Tax Workpaper Prep With AI: Tie the Return Back to Source Documents

Yes, AI can draft a tie-out schedule: a table that lists each figure on a tax return beside the source document, page and amount that supports it. It cannot vouch for the figures. The reviewer still opens the source for each item that matters, and anything with no source gets resolved by a person.

The useful way to think about it is that AI does the clerical matching and the reviewer does the judging. A schedule that is 90% right but looks 100% right is worse than no schedule, so the checks that stop a tidy table from hiding a wrong number matter more than the drafting does.

Follow me on Instagram@sagnikteaches

What a tie-out schedule is, and what it is not

A tie-out connects the return to the paper. Each row says: this line of the return shows this amount; it came from this document, on this page; the document says this amount; the difference is nil or explained. The reviewer reads down the "difference" column and spends time only where it is not nil.

Connect on LinkedInSagnik Bhattacharya

It is not an audit or compilation workpaper. Those document different work under different standards, with their own sign-offs. A tax tie-out is a preparer-and-reviewer tool inside one return, and keeping that boundary clear matters because a model will happily write language that sounds like assurance. Do not let it.

Subscribe on YouTube@sagnikteaches

The columns that make the schedule useful

Notice who fills what. The return amount comes from the tax software, ideally an export, because that is the figure under test and it should never pass through a language model. The difference is a formula. The model's job is limited to finding the source and reading its figure, and the status column is human only.

A worked example: a three-schedule return for a yoga studio owner

An illustration, with invented numbers. Picture a sole trader who runs a yoga studio, with a part-time job, a rented room and a retirement contribution. The source pack has nine documents. The preparer exports the return's figures to a sheet and gives the AI, one document at a time, a prompt to extract the document type, the period, and the labelled amounts.

Read the attached statement (names and account numbers already
redacted). Return a table with: document type, period covered,
each labelled amount exactly as printed, and the page it appears on.
If an amount is unclear or cut off, write UNREADABLE and do not
estimate. Do not add figures together. Do not comment on tax
treatment.

A sample result (illustrative) for a payment processor's annual summary lists gross card receipts of 61,480, refunds of 1,250 and fees of 1,830, page 1, with "net payout 58,400" on page 2. The preparer sees at once the cross-check: 61,480 minus 1,250 minus 1,830 is 58,400, so the document is internally consistent. That is a useful second test and one you can ask the AI to run: "check whether the labelled amounts on this statement add up, and show your arithmetic". Then you verify the arithmetic yourself, because that is where a model slips.

The schedule for the business income line then reads: return shows 61,480 gross receipts less 1,250 refunds; source says the same; difference nil. Rent: return shows 9,600, the landlord's invoices list twelve payments of 800; difference nil. Instructor payments: return shows 14,200, and the payment records sum to 14,700. Difference of 500, status open, note "one payment may be a reimbursement". That single open item is what the reviewer came for, and the whole point of the process is that it surfaced in minutes.

Laid out as the reviewer receives it, the filled-in schedule for those three lines plus the two awkward ones looks like this (illustrative):

The reviewer reads the Status column first. Three rows say Agreed, so they go into the sample pool; the Open and No source rows are opened every time. The Difference column is a formula (amount on return minus amount on source), which is why the instructor line shows -500 rather than a figure the model typed.

Prepare the source pack so the AI can read it

1. One file per document. A merged scan of forty pages defeats page references. Split and name each file with year, document type and issuer.
2. Redact first. Strip names, identity numbers and account numbers before a file goes to a general chatbot, following the redaction method for returns. If your practice platform has AI built in under a data agreement, use that instead.
3. Rotate and straighten. Sideways pages and phone photos taken at an angle produce the worst misreads.
4. Log the pack. A simple index of the documents received is the first thing the tie-out checks against: if a document is not on the index, it cannot be a source.

This is where the intake work pays off. A practice that has already built a document checklist, as in the small-practice intake and chasing workflow, has the index half done.

The extraction prompts, by document type

One generic prompt is fine for a clean statement. For anything with a layout quirk, tailor it. Three that earn their keep:

Wage and income statements

Ask for the employer or payer, the box labels and the amounts, and tell the model to leave box labels in the document's own wording. A common failure is the model "helpfully" renaming boxes into its own categories, so that a reviewer later cannot find the box. Another is reading a corrected statement and the original as two statements. Prompt: "if the document says corrected or reissued, say so at the top".

Bank and card statements

Do not ask the AI to categorise a year of transactions and then rely on the totals. Ask for opening balance, closing balance and the named annual totals the statement itself prints. For a small business with 3,000 transactions, a model can drop or double a row without any sign. If you do need categorised transactions, treat the output as a first sort that a person reconciles to the closing balance.

