AI Software for Tax Professionals: Run a Pilot and Measure Cost per Return

Few AI tax tools publish a per-return price. Juno lists $50 a return whatever the complexity, slightly less at volume; Filed quotes a flat firm-wide fee based on how many taxpayers you serve; Blue J charges $1,498 per user a year for research. Your real figure is software cost plus review time, divided by returns, measured in a four-week pilot.

That formula matters more than any vendor's headline number, because the cheapest-looking tool is often the one that moves cost from the subscription line into your own hours. A pilot gives you the real number from your own returns, with your own clients' messy documents, rather than from a demo that uses clean files. Pricing models also favour different practices: a flat $50 a return is easy to recover when the tool saves an hour on a long partnership return, and hard to recover on a simple return that only takes thirty minutes by hand.

Follow me on Instagram@sagnikteaches

Why a pilot beats a demo for tax work

A demo shows the vendor's best case. A pilot shows yours. Tax work has an unusual property that makes pilots especially informative: every task has a known right answer, because you can compare the output against a return you have already filed. That makes it possible to score a tool without opinion.

Connect on LinkedInSagnik Bhattacharya

The pilot has five parts: one task, one metric, a pass line, a time limit and a stop rule. If you cannot write all five on a single page before you start, you are not ready, and a salesperson will fill the gap with their own measures. Each part gets its own section below.

Subscribe on YouTube@sagnikteaches

Step one: pick a single task with a right answer

Choose from tasks where the output can be checked line by line. Four good candidates, with the check that applies to each:

• Extracting figures from client documents into your tax software. Check against the source documents, line by line.
• Building a document checklist from last year's return. Check against what you actually needed to ask for.
• Drafting a client query email for missing information. Check for accuracy of amounts and dates.
• First-pass review notes on a prepared return. Check against the issues you found yourself.

Leave tax research and advice for a separate test, since they need a different kind of checking. The extract, review and sign-off workflow for small practices sets out the preparation task in more detail, and AI return review compared with software diagnostics explains what each kind of check catches.

Step two: set pass or fail before you see a result

Write the pass line down first, because people move goalposts when they like a tool. For an extraction task, a workable rule has three parts:

1. Accuracy: at least 98 per cent of extracted numeric fields match the source document, with no single error above $100.
2. Time: total time per return, including checking, is at least 25 per cent lower than your baseline.
3. Safety: no figure is ever invented. One fabricated number fails the pilot.

The percentages are illustrative; set your own. The shape is what matters: an accuracy floor, a time target and a zero-tolerance rule for invented data. If the tool passes accuracy but fails on time, you are paying to do the same work with extra steps. If it passes time but fails accuracy, you have swapped typing for proofreading, which is often slower than it feels.

Step three: measure your baseline before you switch anything on

You cannot claim a saving without a before figure. For two weeks before the pilot, log the following for the task you chose, on at least 15 returns:

• Minutes to complete the task, start to finish.
• Errors found at review, and how long each took to fix.
• Client queries triggered by that step.

An illustrative baseline for a three-person practice: document entry on an individual return with a few income sources takes 38 minutes, the reviewer finds an average of 1.4 errors, and each takes 4 minutes to fix. That is roughly 44 minutes per return. Everything the pilot reports gets compared with that 44.

Cost per return for a three-preparer practice, line by line

Here is the sum, built from assumed figures. Prices from vendors are placeholders you must replace with real quotes.

In this illustration the pilot fails commercially: the software costs $10 a return and saves $9.75 of time. It breaks even at a monthly quote of about $292, so a $250 quote would pass, as would a saving of 15 minutes at the original price. That is the useful result. The tool is not bad, but at that price and that saving it does not pay. Notice also that nothing here counts the extra capacity: if saved time lets you take 10 more returns in the busy season at a $350 average fee, the answer changes. Add a second row for that only if you will really take the work on.

Now compare with a research tool, where the cost is per user rather than per return. Blue J's individual plan is $1,498 a year per user, according to its pricing page. If a preparer spends three hours a week on research and the tool saves one hour of it, that is about 50 hours a year, or $2,250 at $45 an hour. The tool pays back about 1.5 times on those assumptions. Different tools need different sums, which is why you pick one task at a time.

