Where AI Actually Fits in the Tax Workflow (and Where It'll Get You in Trouble)

AI is already good at the grunt work in tax, and that's exactly the part that used to eat your week - the document chase, the data entry, the first-pass classification, the rebuilding of the same workpapers every cycle. That's real, and the adoption numbers are finally catching up to it: Thomson Reuters Institute found generative-AI use among tax and accounting professionals roughly doubled in a year, from about 12% in 2024 to 22% in 2025, and the share who think they SHOULD be using it on daily work jumped from 52% to 71% (Thomson Reuters). So most of the profession now agrees the tool belongs in the workflow, and far fewer have actually wired it in, which is the gap worth closing. And the way you close it is by being precise about where AI earns its spot and where it quietly creates risk, because in tax the cost of a confident wrong answer is a lot higher than in most accounting work.

Where AI is genuinely good in tax right now

Start with the document and data layer, because that's where the wins are obvious and the downside is small. Pulling figures off a stack of K-1s, 1099s, brokerage statements, and fixed-asset schedules, then dropping them into a structured workpaper, is the kind of tedious matching that AI handles well and that nobody on your team enjoys doing by hand. Same with first-pass document organization - sorting a client's shoebox of PDFs into the right buckets, flagging what's missing, drafting the request list for the stuff that isn't there. None of that touches a judgment call, all of it is checkable, and it's the part of the job that made the front of every busy season such a damn slog.

The other strong fit is drafting and summarizing. AI is good at turning your notes into a first-draft client memo, summarizing a long engagement thread so you remember what you promised in March, or restating a position in plain language for a client who doesn't speak tax. It's the same move we mapped for the back office in accounts payable is the easiest place to put AI to work - you put the model on the repeatable prep and keep the human on the call.

Where it'll get you in trouble

The danger zone is anywhere the answer depends on the law, and that's most of what makes tax actually tax. Generic models confidently cite code sections that don't say what they claim, mix up tax years, and hallucinate authority that sounds exactly right, and in a field where a wrong position carries penalties and your signature, a plausible-sounding wrong answer is worse than no answer. So treat anything that smells like research or a position as a draft to verify against primary source, never as the conclusion. If a model hands you a citation, you pull the actual code section, reg, or ruling and read it yourself - no exceptions, because the one time you don't is the time it invented it.

The provision is its own minefield. AI can absolutely help you assemble and reconcile the supporting schedules for a tax provision, but the judgment - the valuation allowance call, the uncertain-position assessment, the rate reconciliation that has to tie and make sense - stays with a person who can defend it to an auditor. Use the model to do the assembly and the tie-outs fast, then spend the time you saved on the judgment instead of the gathering.

The pattern that actually works

The teams getting real value aren't the ones asking a chatbot to "do the return." They're the ones who broke the workflow into pieces and handed AI only the pieces that are repeatable and checkable - intake, extraction, organization, drafting, summarizing - and kept the judgment, the research conclusions, and the sign-off firmly human. It's the same logic behind every AI workflow that holds up under pressure, and we walked through it for forecasting in FP&A's real bottleneck is the data gathering: put the model on the gathering, keep the person on the call.

One more thing that separates the teams doing this well - they care a lot about where the tool runs and what it does with client data, because tax data is some of the most sensitive a client will ever hand you. Purpose-built tax software with AI baked in and real data controls is a very different risk profile than pasting a client's return into a consumer chatbot, and the difference matters before a single efficiency gain shows up. If you're sorting through what's actually built for this versus what's a thin wrapper, our Audit Friendly software directory scores tools on features, support, value, and security so you're not guessing.

What I'd do

Pick one repeatable, low-judgment slice of your busy season and put AI on just that - I'd start with document intake and data extraction, because the pain is high, the work is checkable, and you'll feel the time back almost immediately. Build the habit of verifying any output that touches the law against primary source, every time, until it's reflex. Keep the provision judgment and the final review human and loud about it. Then, once that one slice is working and you trust it, add the next one. You don't win this by going all in on a tool you don't trust yet; you win it by handing over the drudgery a piece at a time and earning the trust as you go. If your tax work is increasingly remote, by the way, the demand is real - you can see what's actually open on the remote tax jobs board.

Frequently asked questions

Can AI prepare a tax return on its own?

No, and you shouldn't want it to. AI is good at the prep underneath a return - extraction, organization, draft workpapers, tie-outs - but the positions, the judgment calls, and the sign-off have to stay with a person who can defend them. Treat it as a fast assistant, not a preparer.

Is it safe to put client tax data into an AI tool?

It depends entirely on the tool. Purpose-built tax software with real data controls is a different world than a consumer chatbot, where you have little say over how the data is handled. Check the security and data-handling terms before any client information goes near it.

What's the single best place to start with AI in tax?

Document intake and data extraction. The work is high-volume, painful, and easy to verify, so you get an obvious time win with low downside, and it builds the habit before you point AI at anything that requires judgment.

Will AI replace tax accountants?

It's reshaping the work more than removing it - the gathering and prep compress, and the judgment, advisory, and review become a bigger share of the job. The accountants who lean into the judgment work and let AI handle the drudgery come out ahead.

How do I check an AI's tax research?

Pull the primary source it cites - the actual code section, regulation, or ruling - and read it yourself. If it can't produce a real, checkable citation, treat the answer as unverified. Models hallucinate authority that sounds right, so the verification isn't optional.

The move here isn't to wait for the tools to get perfect, it's to hand over the drudgery you already hate one slice at a time and put the hours you get back into the judgment work that's actually why clients pay you.