Can AI Prepare Workpapers for Review and Compilation Engagements?

Yes, as a drafting aid, if the file stays defensible. AI can build lead sheets from a trial balance, draft a tick-mark legend and suggest review notes. Each output must be labelled as AI-prepared until a named preparer has checked it against source records, and a named reviewer has signed. The sign-offs remain human.

A workpaper is evidence that somebody did the work, and a reviewer can only rely on it if they know who did what. A neat schedule that nobody has traced to the ledger weakens a file, however good it looks. AI can clearly produce the pages. The design question is how to make the review trail show exactly what the tool did and what a person verified.

Follow me on Instagram@sagnikteaches

What review and compilation files actually need

The two engagement types sit at different levels. A compilation has the practice assembling financial information for the client without giving assurance, so the file needs to show the information was compiled from the client's records, the practice understood the client's business well enough, and obvious errors were queried. A review engagement gives limited assurance, built on enquiries and analytical procedures, so the file also needs the analytical work, the explanations obtained and the conclusions. The international standards are ISRS 4410 (Revised) for compilations and ISRE 2400 (Revised) for reviews, and many practices report under a national equivalent. Under the international compilation standard the practitioner is not required to verify the completeness or accuracy of what management provides, but must understand the business and the basis of accounting well enough to do the job, and must ask for additional or corrected information when the records look incomplete or wrong. Your reporting standards set the exact documentation requirements, and they differ by framework, so treat what follows as a pattern to adapt, not a standard to follow.

Connect on LinkedInSagnik Bhattacharya

Common ground across both: a trial balance or ledger extract, lead sheets that group accounts into financial statement lines, supporting schedules, a record of queries and the client's answers, and review notes showing the second pair of eyes. That list is where AI helps.

Subscribe on YouTube@sagnikteaches

The labelling rule that keeps the file honest

Before any prompt, decide how AI-assisted pages are marked. A minimum scheme that works for a small practice:

• A header on each AI-drafted page: "AI-prepared draft, not yet reviewed".
• The preparer replaces it, once they have checked the page against source, with their initials and date, and a line "Tool output verified to source".
• The prompt and the version of the input file are saved with the workpaper reference, so the work can be reproduced.
• The reviewer's sign-off is separate from the preparer's, and it is a different person.

Nothing in the file should show only the AI header at the time of sign-off. A page still marked "AI-prepared draft" at the end of the engagement is a finding for your own quality review. A small practice can add this to its engagement checklist as one line, and it costs almost nothing.

Lead sheets: the best first use

A lead sheet summarises accounts that roll up to one line of the financial statements, with the prior year, current year, adjustments and references to supporting schedules. Building one from a trial balance is repetitive and error-prone by hand, which makes it a good candidate. The input is a trial balance export in a spreadsheet, with account numbers and names, and a mapping of accounts to statement lines that your practice owns.

I will paste a trial balance (account number, name, prior year, current year) and a mapping
table (account number to financial statement line). Using only these, produce:
1. One lead sheet per statement line: list the accounts, prior year, current year, movement
   in value and in percent
2. A check row on each sheet: sum of accounts versus the sum on the trial balance
3. A list of accounts in the trial balance that are not in the mapping
4. Any line where movement exceeds 15% and no explanation is given
Do not change any figure. Do not net anything. Output as tables.
[trial balance and mapping pasted here, client name removed]

An illustrative output for a small car repair garage: a lead sheet for "Trade receivables" listing four accounts, totalling 48,200 current year against 41,900 prior year, movement 6,300 or 15.0%, with a flag that the movement meets the threshold. And a list of two unmapped accounts, such as a new suspense account. What you fix: the arithmetic and the mapping. Models can slip on sums, so recompute every total in the spreadsheet, and treat unmapped accounts as a question for the client, not something to be quietly assigned to a line.

