Tax Preparation
AI Tax Preparation Software for CPA Firms: Judge It by the Draft
Jump to a section
Key takeaways
- Evaluate the draft the Reviewer receives: The Reviewer signs the output, so that's what you test.
- Test 1, trace three figures: Pick them yourself and ask for the document, page, and printed value behind each one.
- Test 2, missing items already asked: Leave one document out of your test folder and see whether the request list catches it before drafting.
- Test 3, the Reviewer signs off: Edit one figure and check that the log records who, what, when, and why.
- Get data answers in writing: Send the vendor questions about storage, training use, and retention before the demo.
- The software drafts; people decide: The preparer prepares and the Reviewer concludes, whatever tool sits underneath.
It's March. Your Reviewer opens a drafted return and starts asking the same three questions: where did this number come from, what's still missing from the client, and who changed what since the draft was built. A good draft answers all three before anyone asks. A weak one sends the questions back down the hall, and the hours you hoped to save just move from the preparer's desk to the Reviewer's.
That's an easy test to skip when you shop for AI tax preparation software, and it's the one that matters most when you weigh AI for CPA firms. You watch a demo. The demo looks fast. But a finished draft reads the same on screen whether its figures are right or wrong, and a demo only shows you the screen. This piece gives you a different test: judge the software by the draft it hands your Reviewer, using a folder you built yourself.
Why the demo is the wrong test
A demo is a sales tool. The vendor chooses the inputs, so the thing you'll live with all season may never show up on screen.
What a demo hides
A demo can run on a clean folder the vendor prepared in advance, where every document is the right year, every page is there, and the figures may have been picked because they extract well. What you may not see is the folder that arrives in real life: a prior-year 1099 mixed in with this year's, a brokerage statement missing page 3, a scanned receipt at an angle.
So bring your own folder. Build a synthetic household (call it the Doe Household) with made-up figures and no client data in it at all. Put in one document from the wrong year and leave one document out.
What the Reviewer sees in March
Think about the file from the Reviewer's chair. They don't care how the draft was built. They care whether they can check it without reopening the whole folder. AI tax software that drafts fast but leaves the Reviewer guessing hasn't saved your firm anything.
Test 1: does every figure trace to a page
In AI tax preparation, a drafted figure is quick to check when it points to where it came from. Without that pointer, your Reviewer has to go back into the folder to find its source.
Ask for the source of three figures
Pick three figures from the draft at random. Don't let the vendor choose. For each one, ask: which document, which page, and what value is printed there? On your Doe test file that might be a wage figure from a W-2 and an interest figure from a 1099-INT, both made up.
A good answer is a pointer you can open in one click that lands on the page with the printed value next to the drafted one. A weak answer is "the system read it from the folder." If the pointer doesn't exist for a figure you picked at random, assume it doesn't exist for the figures you didn't pick. Firms that want a closer look at what a traced figure needs can read our piece on making AI cite the page behind each figure and the traceable evidence page.
What a usable source pointer contains
Four things: the figure, the document name, the page, and the value as printed on that page. Anything less leaves your Reviewer doing the lookup by hand.
Tie-out of totals
Where a document prints a total, the figures pulled from it should add back to that total, and the draft should show the tie-out instead of asserting it. When a total doesn't tie, the draft should flag it for a person rather than quietly adjust a line to make it fit.
What you're testing here is an inspectable process: your Reviewer can see how each number got there and decide for themselves.
Test 2: are missing items already asked
A missing document found late can mean rework. A draft that only discovers the gap at review may need a second pass.
Request list at triage
The missing items should surface when the folder is first read, before any drafting, as a list you could send the client. This is where the document you left out of your Doe folder earns its place. Ask the vendor to show the request list the tool built for your test folder. If the missing document isn't on it, you've learned something no demo would have told you. Same with the wrong-year document: it should be flagged, not drafted.
What is still open
At handoff, the open items belong on the draft itself, each one with what's missing and why. Not buried in a notes field, and not in someone's inbox. Your Reviewer should know exactly what they're signing around.
Test 3: who signs off
This is the test that protects your Reviewer's name. The Reviewer concludes. The software never does.
