decide what the law says. Those are two different jobs with two different failure modes, and the whole trick to drafting a memo with AI is keeping them apart. The model is a strong writer and a weak source of record. Treat it accordingly and you get a memo in a fraction of the time without inheriting the risk. (If you want the mechanics of why models fabricate citations in the first place, that is its own piece.)

The memo has a structure. Use it as the seam.

A tax research memo is the same four moves every time: the facts, the issue, the analysis against authority, and the conclusion. That structure is not just convention. It is the natural seam between what AI should do and what it must not.

Layer	Who owns it	Why
Facts	You	The model was not in the room. Garbage facts produce a fluent, wrong memo.
Issue framing	AI drafts, you approve	Models are good at sharpening a vague question into a precise legal one.
Authority	Retrieval, never the model’s memory	Section numbers and quoted text must be copied from the real Code, not generated.
Analysis	AI drafts, you judge	The model assembles the argument; you decide whether it actually applies.
Conclusion	You	Your name signs it.

The single rule that holds the whole thing together: the model writes the analysis, but every authority in it is retrieved from primary source, with a URL you can open. Where that is true, the memo is fast and defensible. Where it is not, you are back to verifying a stranger’s citations under deadline.

A worked example: the home-office deduction

Take a routine question. A sole proprietor converts a spare bedroom into the only place she does her books, invoicing, and client scheduling. She uses it for nothing else. Can she deduct it?

Here is how the four layers play out.

Facts

Write these yourself, tightly. The model should never invent or soften a fact.

Client is a cash-basis sole proprietor. She uses one bedroom of her home regularly and exclusively for the administrative and management work of the business. She conducts no substantial administrative work at any other fixed location. The room is not used for any personal purpose.

Every operative word there (“regularly,” “exclusively,” “no other fixed location,” “no personal purpose”) is doing legal work. That is exactly why you, not the model, control it.

Issue

Ask the model to frame it. A good prompt:

Frame the precise legal issue for a tax research memo, given these facts. Do not answer it yet. Do not cite anything.

You will get back something like: Whether the taxpayer may deduct expenses allocable to the bedroom under the business-use-of-home rules, given regular and exclusive use as the principal place of business. That is clean. Approve it and move on.

Authority

This is the layer that has to be real. Pulled from primary source, the governing provision is §280A, and the relevant mechanics are specific:

§280A(a) lays down the general rule: no deduction for the business use of a dwelling unit the taxpayer uses as a residence. That is the default, and it disallows everything until an exception applies.
§280A(c)(1) is the exception that rescues our facts. It restores the deduction for the portion of the home used exclusively and on a regular basis as the principal place of business. Both adjectives are load-bearing. A room used exclusively but only occasionally fails. A room used daily but also as a guest room fails the exclusivity test.
§280A(c)(5) is the ceiling people forget. Even when the exception applies, the deduction cannot exceed the gross income from the activity, reduced by the otherwise-allowable deductions (mortgage interest, taxes) and the business deductions not allocable to the home. In a low-revenue year, this cap is the real answer, and a memo that omits it is incomplete.

Notice what the model did here: nothing. The section numbers, the structure, and the quoted thresholds came from the Code. The model is about to write around these, not generate them.

Analysis and conclusion

Now you hand the approved facts and the retrieved authority to the model and let it draft:

Using only the facts and the authorities below, draft the analysis section of a tax research memo. Apply each authority to the facts. Do not introduce any citation that is not in the list I gave you. Flag any fact you would need to confirm.

Because the authorities are fixed inputs, the model cannot wander off into an invented §280A(c)(7). It will apply (c)(1) to the exclusive-and-regular-use facts, reach a favorable conclusion on eligibility, and (if you gave it (c)(5)) note that the deduction is capped at the activity’s income. The conclusion is yours to set, but the draft will be most of the way there.

The verification pass, before it leaves your desk

Grounded retrieval removes the fabrication risk on the authorities. It does not remove your judgment. Run this pass on every memo:

Open every link. Each cite should resolve to the section it claims, on a primary source. A cite you did not personally open is a cite you are taking on faith.
Match the quote to the text. If the memo quotes statutory language, confirm it reads that way at the source, word for word.
Check the clock. Authority has effective dates and sunset provisions. Confirm the version you are citing governs the tax year in the facts.
Look for the provision that hurts. Models draft toward the conclusion you implied. Ask directly: what authority cuts the other way? In the home-office memo, §280A(c)(5) is that provision, and a model steered toward “yes, deductible” will happily skip it unless you make it look.
Confirm the facts are the client’s facts. Reread the facts section against the file. The most dangerous error in an AI-drafted memo is not a bad cite. It is a clean memo answering a question the client did not ask.

Do this and the time saved is real and the exposure is not transferred to you. The model drafted in minutes what would have taken an hour to write. You spent that hour where it belongs: on the facts, the judgment, and the provision that bites.

That is the shape of the work as drafting gets cheap. The memo was never the job; it was the artifact. The job is knowing which facts carry weight, which authority governs, and which provision the model was too agreeable to raise. Hand off the typing. Keep the judgment, because that is the part with your name on it.

This walkthrough is illustrative and not tax advice. Application of §280A turns on specific facts and current guidance; confirm against primary source for any live engagement.