Mortgage, property and invoice documents

Invoices are the hardest because every one has a different layout. Ask for supplier, date, description, net, tax and total, and tell the model to return the description verbatim. A repair invoice that says "replace damaged tiles" is a different treatment question to "new roof", and the preparer needs the original words, not a summary.

The matching step: AI proposes, the sheet checks

Once you have extracted tables for each document, you need to match them to the return lines. There are two workable routes. In route one, the preparer pastes the extracted tables and the exported return figures into a spreadsheet and does the matching by hand, using the AI only for extraction. In route two, the AI proposes matches ("this 14,700 looks like the instructor payments line") and the preparer accepts or rejects each one. Route two is faster and riskier, because a wrong match with a near-identical figure is easy to accept.

The guard is to ask for the match and the reason, and to reject any match where the amounts are not equal unless there is a stated explanation. Sample output (illustrative):

Return line: Business income, 61,480
Proposed source: Processor annual summary, page 1, "Gross card receipts" 61,480
Reason: amounts equal; label matches business receipts
Confidence: high

Return line: Home office, 1,900
Proposed source: none found
Reason: no utility or rent document in the pack that relates to a home office
Confidence: n/a

The second block is the valuable one. A tool that can say "none found" is useful; one that always finds something is dangerous. Test this deliberately: before you trust a prompt, remove a document from a test pack and check that the AI reports the gap rather than inventing a match. If it invents, rewrite the prompt with a stricter instruction and test again.

What a reviewer checks, in order

This is the reviewer's checklist, and it works whether the schedule was drafted by a junior, a spreadsheet or an AI.

1. The index is complete. Compare the documents on the schedule with the checklist and the index. A source that does not appear in the pack is a red flag.
2. Return figures match the software. Compare the "amount on return" column with the final return, not an earlier draft. A schedule built on version two of a return is a common error after late changes.
3. Every difference is explained. A difference of any size needs a status of explained or open, never blank.
4. Sample the nil differences. Pick three nil lines, open the source and read the figure. If any is wrong, the extraction cannot be trusted; widen the sample to everything for that document type.
5. Check the years and the people. Confirm period, tax year and the right taxpayer on each source. Similar documents from the previous year are the usual culprit.
6. Look for what is missing. Read the return for income and deductions that have no row on the schedule. A tie-out proves what is there, never what was left out.
7. Initial and date. The reviewer's sign-off is a record of their own work.

Step six is the one most reviewers skip and the one AI is least able to help with. The schedule only knows about the documents it was given. A missing rental income statement does not show up as a difference, because there is no row for it. That is why the checklist from intake matters: the missing-item test is against what you asked for, not what arrived.

A second illustration: where a wrong tie-out nearly slipped through

A realistic mistake, described generally. A print shop's owner supplies two supplier statements with similar totals: one for the tax year, one for the following year, because the supplier sent a combined PDF. The AI extracts the figure from the later one, which happens to be within 3% of the return's figure. The AI proposes the match with "high confidence". The reviewer, following step five, checks the period on the statement and finds it is the wrong year. The figure was plausible, the match was confident and the document was wrong. Only the period check caught it.

The fix was a prompt change (ask for the statement period first, and stop if it does not match the stated tax year) and a rule: any proposed match where the difference is under 5% but not nil must show the period alongside it. Plausible near-matches are the failure to fear, not wild errors.

How to size a sample and set a materiality line

Reviewers cannot open every document on every return, and for a simple return they should not have to. Set a rule you can defend and write it down:

• All lines above a materiality threshold you choose per return type, for example 5% of the client's total income or a fixed figure such as 500, whichever is lower.
• Every line with a status of open, explained or no source.
• A random sample of one in five of the remaining nil-difference lines.
• Everything, on any return where the first sampled line was wrong.

The threshold is a judgement, so pick it with your reviewer and revisit it after the first season. Say a tutoring centre owner's return has 34 lines on the schedule. Eight are above the threshold, three have open status, and of the remaining 23 you sample five. The reviewer opens sixteen sources instead of 34: a saving of over half, and the sixteen are the ones that carry the risk.

Three more drafts to try on a test return

Before you use any of this on a live file, run it on a return you finished last season and already know is right. Three quick tests, each with a known answer, tell you how the tool behaves.

The reissued statement test

Put the original and the corrected wage statement in the pack. A good prompt makes the model say "document B is marked corrected; document A is superseded" and use only B. If it lists both amounts as income, add the line "if two documents cover the same payer and period, list them separately and say which is later" and run it again. Say the original shows 41,200 and the correction 41,950: a model that adds them gets 83,150 and a schedule that fails to tie by exactly the wrong amount, which is easy to spot, but only if the original return is open next to it.

