AI tax preparation: what it does and how firms use it

AI tax preparation is software that reads a client's source documents, extracts the figures, and enters them into the firm's tax software as a draft return, flagging anything uncertain with the source attached. It takes over the mechanical middle of prep: sorting, keying and first-pass checking. A preparer still reviews the draft and signs it, and every judgment call stays with the firm. The practical change is that staff start from a traced draft instead of a blank return.

If you want the category basics (what the technology is and where it came from), our complete guide to AI tax preparers covers them. This piece is for the firm owner deciding whether and how to put it to work: what happens to a return, where the limits sit, how to test a tool, and how firms roll it out without wrecking a season.

What does AI tax preparation actually do, step by step?

It follows the same path a good preparer does, in five steps, and leaves a trail at each one. The difference from older scan-and-populate tools is the trail: every number should point back to the page it came from.

1. Intake and sorting. The client's documents arrive the way they always do: a consolidated brokerage statement, three W-2s, a K-1 with twelve pages of footnotes, a phone photo of a 1098. The software classifies each document, matches it to the client and the tax year, and builds the binder.
2. Extraction. Figures are read off each document and tied to their source box. A K-1 is where this earns its keep: ordinary income, guaranteed payments, capital account changes and the footnotes that modify them all have to come through intact.
3. Reconciliation. Before anything is entered, the extracted data is checked against last year's return and against itself. The second K-1 that never arrived, the brokerage account that closed, the 1099-NEC with no withholding and no estimated payments on file: these surface now instead of at review.
4. Preparing the return. The figures are entered into the firm's own tax software as a draft. Values the system is confident in go straight in. Values it is not confident in are flagged, with the source document attached, instead of guessed.
5. Review with citations. The draft is checked line by line against the source documents, and the reviewer gets a ranked list of issues, each tied to the page that supports it. Verification becomes a click instead of a hunt.

Two things separate good AI tax prep software from a demo. It flags uncertainty instead of guessing, and it learns: when a preparer corrects how the firm handles, say, 8949 naming conventions, that correction should apply to the next return automatically.

What doesn't AI tax prep do?

It doesn't make judgment calls, and it doesn't sign. Everything that depends on facts outside the documents, or on professional judgment about those facts, stays with the preparer and reviewer.

That covers a lot of ground:

• Elections and positions: whether to elect out of bonus depreciation, whether a position has enough support to take, which safe harbor applies.
• Facts the documents can't show: whether a rental was used personally, whether a client materially participates, whether that "consulting income" is really a hobby.
• Client conversations: explaining a balance due, pushing back on a deduction the client wants, spotting that a life change should reshape the plan.
• Sign-off. The IRS holds the practitioner, not the software, to its standards of diligence under Circular 230. A drafted return is still your return.

‍

Accuracy benchmarks make the same point from the other direction. Filed scores 94% line-by-line on TaxCalcBench, which clears the bar for supervised production use. It does not clear the bar for running unsupervised, and no current system does. The errors that remain are the ones that need a professional.

How much time does AI tax preparation save?

Most of the saving shows up at review, not extraction. Extraction and entry get faster almost immediately, but the bigger gain, and the one worth measuring, is what happens to the reviewer's day.

Firms using Filed report 64% shorter review cycles and 42 minutes saved per return. Those numbers come from the review stage, because that is where senior time goes: a reviewer who has to re-derive every figure from the binder spends hours on clerical checks before reaching the questions that need a partner. When each number already traces to its source, that reviewer starts at the judgment work.

Be skeptical of any "minutes per return" figure that only describes extraction. A tool can populate a draft in ten minutes and still leave an hour of fixing for your team. Measure from upload to sign-off.

How should a firm evaluate AI tax software?

Test any AI tax preparation software on your own worst returns, and judge it by what it does when it's unsure. A vendor's accuracy claim describes their test set, not your client base.

