AI Tax Return Prep for Small Practices: Extract, Review, Sign Off

Use AI for three jobs: extracting figures from source documents into a worksheet, running tie-out checks that compare the return with those documents and last year's return, and drafting review notes. The preparer still approves every figure that lands on the return, a reviewer signs off against a written checklist, and the client has consented to the tools used.

Speed comes from the first two jobs, not from skipping review. A draft that staff review instead of re-keying is faster only if the review is real: each figure traceable to a page, and a sign-off that records who checked what. Vendors quote big gains, such as the "Prep 4x Faster" headline on Canopy's Smart Prep page, which gives no method or baseline. Time a pilot on your own returns and let that number decide.

Follow me on Instagram@sagnikteaches

What Canopy's Smart Prep and TaxDome AI do, and don't

Both products are live, but they do different jobs, and neither does the preparer's. Canopy's Smart Prep page describes documents being synced from Canopy to third-party AI partners, which extract, classify and map the source documents and generate a draft return. Status, errors and completed workpapers come back into Canopy's work lists. The page names three partners, Filed, Byron and Juno, and says each needs its own subscription, with pricing depending on the partner and the return volume. It says Smart Prep sits alongside tax software such as Drake, CCH Axcess, ProConnect and UltraTax instead of replacing it, and that your team reviews and approves the return before it is finalised.

Connect on LinkedInSagnik Bhattacharya

Canopy's AI FAQ separately says that Canopy AI does not directly prepare tax returns and works on documents, tasks and client communication, which fits a picture of the partner doing the extraction. On security, the FAQ says Canopy AI operates within Canopy's SOC 2 Type II framework with 256-bit encryption. The Smart Prep page says data sent to partners is encrypted in transit and at rest and that partners must meet SOC 2 standards. It does not include a client consent statement.

Subscribe on YouTube@sagnikteaches

TaxDome AI launched in August 2024 to tag and rename client uploads, and it now covers more of the season while still stopping short of preparation. TaxDome's help centre lists checklist generation from the prior-year return, matching each upload to its checklist section, renaming and tagging files by document type, and help with delivering the finished return: grouping its pages, filling in the tax summary and placing signature fields for staff to check before it goes out. It also lists an API and an MCP connection for linking an assistant such as ChatGPT or Claude to your TaxDome data. Some features come with the subscription and others need an AI add-on, availability depends on where the firm is based, and nothing runs until a firm owner or admin switches it on. On data, the same page says the features run on enterprise services from OpenAI, AWS and Google, that TaxDome does not train AI on personal data, and that providers keep nothing beyond processing, except that Google may hold some data for up to 30 days to detect abuse. None of the features it lists drafts the return itself.

Before any client document goes into a vendor's tool, put these questions to the vendor and keep the answers on file:

1. Where are documents processed, and by which partners or sub-processors?
2. Is client data used to train models, by you or by a partner, and is that written in the contract and not only on a web page?
3. How long are documents and extracted data kept, and how is deletion confirmed?
4. Does every extracted figure carry a page reference I can check?
5. Can I export an audit trail showing who changed what?
6. What happens when a document is corrected or reissued after extraction?
7. What is the price per return at my volume, and are there minimums or seasonal limits?

The practical rule is to treat whatever either tool produces as unreviewed data entry by a fast but fallible junior. The tool saves typing. It does not take over the looking, the checking or the signature. The sections below set out where each of those happens, using one invented example: a client with a pay statement, two bank interest statements and a consolidated brokerage statement that arrived twice, once as an original and once as a corrected copy. The documents reach this stage through the process in the tutorial on building an organiser that chases missing documents.

Extraction: every figure gets a page reference

The rule that makes AI extraction safe is that no figure exists without a source. Ask for the value, the page and the exact text it came from, and tell the assistant never to calculate or infer. Then a preparer can check any figure in seconds by turning to the page, instead of hunting.

Extract the fields below from the statement text that follows.
For every value return: Field | Value | Page | Exact text it came from.
If a value is not on the document, write NOT FOUND. Never calculate,
round, infer or combine figures. If the document says it is corrected
or supersedes another, say so on the first line.