Per-return pricing: how many minutes must $50 buy?

Juno's pricing page sets one list price per return, $50, for a simple wage-earner return and a 40-page partnership return alike. The price falls by up to $5 a return as volume approaches 2,500 returns, and a first order can carry an introductory discount. Turn that into the minutes the tool has to save. At a loaded preparer cost of $45 an hour, $50 buys about 67 minutes; at $80 an hour, the rate Juno's own savings calculator assumes, it buys about 38. A tool priced this way pays on the long returns and struggles on the short ones, so put your complex work through it in the pilot and keep a separate tally for the simple returns.

Two more details change the sum. Purchased returns are valid for the period you bought them for, one year for a new customer, so unused returns are a cost: buy 300 at $50, use 240, and the real price is $62.50 a return. And logins scale with volume, two users for every 100 returns, which matters if four people need access in a 150-return practice. Ask any per-return vendor the same two questions: what happens to unused returns, and how many users the purchase includes.

What vendors publish and what you must ask

When I checked in November 2026, pricing transparency was patchy. Blue J published its individual price and a seven-day free trial, and Juno its per-return list price and five free preparations. Filed's pricing page shows custom quotes based on the number of taxpayers you serve, charged as a flat monthly fee for the whole firm with unlimited seats and no per-return charge. TaxGPT listed its Pro, Pro Plus and Enterprise tiers but no price, and asked for a demo first. Canopy's Smart Prep page says each of its AI preparation partners (Filed, Byron and Juno) needs its own subscription, priced on the partner and the number of returns processed. For preparation tools, send each vendor the same written questions.

1. What is the billable unit: user, return, client or document?
2. What is the price for my volume, say 300 returns a year and three users?
3. What is included in the pilot, and is it free?
4. What happens to my data after the pilot, and is it used to train models?
5. Which tax software can your output flow into, and by what route?
6. What are the cancellation terms after a trial or annual signup?

Ask for the answers in an email and keep it with the pilot file. A vendor that replies with a price against your volume in a day is easier to work with than one that insists on a call first. For a view of platforms that bundle AI inside practice management, see TaxDome, Canopy and ChatGPT compared.

Consent and WISP check before any client data moves

Do this before the pilot, not after. Two documents need attention.

Client consent. Rules on sharing return information with outside software differ by place, and many require written consent before client data goes to a third party, including an AI provider. Check the rules that apply to you and read your engagement letter. An illustrative clause you might discuss with your adviser:

"We may use secure software, including tools that use artificial
intelligence, to prepare and review your return. Your information is
processed only to provide our services, is not used to train
third-party models, and is protected under our information security
plan. You may decline this use and we will prepare your return
without it. Please sign below to confirm your choice."

That wording is a starting point for a conversation with a qualified adviser, not approved text. Keep a signed copy per client in the pilot file.

Written information security plan (WISP). A WISP is the written plan that describes how you protect client data. Before the pilot, add a short section for the new tool: who can access it, how logins are protected, where data is stored, how long it is kept, and how you delete it. Ask the vendor for its security documentation, such as a SOC 2 report (Blue J says its service has SOC 2 Type II), and attach it. If a vendor cannot show how it handles data, do not feed it a live return.

Running the four weeks

A schedule that works for a small practice:

• Week 0: baseline logs, consent letters, WISP update, vendor questions answered.
• Week 1: five returns with the tool, every figure checked. Fix settings, not people.
• Week 2: ten returns, two preparers, same checking.
• Week 3: ten more, including the two messiest clients.
• Week 4: five returns timed under normal conditions, then the decision meeting.

Record for each return: client type, documents received, minutes, errors, errors found after review, and whether the client queried anything. A simple sheet with those six columns is enough. The pilot also needs a person who owns it. In a small practice that is usually the owner, because the buying decision sits with them.