Tying the lead sheet to the trial balance

The preparer's check is simple and non-negotiable: the total of each lead sheet agrees to the trial balance, and the sum of all lead sheets equals the trial balance total. Do it by formula in the spreadsheet rather than by reading the assistant's statement that it agrees. If the check fails by a rounding amount, find why instead of adjusting it. Rounding differences in a lead sheet often hide a missing account.

Tick-mark legends and their misuse

A legend defines the symbols used to show what was done: agreed to bank statement, agreed to invoice, recalculated, traced to prior year file, and so on. AI can draft a consistent legend for the practice and apply a house style across engagements. The danger is different. A tick mark claims that a procedure was performed, so a tool that drops tick marks onto a schedule is generating false evidence.

The rule is firm: AI may draft the legend and the schedule layout, but only a person places a tick mark, and only after doing the thing. If a template arrives with marks pre-filled, strip them before use. An illustrative legend for a compilation file:

✓ Agreed to client's bank statement
∑ Footed (added) by preparer
T Traced to prior year file
R Recalculated by preparer
Q Queried with client; see query log ref
N No exception noted on compilation procedures performed

Have the assistant propose the legend, and then reduce it. Small practices do better with six symbols that everyone understands than 20 that people use inconsistently. Keep each symbol's definition to one line, and check that the wording describes an action, not a conclusion. "Agreed to bank statement" is an action. "Verified" is a conclusion without a procedure behind it.

Analytical review notes for review engagements

In a review engagement, analytical procedures compare the current period with expectations, such as prior year, budget or industry patterns, and the practice enquires about differences. AI is useful for the first sweep: lining up year-on-year changes and ratios across the lead sheets and suggesting where an enquiry is warranted. It is not useful for supplying the explanations, which must come from the client and be corroborated.

From these lead sheets, list every line where the change from prior year exceeds 10% or
5,000, whichever is lower. For each, write a neutral enquiry for the client
("Revenue increased by 18%. Please explain the main reasons and provide supporting detail.").
Do NOT suggest explanations. Do not conclude anything.
[lead sheet summary pasted here]

A sample output, illustrative: nine enquiries, each neutral and specific. What you fix: any enquiry that leads the witness ("Is this due to the new contract?") and any line the assistant skipped. Then send the enquiries, record the answers in the query log, and judge whether each explanation is plausible and consistent with other information you hold. The conclusion is the reviewer's.

Review notes that survive scrutiny

A reviewer's notes show what was looked at and what was resolved. AI can draft notes from a reviewer's own bullet points, and can check a set of notes for open items. It should not generate notes that imply a review that did not take place.

Here are my review points in my own words. Tidy them into numbered review notes:
reference, observation, action required, owner. Keep my meaning exactly.
After the list, show any point I left without an action.
[reviewer's bullets pasted here]

Compare a before and after:

Before (reviewer's bullet): receivables up a lot, check aged list, why no provision
After (tidied note):        Ref C2. Trade receivables up 15% on prior year. Obtain aged list at
                            year end and the client's reasoning for no provision. Owner: preparer.
                            Status: open.

Notes must close. At the end of the engagement, every note carries a status and, where cleared, a reference to the evidence that cleared it. A prompt that lists open items from the note log takes seconds to run and catches the one note everyone forgot.

A worked file: a compilation for a six-person furniture maker

Take an illustrative annual compilation for a furniture maker with 60 trial-balance accounts, one senior preparer, and a partner as reviewer. Budgeted time is 14 hours without AI.

Step one (about 1 hour). The preparer exports the trial balance and uses the practice's mapping table. The assistant builds 11 lead sheets and flags two unmapped accounts: "Workshop deposit" and a new asset account. The preparer asks the client about both. The deposit is a customer deposit, so it maps to deferred income. The new asset is a purchased machine, and the preparer requests the invoice.

Step two (about 3 hours). The preparer ticks the sheets by hand: bank balances agreed to the bank statement, loan balance agreed to the lender's statement, fixed asset additions agreed to invoices. The assistant is not involved except to draft the query list: four questions about stock and one about a director's loan.

