AI Tax Prep Software for CPAs: What to Test Before You Switch

Before switching, run ten of your own filed prior-year returns through each candidate tool and score field accuracy, correct mapping into your tax software, traceability to the source document, reviewer time, and what the tool does when it is unsure. Add planted errors and a security check. Switch only if it beats your pass line.

The reason to test rather than trust a demo is simple: every filed return is a marked exam paper. You already know the right answers, so a tool can be scored without anyone's opinion. Most roundups list features and skip this step, and most vendors do not publish prices, so the test is also how you find out what a tool costs you in review time. And it exposes the behaviour a demo never shows: what the tool does with a blurred figure or a corrected form.

Follow me on Instagram@sagnikteaches

What you are choosing between

The category is moving quickly, so first sort the options by how they fit into your workflow. Facts below were read from vendor pages in November 2026. Only Juno publishes a per-return price; for the others, each row says what to ask.

Connect on LinkedInSagnik Bhattacharya

Two points shape the test. First, Smart Prep is a framework with partners, so you can trial more than one extraction engine without leaving your practice platform. Details of what sits inside Canopy are in Canopy's AI features for a tax practice, and cost mechanics are in Canopy pricing, client credits and add-ons. Second, partner tools bill separately from the platform, so your cost per return has more than one line. For turning quotes into a figure, use the method in running a pilot and measuring cost per return.

Subscribe on YouTube@sagnikteaches

Choosing the ten returns

A test that uses ten easy returns will pass everything and tell you nothing. Build a set that includes the cases that cost you time. A balanced set for a small firm might look like this:

Pull the original client documents and your filed return for each. Together they are your answer key. If you cannot locate the original documents for a return, replace it, since without them you cannot score extraction.

Protect client data while you test

Using prior-year returns in a trial means sending real client data to a vendor. Handle that before the first upload.

1. Get consent or redact. Either have written consent from the clients concerned, or redact names, identification numbers, addresses and account numbers. Redaction affects results slightly, so note it in your scoring sheet.
2. Read the trial terms. Ask how long the vendor keeps uploads, whether they are used to train models, and who can see them. Get the answers in writing.
3. Ask for security documents. A SOC 2 report or equivalent, a list of sub-processors and the data location.
4. Update your written information security plan. Add the tool, who has access and how data is deleted at the end of the trial.
5. Delete after the trial, and ask the vendor to confirm in writing that it has.

If a vendor will not answer the first two questions, stop there. No feature list is worth a data problem.

The scoring sheet: what to record for every return

Use one row per return per tool. The columns that matter:

• Fields extracted and fields correct. Count every numeric field the return needs, then count matches against the source document. Divide for a percentage.
• Largest single error in dollars. One $5 slip and one $5,000 slip are not the same.
• Mapping accuracy. Of the correct figures, how many landed on the right line of your tax software?
• Traceability. Can a reviewer click from a figure to the page it came from? Score yes, partly or no.
• Behaviour when unsure. Did it leave a field blank and flag it, or fill it in with a guess?
• Reviewer minutes from opening the draft to being ready to sign off.
• Final tax figure difference against your filed return.

The behaviour-when-unsure column is the one most firms forget and the one that matters most. A blank flagged field costs a minute. A confident wrong number can cost a client a penalty.

Two tools, ten returns, 612 fields: a scored run

Here is an illustrative scoring run for a four-preparer firm testing two tools. All figures are made up to show the method. Across the ten returns there are 612 numeric fields to extract and a baseline of 52 minutes a return to key in and check those figures by hand.

Neither tool passes. Tool B is clearly worse: six guesses and a $4,380 error mean it fails the safety rule outright. Tool A is the better engine but misses on accuracy and on the two returns where the final figure differs. The right reaction is not to buy Tool A. It is to examine its 23 wrong fields. If 15 of them come from the same two document types, ask the vendor whether a fix exists, then rerun those returns. If the errors are scattered, the tool is not ready for you.

Notice what the numbers show that a demo could not. Tool A saves 18 minutes a return against baseline (52 minus 34), which is about three hours across ten returns. That is real, but only if the reviewer trusts the output enough not to re-check every field, and the guessed-field count is why that trust has to be earned.