FIELDS: ordinary dividends; interest; short-term proceeds; short-term
cost basis; long-term proceeds; long-term cost basis; total proceeds
and total cost basis as printed.

STATEMENT: [paste text]

Here is an illustrative result for the corrected brokerage statement, with what a careful preparer notices:

One transposed pair of digits changes the short-term gain from 1,160.00 to 530.00, a difference of 630.00 that nothing on screen would flag. A glance at page 7 does catch it, which is why the page reference exists. The next section shows a second net for the same error.

Set the depth of checking by document risk, not by how confident the tool looks. Brokerage statements, business and rental income, anything running past a couple of pages and any corrected or reissued document should have every extracted figure checked against its page. Single-figure statements, such as a lender's year-end interest total, need a quick look at the referenced page next to each number. In every tier the preparer verifies each figure that is entered on the return. The model-reported confidence in an output is a hint for choosing what to read first, not a probability.

Layout is the biggest variable. Clean digital PDFs usually extract better than phone photos of crumpled pages, multi-column brokerage layouts can swap columns, and image-only or password-protected files may need preparing first. Keep a running list of senders and document types where your tool has missed, starting from the pilot below, and send those to full-page verification or manual entry by default. For handwritten notes, a client's own rental ledger and anything unusual, skip extraction and key it by hand. A tool that confidently reads an unfamiliar layout is more dangerous than one that visibly fails.

The tutorial on moving source documents into tax software with AI covers the hand-off in more depth.

Tie-out checks: the arithmetic belongs to a spreadsheet

A tie-out is a check that figures agree with each other or with a source. Models are unreliable at adding up long columns, so use them to read documents and write explanations, and use a spreadsheet for every sum and difference. This table lists the checks worth running on a typical individual return:

The first row rescues the transposed digits. The statement prints a total cost basis of 36,390.00. The extracted components sum to 17,920.00 plus 19,100.00, which is 37,020.00, so the check fails by 630.00 and sends the preparer to page 7. The fourth row handles the other trap in the running example: the original statement and the corrected copy both sit in the folder, and an extraction run on the first one would use the superseded basis. The rule is simple. The tracker keeps both files, marks the older one superseded, and extraction runs only on the latest.

The variance check is the one that finds missing documents. Here is the running client with an illustrative rule: flag a line when it moves by more than 25% and by more than 200.

Flag column (B = last year, C = this year):
=IF(B2=0, IF(C2<>0,"NEW LINE",""),
   IF(AND(ABS(C2-B2)>200, ABS((C2-B2)/B2)>0.25),"EXPLAIN",""))

The interest line is the useful one. A drop from 355 to 41 usually means a statement from the second bank has not arrived, not that the client earned less. The formula's thresholds are mine, for illustration, so set yours from last season's returns. The tutorial on tying the return back to source documents goes further on the workpaper side, and AI return review against software diagnostics shows what each layer catches, so you do not rely on one.

Review notes the AI drafts and the preparer decides

Once the tie-outs have run, an assistant can turn the results into review notes. Give it the tie-out sheet and the list of flags, and ask for three things: open items, items needing the preparer's judgement and questions for the client. Never give it the full return to "review" and accept the verdict.

Below are the tie-out results and variance flags for one return, with
names and ID numbers removed. Write review notes in three parts:
1. Open items (a check failed or a document is missing)
2. Judgement items for the preparer (anything that depends on the client's
   circumstances or a professional decision)
3. Questions for the client, in plain English
Use only the data provided. If something looks wrong but you cannot tell,
say so. Do not state a conclusion about tax treatment.

An illustrative extract of the result for the running client: open items would include "short-term cost basis fails to tie by 630.00; check page 7 of the corrected statement" and "interest down 88.5%; second bank's statement not on file". The questions for the client would include "Do you have receipts for the additional 2,700 of giving?" Judgement items would be left for the preparer, such as how the business expense increase should be treated. The last line of the prompt matters. A notes draft that announces how something is taxed is exactly what the preparer must not outsource.

The reviewer reads the evidence, not just the memo. A memo is a convenient index to the tie-outs and the extraction sheet, and a reviewer who reads only the memo has reviewed the AI's opinion.