A short checklist:

1. Your tax software, unchanged. The tool should work inside what you already run (CCH Axcess, UltraTax, Lacerte, Drake or ProConnect) with no migration. Switching tax software to adopt AI is a multi-season project most firms can't absorb.
2. Your return mix. Check coverage for everything you file: 1040, 1065 with K-1, 1120, 1120-S, 1041 and 990. A 1040-only tool helps less than it looks if half your hours go to entities and trusts.
3. A citation on every number. If you can't click a figure and see the source box, your reviewer is back to checking everything by hand.
4. Flagged, not guessed. Run ten real binders from last season, including the shoebox client and the 300-page brokerage statement. Count the values it got wrong without flagging. Uncertain-and-flagged is fine; wrong-and-confident is the risk.
5. Corrections that stick. Ask what happens the second time the same issue comes up. If the answer is "the preparer fixes it again," the tool won't get better with use.
6. Data handling in writing. Ask for the vendor's SOC 2 report, where data is processed, and whether client data trains shared models. Also ask how their handling fits the IRS rules on disclosure and use of return information under Section 7216.

For how specific platforms stack up on these criteria, see our comparison of AI tax prep platforms.

How do firms roll out AI tax preparation?

In stages, starting outside busy season, with review before prep. The firms that get the most out of it change one thing at a time and measure each step.

A rollout that works for most mid-sized firms:

1. Start with review. Run AI review on returns your team has already prepared. It needs no change to how returns are built, it can be running within a day, and it shows reviewers how the citations work before anyone relies on a draft.
2. Pilot prep on last season's binders. Pick a slice of real clients, prepare them with the tool, and compare against what your team filed. You get accuracy data on your own work with no client risk.
3. Expand by complexity. Go live first on returns heavy in W-2s and 1099s, then Schedule C and rental clients, then K-1 heavy returns and entity work. Each tier shows you where the tool needs teaching.
4. Redefine the roles. Admin staff own document upload and the missing-items list. Preparers move from keying to reviewing drafts. Reviewers work from the ranked issue list instead of the full binder.
5. Teach it as you go. Every correction your team makes should become a firm rule. The firms that see the biggest gains in season two are the ones that treated season one corrections as training, not cleanup.

The timing matters as much as the sequence. Set up and pilot between October and January. Nobody should be learning a new workflow in March.

How does Filed handle AI tax preparation?

Filed does AI prep and review for mid-sized and high-volume firms, typically those filing 2,000 to 10,000 returns a year. Filed Prep reads the client's documents and prepares the return inside CCH Axcess, UltraTax, Lacerte, Drake or ProConnect, across 1040, 1065 with K-1, 1120, 1120-S, 1041 and 990 returns. Filed Prep enters the figures it is certain of, flags the ones it isn't, and links every number to the source box it came from.

Filed Reviewer compares source documents against the draft return line by line and gives the reviewer a ranked, cited list of issues. Prowork, a desktop app and Excel add-in, sits alongside the tax software and reviews inline as the preparer works. Prowork ships with about 35 pre-built skills, and when a preparer corrects it once in plain language, Prowork hard-codes that fix permanently. Filed does the analytical work so your CPAs can spend their time on the judgment work.

Frequently asked questions

Is AI tax preparation accurate enough to rely on?

Accurate enough to draft and to check, not to sign unsupervised. Current AI tax software extracts and enters standard documents well and flags what it can't read confidently. Every figure still needs to be verifiable against its source, and a credentialed preparer still reviews and signs the return.

Can AI tax software prepare business returns?

Yes, though coverage varies a lot by vendor. Some tools handle only 1040s. Others read partnership, corporate, trust and exempt-organization documents and prepare those returns too. Entity work is where document loads are heaviest, so check coverage for your full return mix before you buy.

Does Filed work with my tax software?

Filed works with CCH Axcess, UltraTax, Lacerte, Drake and ProConnect, and supports 1040, 1065 with K-1, 1120, 1120-S, 1041 and 990 returns. Filed runs inside the tax software your firm already uses, so there is no migration, and Filed Reviewer can be reviewing returns within a day.

Is it safe to send client documents to AI tax prep software?

It can be, if the vendor can show its controls. Ask for a SOC 2 report, confirm where data is processed and stored, and get in writing that client data does not train shared models. Filed is SOC 2 compliant, with ISO 27001 certification in progress, and client data is never used to train shared models.

If you want to see AI tax preparation on your own returns rather than a sample set, book a demo and bring your hardest binder.

‍