Planted errors: test the reviewer as well as the tool

Extraction accuracy is one half. The other is whether your reviewer, working from the AI draft, spots mistakes. Before a reviewer opens a draft, plant five known problems in the source set. For example:

1. A transposed digit in one figure ($3,617 typed as $3,671).
2. A statement from the wrong tax year placed in the folder.
3. A corrected document alongside the original.
4. A missing statement, such as one investment account with no document.
5. A duplicate upload of the same statement.

Run the reviewer through the draft without telling them which problems exist. Record how many they find and how long it takes. An illustrative result: the reviewer catches 3 of 5, and misses the wrong-year statement and the duplicate. That is not the reviewer's fault. It shows the draft hid those problems. A good tool flags a mismatch between the document year and the return year, and flags duplicates. A poor one passes them through, and the review process has to catch them by hand. For what an AI review catches compared with your software's diagnostics, see AI return review vs software diagnostics.

A cheap baseline: what a general assistant does with a pay statement

Run one return through a general assistant as a comparator. It shows what you are paying a specialist tool for. A prompt and an illustrative output:

From the attached pay statement, return a table with columns: field,
value, page, confidence. Include gross pay, tax withheld, retirement
contributions and the tax year. If a value is not visible, write
"not shown". Do not estimate.

A typical result gives a clean table of five rows. What you would fix: the assistant may report year-to-date figures as annual figures, round to whole dollars, or mark a smudged value as high confidence. Compare each row against the page. If the general assistant already matches a specialist tool on your simple returns, the specialist has to earn its fee on returns 3 to 10, where long tables, scans and carryovers appear. A general assistant also lacks the link into your tax software, which is where the real time goes.

The integration test: does the draft land in your software?

Extraction is only useful if the data flows into the tax software your firm already uses. Canopy's page names Drake, CCH Axcess, ProConnect and UltraTax as examples of software Smart Prep works alongside, and Juno says it pushes data to Drake, Lacerte, ProConnect, ProSeries and CCH Axcess, among others. Neither statement tells you how clean the handoff is on your setup. Test it by checking five things on each return:

• Does the data arrive in the right input screens, or in a worksheet someone must copy across?
• Do names, identifiers and dependants map without retyping?
• Do the totals in your software match the source documents before any review?
• Do your software's diagnostics run clean, or throw errors the AI draft caused?
• Does the return's final figure match your filed return?

Count the retyping. If a preparer still has to rekey 20 per cent of the fields, the tool's accuracy score overstates the saving. A tool at 96 per cent accuracy with clean mapping often beats one at 98 per cent that needs manual transfer.

Silent failures to look for in each draft

Run through that table with each returned draft. The most common surprise is the corrected statement: tools often read whichever page arrived first. If a vendor says it flags prior-year changes, as Juno's pages do, test that claim on returns 8 and 10 rather than taking it on trust.

What 18 minutes a return is worth, and the ceiling it sets

Turn the saving in the worked example into a ceiling on what you should pay. Say the firm files 300 returns a year with the tool, and Tool A saves 18 minutes on each. That is 5,400 minutes, or 90 hours. At a loaded preparer cost of $45 an hour, the time is worth $4,050 a year. Divide by 300 and the break-even is $13.50 a return. Any combination of partner subscription and platform fee above that, per return, means you are paying more than the time is worth. Any figure well below it, say $6 a return, leaves $2,250 a year of room, which is before you count the capacity gained. Ask each vendor for a quote and compare it with your own ceiling instead of with another vendor's number. If a quote is per user, divide it by the returns each user handles to get the same figure.

Published prices make the gap concrete. Against Juno's list price of $50 a return, Tool A's 18 minutes, worth $13.50, leaves a shortfall of $36.50 on every return. A per-return price only pays where the tool saves more than an hour at that hourly cost, which is why returns 3 to 10, the long ones, decide the matter. A flat firm-wide fee, Filed's model, works the other way: divide the annual quote by every return you file, simple ones included.

When the test set is too kind to the tool

Picture a firm of three preparers that tests a tool on eight returns, all simple employee returns with one or two documents. The tool scores 99.1 per cent and the firm signs for a year. Then the first self-employed client of the season arrives with 140 lines of bank transactions and a shoebox of photographed receipts. Accuracy on that return falls below 90 per cent, and the reviewer spends longer than the old manual method took. Nothing is wrong with the tool; the test never asked the hard question. The fix is built into the ten-return table above: at least four of the ten should be the kind of return that eats your time today. If you cannot name which returns those are, look at where your reviewers spent their longest hours last season, and pick from there.