The missing-page test

Remove page two of a four-page brokerage summary. Ask for the annual totals. A reliable tool says the totals are not on the pages supplied. An unreliable one returns plausible totals. Whatever your tool does, you have learned it before a client file depended on it.

The reviewer-note draft

Once the schedule is done, ask the model to turn open and explained rows into a short note for the reviewer, using only the status and note columns. For the yoga studio example it should give you two lines: the 500 difference on instructor payments, noted as a possible reimbursement awaiting the client's answer, and the home office deduction with no source, awaiting a rent or utility document. If it adds a recommendation ("this deduction is probably not allowed"), delete it. Tax treatment is the preparer's call and sits outside the tie-out.

Keep the schedule as the file's record, not a chat history

A tie-out is only worth having if it can be re-read next year, or when a notice arrives. Save the final schedule as a spreadsheet or PDF in the client's file, with the source pack index, the date, and the names of preparer and reviewer. Do not rely on a chat transcript as the record: chats can be deleted, they are hard to search, and they hold the working, not the conclusion.

It also helps to save the prompt you used, the model's name and plan, and a one-line note on what was redacted. If you ever have to explain how a figure got onto a return, "prepared by a person, reviewed by a person, schedule drafted with a tool under these rules" is a far better position than an undocumented process. Where you are unsure what your professional body expects around tools and record-keeping, ask it or an adviser; do not infer it from a chatbot.

Where AI tie-outs fail, and how each failure shows up

• Transposed digits. The model reads 61,840 for 61,480. It shows up as a difference you cannot explain. Look at the document, not at the model's second attempt.
• Rounded figures. Reading "about 14.7k" from a chart or summary. Insist on printed amounts only.
• Merged lines. Two statement rows read as one, so the total is right and the detail wrong. Check row counts against the source for any line you sample.
• A plausible invention. The model returns a figure for a document that was not in the pack. The "none found" test above catches this.
• Silent redaction damage. Redaction that removes a digit or a decimal point. Redact by replacing identity fields only, and look at the redacted file once.
• Over-trust. The schedule looks finished, so everyone stops looking. The sampling rule, written down, is the defence.

For the stage before this one, where extraction feeds the return itself, see the tutorial on extract, review and sign off, and for what an AI reviewer can and cannot catch compared with your software's own diagnostics, that comparison is the natural next read. Both assume the same principle: AI drafts, a person decides, and the file shows who did what.

Questions reviewers ask about AI-built tie-outs

Is a tie-out the same as an audit workpaper?

No. A tax tie-out links each figure on a return to the document that supports it, so a reviewer can check it quickly. Audit and compilation workpapers follow different professional standards and document different work. Only the return-to-source schedule is in scope here.

Should the AI be allowed to total the schedule?

Let it list and match, but total with a spreadsheet formula and check the total against the return yourself. Language models can misadd, and a total that looks right is exactly the thing a tired reviewer skips.

How much of the tie-out should a reviewer re-check?

Start with every line for the first dozen returns, so you learn how it fails. Then sample: all figures above a materiality line you set, all items with no source, and a random one in five of the rest.

What if the source document is a poor scan?

Mark the line as unreadable rather than letting the tool guess. Ask the client for a better copy. A confident figure read from a blurred page is the most dangerous output in the process.

Further reads

• ChatGPT Prompts for Tax Preparers: Client Emails and Checklists — Prompts for client emails and checklists.
• AI Tax Research Tools Compared: Citations, Cost and Checks — Checking AI tax research for citations.
• AI for Small CPA Firms: Chasing Missing Client Documents in Tax Season — Get the documents in before you tie anything out.
• Free AI for Tax Professionals: Where the Free Tier Stops at Client Data — Where free AI tiers stop for client data.
• Section 7216 Consent for AI Tools: What Clients Must Sign — Client consent before AI tools see return data.
• Using AI in Financial Planning: Tax Return Data Into a Client Plan — Using the same return data in a client plan.
• 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.
• Can AI Help Prepare an R&D Tax Credit Claim? Narratives and Costs — How to use AI to draft R&D claim narratives from project notes and cross-check cost schedules, with eligibility left to the specialist.
• Can AI Prepare Workpapers for Review and Compilation Engagements? — How a small practice can use AI to prepare lead sheets, tick-mark legends and review notes while keeping the file labelled, traceable and signed by people.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: practice-software and AI provider documentation on document handling and data use; the author's workflow design. Figures and documents in examples are invented and illustrative.