Most accounting firms have now tried some form of AI. Very few have changed how a return moves through the office because of it. The gap between those two facts is where most of the useful lessons from the last two tax seasons sit.
This piece is for firm owners and operations leads who have run a pilot or two and are trying to decide what deserves a real place in the workflow.
General chat assistants are good at talking about tax, not doing it
A general-purpose assistant can explain the difference between a 1065 and an 1120-S, draft a client email about missing documents, or summarize a revenue procedure. That is useful, and most staff already use it that way.
What it cannot do on its own is see the client’s documents, the draft return in the firm’s tax software, or the firm’s own review standards. Without those three things, it has nothing to check. Anything it says about a specific return is a guess dressed in confident prose, which is the worst possible output for a profession built on signing your name to the numbers.
The work that benefits most is comparison work
The stage where AI has earned its place fastest is review. Review is mostly comparison: does the W-2 in the binder match wages on line 1a, does every 1099-B lot show up on Form 8949, did the K-1 figures carry to the right schedule, is anything from last year’s return missing this year.
That work is repetitive, it is high-stakes, and it is exactly what a senior reviewer should not be spending the last week of March on. Software that reads source documents and compares them line by line against the draft return does this well, as long as it shows its evidence. A flag that says “this number looks wrong” is noise. A flag that says “the brokerage statement on page 4 shows $12,480 in proceeds and the return shows $1,248” is something a reviewer can clear in ten seconds.
If you are evaluating tools in this category, what tax return review software actually needs to do is a reasonable checklist to start from.
Three questions to ask any AI tool before it touches client data
- Does it read the source documents, or only the data someone typed in?A tool that only checks entered data inherits every keying mistake.
- Can a reviewer trace every output back to a page?If the answer is “trust the model,” the answer is no.
- Does it work with the tax software you already run?Mid-season software migrations cost more than any tool saves. Firms on CCH Axcess, UltraTax, Lacerte, Drake, or ProConnect should not have to switch to adopt AI.
Security belongs on that list too. The IRS spells out the baseline safeguards tax professionals are expected to maintain in Publication 4557, and any vendor handling client files should be able to show a SOC 2 report without much prompting.
Where AI still falls short
Judgment. Whether a client qualifies for a position, how aggressive a firm wants to be, what a client is going to be upset about next April. None of that belongs to software, and tools that pretend otherwise tend to create more review work, not less.
The realistic 2026 picture is narrower and more useful: AI handles the reading, extracting, and comparing, and the CPA spends the recovered hours on the parts of the job that need a CPA. At Filed, that is the design principle we build around, and it is the test worth applying to any tool you are considering.
The short version
Start with review, because the evidence is easiest to check there. Demand traceability. Keep your tax software. And judge the pilot by one number: how many hours your senior people got back in the weeks that matter most.