Step three (about 1 hour). Analytical comparison: the assistant lists the six lines moving by more than 10%. The preparer records the client's answers.

Step four (about 3 hours). The preparer completes the file, replacing every "AI-prepared draft" header with initials after checking the page. A check that all 11 headers are gone is part of the file-closing list.

Step five (about 2 hours). The partner reviews, re-performs a sample of five balances from source to lead sheet, and writes four review notes, which the assistant tidies. The preparer clears them.

The total is about 10 hours against 14 (illustrative). The saving sits in the lead sheets, query drafting and note tidying. The ticking and the review did not get quicker, and should not.

Supporting schedules: bank reconciliations and fixed assets

Below the lead sheets sit the supporting schedules, and two of them suit a supervised first draft: the bank reconciliation layout and the fixed-asset movement schedule. For the reconciliation, the preparer enters the statement balance, the ledger balance, and the list of unpresented items. The assistant can build the layout and check that the reconciling items explain the difference, though the figures themselves come from source documents and the preparer's tick marks show they were agreed.

For fixed assets, paste the opening balances, additions and disposals with dates and ask for a movement schedule: cost, additions, disposals, depreciation charge and closing net book value, with a check row. Ask it to list any asset whose depreciation charge differs from the stated policy by more than a rounding amount. In the furniture maker example, the schedule might flag that a new CNC router bought in month 10 was depreciated for a full year. The preparer confirms with the invoice date, corrects the charge, and notes the correction in the query log. The model found the question; the person checked the invoice and made the correction.

The query log, filled in

A query log records what you asked the client and what they said. It is one of the most useful pages in a review file, and one of the easiest for an assistant to keep tidy. A filled-in extract (illustrative):

Use the assistant to maintain the table from the preparer's notes and to produce the list of open items on demand. The column that matters is "Evidence". A response from the client with nothing behind it is a statement, not evidence, and for a review engagement the reviewer should ask whether it is consistent with everything else in the file.

The file-closing checklist

A short closing checklist is where the labelling rule becomes routine. Add these lines to the practice's existing list:

• No page carries an "AI-prepared draft" header; each has initials and date.
• Each AI-assisted workpaper has its prompt and input file saved against the reference.
• Lead sheets agree to the trial balance by formula, and the check cell reads zero.
• Every tick mark on the schedules was placed by a person, and the legend matches the symbols used.
• The query log has no open items, or each open item is explained and approved by the reviewer.
• The reviewer's sample is recorded: which balances, what was traced and the result.
• Preparer and reviewer are different people, each initialled and dated.

Print it with the file, or keep it as a final tab. It takes five minutes, and it is the page a quality reviewer will look at first. Where the practice does a quality-management check on closed files, this list also tells the inspector exactly where to look.

What the reviewer does differently

When part of a file was drafted by a tool, the reviewer's sample becomes the control that matters most. Three practical changes:

1. Re-perform, don't reread. Choose a handful of balances, trace each from source document to lead sheet, and see that the numbers and the tick marks are real.
2. Check the labels. No page should reach sign-off with an unreplaced AI header, and every prompt and input should be stored with the reference.
3. Widen when wrong. If the sample finds an error, extend it. An error in a tool's output may repeat across similar accounts.

Document the sample in the review note: which balances, what was traced, the result. That one paragraph turns the sign-off from a signature into evidence. See tying a tax return to source documents for the same discipline applied to tax work.

Where it goes wrong: four short cases

The helpful rounding. A draft lead sheet shows a total of 112,000 where the trial balance accounts add to 111,962. The assistant rounded the sub-totals. The check row catches it, and the preparer finds that the difference is a missing account.

The pre-filled tick. A template generated for a pest-control firm file arrives with "agreed to bank statement" marks on every cash line. The reviewer finds the bank statements are not in the file. Rule: only a person places a tick.