What stays with the preparer, whatever the tool does

Extraction and tie-outs cover what is on the page. A large share of a return's risk sits in what is not, and that part stays human:

• Eligibility and status questions, such as whether a dependant qualifies or which filing status applies. They depend on facts only the client and the preparer know.
• Classification judgements: personal or business use, repair or improvement, revenue or capital.
• One-off transactions, including a home sale, an inheritance, cancelled debt or income from an unusual source, where the document does not tell the story.
• Choices the client can make that change the result, which need explaining to the client in plain words.
• The interview: the questions the documents cannot answer, because nobody uploads a document for something they forgot.
• Due diligence on credits and other claims, which the preparer must document.
• The signature, authorisation and delivery of the return.

Put these on the sign-off sheet as items the preparer ticks personally. It keeps the division of labour explicit: the tool prepares evidence, and the person decides.

The sign-off sheet: who checked what, and when

Sign-off is where the practice proves the preparer's review happened. A short sheet, kept with the return file, records it. If you are the only preparer, review the next morning against the printed tie-out sheet instead of straight after finishing, because fresh eyes catch more, and consider a swap with another small practice for the final read.

RETURN: [client code], [year]     PREPARER: [initials]     REVIEWER: [initials]
CONSENT TO AI TOOLS ON FILE: yes, [date] / no, manual route used
TOOLS USED AND FOR WHAT: [extraction tool] for brokerage and pay statements;
  [assistant, business plan] for review notes
DOCUMENT CHECKLIST: 14 of 14 received or confirmed not applicable
EXTRACTION: every figure entered on the return traced to its page by the
  preparer: yes
TIE-OUTS RUN: components to totals; withholding; wages; latest version only;
  year and payer; identity data twice; refund account twice
VARIANCES OVER THRESHOLD: 3 flagged, 3 explained (notes attached)
CORRECTED OR SUPERSEDED DOCUMENTS: 1 found, latest version used
AI FLAGS RAISED: 5, all read by the preparer, 2 changed the return
REVIEWER ACTIONS: [list]
SIGNED: [initials] [date]

Two lines on that sheet matter beyond the checks. "Tools used and for what" gives you an audit trail if a client or a regulator asks later. "AI flags raised and read" proves a person looked at every flag, which is the difference between a tool that assists and a tool that decides. Keep the extraction sheet, the tie-out sheet and the notes with the return as the working record, in whatever form your file retention requires.

Consent and engagement wording before documents go in

Many professional and privacy rules require the client's consent before return information is disclosed to, or used by, someone outside the practice. Whether an AI vendor counts depends on the rules that apply to you, how the tool processes the data and the vendor's terms. That is a question for your professional body or an adviser, and the tutorial on consent wording for AI tools goes through what clients may need to sign. Neither Canopy's Smart Prep page nor TaxDome's AI help article addresses client consent, so treat it as the practice's job.

Whatever the legal answer, three habits are cheap. Tell clients which tools handle their documents, including partners, since Canopy's page says documents are synced to separate third-party partners. Get the consent in writing and file it with the return. Offer a manual route for any client who declines. Sample wording for an adviser to review:

[Practice name] uses AI-assisted tools to read your tax documents and
prepare a draft of your return. The tools are provided by [vendor names],
who process your documents on our behalf under written security terms.
A qualified preparer at our practice checks every figure and signs off
every return. If you prefer, we will prepare your return without these
tools, which may take longer.

Consent: [client name], [date]

If you paste material into a general assistant instead of a practice platform, use a business plan. ChatGPT Business is $25 per user per month on monthly billing, or $20 on annual billing with a two-seat minimum, and Claude Team is $20 on annual billing and $25 monthly. Neither trains on business content by default. Add the tools to your written security plan, and delete working files and conversations at the end of the season.