Reviewer control
Ask three things. Can the Reviewer reject a drafted figure? Can they send it back to the preparer with a note attached? Can they hold sign-off until the open items clear? Any tool that suggests review is optional, or that the draft is ready to file as-is, fails this test. Your firm's Reviewer sign-off has to stay a real decision.
Audit log of changes
During the demo, change one figure yourself. Then ask to see the log. It should record who made the change, the prior value, the new value, the reason given, and the time. If a figure can move without a reason on the record, your Reviewer can't tell a fix from an accident.
Data handling questions to ask
For the rules on client data, see the post on Section 7216 and AI tools. This section is about the questions to ask, not the rules.
Where documents go
You need to know where client documents are stored and processed, whether they're used to train any model, who at the vendor can see them, and how long they're kept after a file closes.
What the firm keeps
Ask who keeps the drafts, the source pointers, and the change log, and what your firm still holds if the subscription ends. The workpapers behind a return are your firm's record, so the answer should be clear before you sign anything.
Questions to send a vendor before a demo
Paste these into an email and ask for written answers before you book the demo.
For any figure in a draft, can the Reviewer open the source document at the exact page in one click? Please show an example from a test file we provide.
When a document prints a total, does the draft show the tie-out of the extracted figures to that total? What happens when they don't tie?
At what point does the software list missing client documents, and can that list be sent to the client as-is?
Can a Reviewer reject a drafted figure, send it back with a note, and hold sign-off until open items are cleared?
What does your change log record when someone edits a figure? Can we see one?
Where are client documents stored and processed, and are they used to train any model?
Who at your company can access our client documents, and for how long are they kept after a file closes?
If we end our subscription, what do we keep: drafts, source pointers, change logs?
The three-test demo checklist
Run this in every demo, on the same synthetic folder.
- Bring the Doe Household folder: made-up figures, one wrong-year document, one document left out.
- Test 1: pick three figures at random and open the source page behind each.
- Test 1: check that one printed total ties out to the extracted figures.
- Test 2: find the missing document on the request list built at triage.
- Test 2: confirm the wrong-year document is flagged, not drafted.
- Test 3: edit one figure and read what the log records.
- Test 3: have the Reviewer send one item back with a note and hold sign-off.
- Get the data-handling answers in writing.
Where TracePrep fits
TracePrep is tax-prep software for CPA and EA firms, built around the outcome this piece describes: return drafts ready for your own Reviewer to review. The plugin drafts; the Firm's preparer prepares; the Reviewer signs off. Access to the plugin is by waitlist. You can see how the plugin drafts or start at the TracePrep home page.
Bring your own folder
Build the Doe folder once, then run the three tests on every tool you look at. Judge what lands on your Reviewer's desk, and join the TracePrep waitlist when you're ready to see a draft built that way.
<!-- lint-ignore: missing-required-reciprocal-anchor per brand form v4.0 plugin-copy rule -->
Not tax advice. TracePrep is a software product from TracePrep Inc., not a CPA firm. This content is operator experience for educational purposes; it is not tax advice and has not been reviewed by a licensed CPA. Consult a qualified tax professional for guidance on your specific situation. Source citations in TracePrep workpapers support audit defense but do not guarantee IRS outcomes.
Frequently asked questions
Does AI tax software prepare the return?
Can I trust a figure the AI extracted?
What should I bring to a demo?
Does AI tax software replace the Reviewer?
The TracePrep Team writes the TracePrep blog for CPA, EA, and accounting firms. TracePrep is a tax-prep platform, and it also produces the Reasonable Compensation Study.
Related reading
- Tax Preparation
1040 Review Checklist: What the Reviewer Checks Before Sign-Off
A 1040 review checklist for the Reviewer before sign-off: folder and open items, source ties, prior-year changes, and review notes a preparer can clear.
TracePrep Team12 min read - Tax Preparation
Tax Season Capacity: Measure Returns per Preparer Before You Hire
How a CPA firm owner measures returns per preparer and where preparer hours go per return, with a capacity worksheet to fill in before a seasonal hire.
TracePrep Team9 min read - Tax Preparation
Tax Workpaper Software: Every Figure Traced to Its Source Page
How tax workpaper software should trace every 1040 figure to its source page, with a synthetic workpaper index and what to require before your Reviewer signs.
TracePrep Team8 min read