The invented explanation. Asked to "explain" a 22% revenue rise, a draft cites new customers. The client later says a one-off contract caused it. The explanation had no source, and it shaped how the file read. Rule: enquiries only, never explanations.

The lost trail. A preparer iterates ten prompts in a chat window and pastes the final version into the file. Nobody can reproduce it. Rule: save the final prompt, the input and the output with the reference.

Confidentiality and where to run it

Trial balances and ledgers are client confidential information. Use tools covered by a data agreement and no training on your content (OpenAI and Anthropic both state that for their business tiers; on a consumer plan the training setting has to be checked and switched off), remove client names, and tell clients in your engagement terms that AI-assisted tools may be used under supervision. Many practices prefer AI features inside software they already use for workpapers or a secure extension of their spreadsheet tool, since the data agreement is already in place. Best AI tools for small accounting firms compares options, and the acceptable use policy template for a professional firm covers the written rules.

For request lists and workpaper notes in an audit setting, see ChatGPT prompts for auditors, and using AI to spot errors in client books for the bookkeeping side. The AI hub lists the wider series.

A last practical point on version control: keep one master copy of each workpaper and name versions plainly, with the date and the preparer's initials. When a tool regenerates a lead sheet after a late adjustment, the old version should be kept and marked superseded rather than overwritten, so the file shows how the numbers moved and why.

How to check the process works

Pick three closed files each quarter and run an internal inspection. Is every AI-assisted page either initialled or still labelled, with the label explained? Are prompts and inputs saved? Did the reviewer's sample appear in the review notes? Do the lead sheets agree to the trial balance by formula? Count the exceptions. If the same type recurs, fix the template, not the people. And ask your own team whether the tool is saving time on the right pages; if preparers spend more time correcting lead sheets than building them, simplify the mapping or switch the step off. Check your professional body's current guidance on AI and documentation too, because expectations are changing, and treat this tutorial as a starting design, not a standard.

Questions practices ask about AI in engagement files

Can AI-prepared workpapers be relied on in a review engagement?

Only after a person has checked them against source records and signed as preparer. The workpaper is the practice's evidence, so its reliability comes from the checking, not the tool. Label AI-prepared items until a named person has verified them and initialled the page.

Does the file need to say AI was used?

Your quality-management policy should say how AI-assisted work is recorded. A short label on the workpaper and a line in the engagement file is a sensible minimum. Check your professional body's current guidance, because expectations are being updated as the technology spreads.

Can I upload a client's trial balance to a chatbot?

Only if the engagement terms and the tool allow it. Use a business plan with no training on your content, minimise identifiers, and check the client's consent. Many practices prefer AI features inside their existing workpaper software, where a data agreement is already in place.

What should the reviewer do differently with AI-prepared files?

Re-perform a sample. Pick a few balances and trace them from source to lead sheet yourself, and check that every tick mark was actually carried out. If the sample shows errors, widen it. The reviewer's sign-off is the control, so it should not be a skim.

Will AI weaken the audit trail?

It can if outputs are pasted in without labels or sources. Keep the prompt, the input file and the output version with the workpaper reference, and date every step. A reviewer should be able to see who prepared, what the AI did, and who checked.

Further reads

• How Small Tax Practices Use AI Through the Busy Season — Running a small practice through its busiest weeks with AI.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven AI mistakes that put client trust at risk.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check the firm is ready before adding AI to client work.
• How to Chase Missing Client Records With AI Before Deadlines — Chasing missing client records before deadlines.
• How Much Does AI Cost a Small Accounting Firm? — What AI costs a small accounting firm.
• What an AI Implementation Looks Like in a Small Accounting Firm — What an AI rollout looks like in a small practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: IFAC material on ISRS 4410 (Revised) compilation engagements and on review engagements as limited assurance; OpenAI business-data page and Anthropic privacy centre on training defaults. All read 4 November 2026. Your own reporting standards and quality-management policy govern the detail.