A ten-return pilot before you trust the timings

Pick ten returns of mixed difficulty before the busy season. Prepare each the usual way and again with the assisted workflow, with a different person doing the second pass, and log the minutes by stage. The tutorial on running a pilot and measuring cost per return covers the design in more detail. These are illustrative results:

Notice that the assisted route spends more on review, not less. The saving sits in the first row. Log every miss as well, because the misses decide whether the process is safe. In the same illustrative pilot the preparer's checks found 14 AI misses across 8 of the 10 returns:

Counting by who caught each one gives 8 caught by the preparer's page check, 5 by tie-outs and 1 by the reviewer. The one that reached the reviewer is the number to study: it means the preparer's checklist had no "right year" line, so add one and re-run. The pilot passes when no miss reaches the client. It does not pass because the average time looks good.

Add a seeded-error test as well. Take two returns from the pilot and, before the review, plant two errors in each: change one digit in a figure and swap a corrected statement for the original. Then see whether the review catches them. If it does not, the checklist has a gap, which is far better found in a pilot than in a client's refund. Repeat the test once a season, because attention fades when a tool is usually right and neat output is easy to trust.

The sum is short. Thirty minutes saved on a return is worth $30 at a loaded preparer cost of $60 an hour, which is an assumption to replace with yours. Across 300 returns it is 150 hours. Any extraction subscription that costs less than $30 a return leaves a margin before you count pilot and training time. Ask each vendor for a quote per return at your volume, then compare it with your own pilot figure instead of the vendor's headline.

Questions about AI in return preparation

Can AI prepare a tax return without a human review?

No tool should be the last pair of eyes. Treat anything an AI extracts or drafts as unreviewed data entry. The preparer traces each figure to its source, runs tie-out checks, and a reviewer signs off against a written checklist. Check the professional rules that apply to your practice, since responsibility for the return stays with the preparer and the firm.

What should a reviewer check on an AI-extracted figure?

Four things: that the figure matches the page it came from, that components add up to the printed total, that the document is the latest version rather than a superseded one, and that it belongs to the right payer and the right year. Those four cover the misses a pilot is most likely to find.

Do clients have to consent to AI tools?

Many professional and privacy rules require the client's consent before their return information is disclosed to, or used by, someone outside the practice, and an AI vendor may count. The answer depends on the rules that apply to you and on how the tool processes data, so ask your professional body or an adviser, and get written consent with a manual alternative.

How do I know whether AI really saves time?

Run a pilot of about ten returns both ways, time each stage and log every miss by type. Convert the minutes saved into a value per return and compare it with the tool's cost per return. A vendor's headline speed-up is a claim about someone else's returns, while the pilot measures yours.

Further reads

• Canopy AI: Smart Tax Prep, Document Renaming and Beta Features — What Canopy's AI modules and beta features actually do.
• AI Tax Prep Software for CPAs: What to Test Before You Switch — What to test before switching tax-prep software.
• ChatGPT Prompts for Tax Preparers: Client Emails and Checklists — Prompt wording for client emails and review checklists.
• WISP for Tax Preparers: Add Your AI Tools Without Gaps — Adding AI tools to your written security plan.
• Circular 230 and AI: Due Diligence When a Tool Drafts the Work — Due diligence duties when a tool drafts part of the work.
• How Small Tax Practices Use AI Through the Busy Season — Where AI fits across a whole busy season.
• Canopy vs TaxDome for Tax-Heavy Practices: Which Fits the Season? — The two practice platforms compared for a busy season.
• Accounting Software for CPA Firms: Automation Features Compared — TaxDome, Karbon and Canopy compared on real automation, with list prices, plan gating and what is still beta, plus a five-seat cost sum.
• Tax Preparation AI Software: TaxDome, Canopy and ChatGPT Compared — TaxDome, Canopy and ChatGPT sit in different product categories. Compare their AI, their cost models and the consent step before return data goes in.
• 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.
• Client Tax Organiser: Build One With AI and Send It by Portal — A conditional tax organiser example with client questions, portal instructions, staged reminders and a completeness review.
• Can AI Check a VAT Return Before You File? Coding Error Checks — Export your quarter's transactions, run VAT-specific checks for rate, zero-rated and reverse-charge coding errors, and sign off the return yourself.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Canopy Smart Prep, AI and Smart Intake pages; TaxDome help centre article on TaxDome AI; CPA Practice Advisor coverage of the TaxDome AI launch (August 2024); ChatGPT Business and Claude plan pages (checked October to November 2026).