A scored return: how to mark one

An example of marking a single extraction. A client's document set has 42 numeric fields. The tool fills 41. Of those, 39 match the source, one has a transposed digit (a $1,284 figure entered as $1,248) and one has a figure pulled from the wrong box. The field it left blank is a number the document does not show clearly. Score: 39 of 42 correct is 92.9 per cent, below the 98 per cent line, and the $36 difference is under the $100 cap. Result: accuracy fail, safety pass. Record the type of each error, because patterns matter. If the same box confusion happens on every form from one issuer, you can write a rule to check that box by hand and rerun.

A second task in the pilot: the missing-document email

If extraction is your main task, a small second task helps you see whether the tool earns its keep in a different part of the workflow. A common one is the email that asks a client for what is missing. Give the assistant last year's document list, this year's received list and the client's name placeholder, and ask for a short message. Here is a prompt with an illustrative result.

Last year we received: employer pay statements (2), bank interest
statement, mortgage interest statement, charity receipts, childcare
invoices. This year so far: one employer pay statement, bank interest
statement. Write a friendly email asking for what is missing. Do not
guess amounts. Keep to 110 words. Use [first name] and [your name].

A typical output asks for the second pay statement, the mortgage interest statement, the charity receipts and the childcare invoices, with a polite deadline. What you would fix: the tool sometimes adds a document that was never on last year's list, such as "your pension statement", because it is common. Mark any item not on the source list as a failure of the safety rule, since an invented request confuses clients and wastes their time. If the draft keeps to the list across ten emails, count it as a pass for this small task, and note the minutes saved, usually five to eight per client.

Seasonal seats against a twelve-month contract

Seat pricing interacts with the calendar. Suppose a vendor quotes $90 per user per month on a twelve-month contract, and your practice files 300 returns a year, 255 of them (85 per cent) in four months. Three users cost $3,240 a year: $12.71 a return if the tool handles only the 255 in-season returns, or $10.80 if it handles all 300. If the vendor offers a four-month seasonal seat at $120 a month for the same three users, the cost is $1,440, or $5.65 for each of the 255 in-season returns. The same tool costs about half as much per return under the second contract. So ask for seasonal or monthly seats in writing, and run your cost-per-return sum under both contracts before you decide.

The pilot return where the tool read the wrong year

A pilot shows failures that a demo hides. Picture a client who uploads a prior-year pay statement by accident next to the current one. The tool reads both and picks the larger gross pay figure, which belongs to the wrong year. Nothing in the output flags the problem, and the preparer, trusting the extraction, enters it. The error surfaces at review because the withholding does not match. The lesson is to add a rule to your checking step: confirm the tax year printed on every source document before comparing figures. Also check whether the tool shows where each figure came from, because a number with no source reference takes far longer to verify than one with a link to the page it was read from. Record both the error and the rule you added, and include them in the decision page.

Telling your team what you found

Share the pilot results with everyone who touched it, in plain terms. A single page works: the task, the pass line, the three numbers, the safety result and your decision. If the answer is to adopt, explain which steps change and which checks stay. If you stop, explain why, so no one buys the same tool next year on a whim. Keep the pilot file, with its consent copies, vendor emails and security documents, for as long as your record-keeping rules require, because it shows how you assessed the tool before relying on it.

Edge cases that change the answer

• Introductory prices. If the pilot runs on a first-order discount or a free allowance, such as Juno's five free preparations, base the decision on the list price you will pay in year two, not on the pilot price.
• Clients with unusual returns. Run a few in the pilot, not only the easy ones. If accuracy collapses on partnership returns, you know the limit.
• Staff who dislike the tool. Ask them to log why, not just to rate it. A repeated complaint, such as "it hides where it got the number", is a design point worth reporting to the vendor.
• Review time creep. If checking takes longer each week because trust drops, the saving shrinks. Compare week one's checking minutes with week four's.

Counting setup and training time as a cost

The cost-per-return sum is incomplete until you add the hours that went into getting the tool running. Log them honestly: reading the vendor's security documents, updating the engagement letter, connecting the tool to your tax software, and training two colleagues. An illustration: 12 hours for the owner and 6 hours for each of two preparers comes to 24 hours, or $1,080 at $45 an hour. Spread across 300 returns in the first year, that adds $3.60 to every return, and the figure falls in year two. For a pilot that barely breaks even on the monthly numbers, the setup cost can decide the matter, so put it on the decision page as its own line rather than hiding it in the average.