Writing up the result for the partners

End the test with a one-page note that a partner can read in three minutes. Put the pass line at the top, then the scoring table for each tool, the break-even figure, the security status and one recommendation: adopt, retest after a fix, or decline. Add the open questions you sent the vendors and any answers. Keep the note, the scoring sheets and the vendor emails together, so that a year from now, when someone asks why you chose this tool, the answer is a file rather than a memory.

Timing the test around your calendar

Do the test early. A sensible plan for a small firm runs over three weeks, a few months before the busy season.

• Week one: pick the ten returns, locate the originals, redact or obtain consent, and write the pass line.
• Week two: run both tools on all ten, record the sheet, then run the planted-error review.
• Week three: send each vendor your scored results and ask them to explain the misses. A vendor that engages with specifics deserves another round; one that repeats a pitch does not.

Then decide. If the best tool passes, run a parallel period: the first ten live returns are prepared the old way as well, and the two are compared. If none passes, you have lost three weeks and learned what to ask for next time, which is cheap compared with a year's contract.

What to ask in the contract once a tool passes

1. Billing unit and price for your volume. Per return, per user, per client, or a mix? Get a figure in writing.
2. Partner billing. If the tool runs inside a platform, who bills, and are there separate minimums?
3. Availability. Is it generally available, or still a pilot? Pilots can change terms.
4. Return types covered. Written confirmation for the types in your set, not a general statement.
5. Exit terms. Can you cancel after the season? What happens to stored data?
6. Accuracy remedies. What does the vendor do when its extraction causes an error that you catch late?

Review the tool again after 90 days of live use; the routine in reviewing an AI tool after 90 days: keep, fix or cancel fits well. The wider picture of practice software, including how much automation each platform brings, is in accounting software for CPA firms compared. If you would rather work out the test with someone, that is what my AI implementation consultation covers, and the AI guide lists the rest of this series.

Questions CPAs ask before testing AI tax prep

When in the year should a firm run this test?

Run it well outside the busy season, ideally a few months before it starts. You need time to repeat a failed test, negotiate terms and train staff. Switching tools in the middle of filing season turns every teething problem into a missed deadline.

Can I test with real prior-year client returns?

Only with care. Prior-year returns hold sensitive data, so get written client consent, or redact names, identification numbers and account numbers first. Check the vendor's trial terms on retention and model training, and delete test data afterwards in writing.

How many returns make a fair test?

Ten is the minimum that exposes different return types without becoming a project. Choose them to include at least one messy client, one scanned or phone-photo set and one with a corrected document. Ten easy returns will flatter any tool.

Should I replace my tax software to get AI prep?

Usually not as a first step. Several AI prep tools sit alongside the software you already use and fill it in, so you can trial them without moving your whole workflow. Replacing core software is a separate, larger decision that deserves its own test.

Further reads

• Tax Preparation AI Software: TaxDome, Canopy and ChatGPT Compared — Compare practice platforms that bundle AI.
• Canopy vs TaxDome for Tax-Heavy Practices: Which Fits the Season? — Which practice platform fits a tax-heavy firm.
• AI Tax Return Prep for Small Practices: Extract, Review, Sign Off — Extract, review and sign off once you have chosen.
• AI for Small CPA Firms: Chasing Missing Client Documents in Tax Season — Fix document chasing before you automate preparation.
• AI Tax Organiser: Build One That Chases Missing Documents — Build an organiser that chases what is missing.
• How Small Tax Practices Use AI Through the Busy Season — Plan AI use around the busy season.
• How to Set Up Human Review for AI Work Without Slowing Down — Design the human review step so it stays quick.
• AI Tax Research Tools Compared: Citations, Cost and Checks — Blue J, CoCounsel Tax, TaxGPT, Bizora, ChatGPT and Perplexity compared on cost and citations, plus a four-point citation check and a ten-question test.
• Best AI Tools for US Tax Consultants: Research, Notices, Letters — General assistants against tax-research tools for notices, research memos and client letters, with prices, consent rules and a one-week test.
• Can AI Summarise IRS Transcripts for a Tax Resolution Case? — A small-practice workflow: redact an IRS transcript, turn its codes into a plain-English case summary, check it, and keep the client updated on a fixed cadence.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Canopy's Smart Prep page and its January 2026 announcement of the Filed integration; Juno's home and pricing pages; Filed's home, pricing and integrations pages; TaxGPT's pricing page; all read in November 2026.