Making the call and writing it down

At the end, put three numbers on one page: accuracy, minutes saved per return, and net cost per return. Add the safety result and the consent and WISP status. Then decide among three outcomes: adopt, extend the pilot for another 15 returns with a specific fix, or stop. Stopping is a legitimate result, and a cheap one.

If you keep the tool, run the same measures again after the busy season, because first-month results often flatter a tool. For help running a pilot like this on your own returns, I cover that in an AI implementation consultation, and the AI guide links the rest of this series. Plan the rest of your season with how small tax practices use AI through the busy season, and for a task that suits a first pilot, look at an AI tax organiser that chases missing documents.

Questions before you start a pilot

How long should the pilot run?

Four weeks is enough if you run at least 25 comparable returns or tasks through it. Shorter than that and one odd client skews the average; longer and you are paying for a tool you have not yet decided to keep.

Which task should a small practice pilot first?

Pick the repetitive task with the clearest right answer, such as extracting figures from client documents into your tax software. Avoid starting with anything that gives advice, because you cannot mark it right or wrong in the same way.

Do I need client consent before using AI on returns?

Often you do. Rules on using return information with outside software vary, and many require written client consent before data goes to a third party. Check the rules that apply to you, update your engagement letter, and ask a qualified adviser if unsure.

What if the vendor will not publish a price?

Ask for a written quote based on your return count, and ask what counts as a billable unit, what happens above the cap and whether the pilot is free. Treat a refusal to quote against your volume as a reason to pause.

Further reads

• Tax Software With AI: What a Small Business Should Check — A checklist of what to check in tax software with AI.
• Accounting Software for CPA Firms: Automation Features Compared — Practice software for CPA firms, automation compared.
• AI for Small CPA Firms: Chasing Missing Client Documents in Tax Season — Chase missing client documents in tax season.
• AI for Tax Advisory: Drafting Client Advice Memos Safely — Draft client advice memos without losing control.
• AI in Bookkeeping and Tax: The Year-End Handoff to Your Preparer — Tidy the year-end handoff from bookkeeper to preparer.
• Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare and what it cannot.
• AI Tax Prep Software for CPAs: What to Test Before You Switch — A pre-switch test for CPA firms: ten filed returns, a scoring sheet, planted errors and the checks that separate Canopy Smart Prep partners, Juno and others.
• 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.
• Intuit Accountant Suite Pricing: Is Accelerate Worth It for a Small Firm? — Core is free, Accelerate is $149 a month from 20 January 2027. Break-even sums, a four-person worked example and a plan for the free beta.
• AI-Driven Tax Software: How to Test Blue J and ChatGPT on Your Own Cases — Compare Blue J with ChatGPT using a ten-question test built from your own closed cases, a scoring sheet, a citation check and a break-even sum.
• 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.
• Tax Preparation Automation: Source Documents to Software With AI — Map W-2, 1099 and K-1 fields, set accuracy checks and handle scans, so AI extraction feeds your tax software without re-keying errors.
• Tax Services AI for a Small Practice: Intake, Chasing and Review Notes — A three-part workflow for a small tax practice: intake checklist, document-chasing emails and reviewer notes, with a named person owning each step.
• TaxDome Automations: What to Automate First in a Small Practice — Which TaxDome automations to set up first in a small practice: onboarding, document requests and reminders, in order, with what to leave manual.
• Using AI in Financial Planning: Tax Return Data Into a Client Plan — A planner's source-to-plan workflow for tax-return figures, with a checked example and a client question log.
• Can AI Draft an IRS Notice Response? What to Check Before Sending — A practitioner's routine for AI-drafted IRS notice responses: triage by type and deadline, redact first, verify every code, amount and date, then sign.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Blue J, Juno, Filed, TaxGPT and Canopy product and pricing pages, and Canopy's January 2026 announcement of its Filed integration, read in November 2026.