Lesson objective

Run an AI-assisted accounting task through seven stages, from assembling source documents to a reviewed and released deliverable, with a gate at each stage and a visible reviewer step.

Plain-language explanation

Why this matters now

Triage decides whether a task goes near AI. Once it does, the risk moves from whether to how. A task that was sensible to assist can still produce a wrong deliverable if the inputs were unverified, the checking was done against the draft instead of the source, or the reviewer saw only a polished result.

The expectations are now written down in several places. Professional guidance: the IRS Office of Professional Responsibility's Alert Issue Number 2026-19, distributed June 24, 2026, says that under Circular 230 § 10.22 practitioners must thoroughly review AI-created documents before delivery to a client or the IRS, verifying facts, citations and calculations. Its best practices include treating AI-generated text as drafts. The alert applies existing Circular 230 provisions, which bind practitioners before the IRS; it is not a new regulation.

Professional standard: the AICPA Statement on Standards for Tax Services No. 1, section 1.4, binding AICPA members in tax services from January 1, 2024, explains at 1.4.7 that a member using tools remains responsible for the completed work product, and at 1.4.8 that tools should enhance the member's understanding, "not to supplant the member's professional judgment."

Professional guidance: the Texas State Board of Public Accountancy's August 2026 newsletter article says a Texas CPA remains responsible for all work product and must review and evaluate AI output before relying on it. For attest work it states that "AI-generated analyses do not constitute audit evidence" and do not replace required procedures. That article is not a Board rule and speaks to Texas licensees, but the rules it cites bind them.

None of these sources sets out a workflow. They set the outcome a workflow must produce: a deliverable a qualified professional has checked and owns.

A question to sit with

If your reviewer received only the finished deliverable, could they tell which sentences came from the client's documents, which came from you, and which came from a tool?

What goes wrong

Starting from the tool instead of the source documents. The preparer asks the tool a question and then looks for documents that fit the answer. The direction of travel is reversed, and whatever the tool asserted becomes the frame for the work.

Feeding the tool unverified inputs. A figure typed from memory, a client-prepared schedule never agreed to the ledger, last year's rate carried forward. An error in the input comes back out as a confident, well-written error.

Checking the output against itself. Asking the tool "are these figures correct?" produces an answer, not a check. Professional guidance: the OPR alert's statement that "Practitioners cannot rely solely on AI" rules this out for tax practice.

The reviewer sees only the polish. A fluent draft with no trace of where each statement came from invites the reviewer to read for tone and flow. The review becomes an edit.

Treating an AI analysis as evidence. On an audit, an AI-produced analysis of a population is a lead to follow, not support for a conclusion. The Texas Board's statement above is explicit for Texas licensees; apply the standards that govern your engagement.

Releasing without an owner's decision. The draft leaves the firm because the deadline arrived, not because the professional who owns it decided it was ready.

The practical answer

Seven stages, each with a gate: something that must exist before the next stage starts. The workflow is this programme's recommendation. The duties behind it are cited above.

Stage Who What happens Gate: what must exist before moving on
1. Assemble sources Preparer Collect the source documents from the client file or system of record, not from memory A source index listing each document and its date
2. Verify and prepare inputs Preparer Agree figures to the ledger or tax software; confirm the data class is cleared for the tool; remove what the task does not need A verified input packet
3. Run the task Preparer Use an approved tool and, where one exists, the firm's prompt for this task; save the prompt and output A draft headed DRAFT FOR REVIEW
4. Preparer verification Preparer Trace every fact, figure and authority to a source; recompute every number; mark each item A verification grid
5. Reviewer step Named reviewer Review the draft against the sources, not against the draft's own reasoning; raise and clear review notes Review notes cleared and the reviewer's sign-off
6. Release Owner The qualified professional who owns the output decides it is ready, and signs where a signature is required The final deliverable
7. Record Preparer, confirmed by reviewer Complete the engagement AI use record entry (Lesson 4) A record entry

Stage 2 is where most errors are prevented. Everything the tool receives should already be something you would be willing to put your initials beside. If the tool will be asked to write about revenue, the revenue figures it receives come from the trial balance, not from an earlier email.

Stage 4 uses three marks. For each statement in the draft, mark it S (supported: traced to a source and, for figures, recomputed), C (corrected: wrong, now fixed from the source) or R (removed: no source exists). Anything without a mark is not ready for review.

Here is an extract from a fictional verification grid for a draft variance note to the controller of Kestrel Joinery LLC, a fictional company.

Draft statement Source Preparer Reviewer
Q2 revenue fell from $1,240,000 to $1,178,000 Trial balance, Q2 2025 and Q2 2026 S; change recomputed as $62,000, or 5.0% Agreed to trial balance
The fall reflects the end of the Marsh Street contract Controller's email of July 12, 2026 S Read email; agreed
Gross margin was 34.0% Trial balance: cost of sales $777,480 S; ($1,178,000 − $777,480) ÷ $1,178,000 = 34.0% Recomputed; agreed
Gross margin rose by three points Q2 2025 margin per trial balance: 32.0% C; the draft said "three"; corrected to two points Agreed correction
Freight costs will fall next quarter None R Confirmed removal

Where the task is attest work, the workflow does not change what counts as evidence. Required procedures are performed and documented under the standards that govern the engagement. An AI-assisted analysis may point to items for testing; it is not the test.

Remember this

  • Work from the source documents to the draft, never from the draft back to documents that fit it.
  • Verify inputs before the tool sees them; a verified input packet is the cheapest control in the workflow.
  • Mark every statement supported, corrected or removed, and recompute every figure yourself.
  • The reviewer reads the sources first, checks the marks and looks for what is missing.
  • An AI-assisted analysis is not audit evidence; required procedures still apply.

Accounting example

Harrow & Pell CPAs is a fictional firm preparing a quarterly reporting pack for the fictional client Kestrel Joinery LLC. Staff accountant Owen Maddox is asked to draft the variance note for the controller. He triages it as Route B, because the note goes to the client. At stage 1 he builds a source index: the Q2 2025 and Q2 2026 trial balances, the budget, and an email from the controller dated July 12, 2026 explaining the lost Marsh Street contract. At stage 2 he agrees the trial balance totals to the general ledger and removes the payroll detail the note does not need. At stage 3 he uses the firm's approved assistant and its variance-note prompt, pasting only the verified packet, and saves the output headed DRAFT FOR REVIEW. At stage 4 his grid shows the draft said gross margin "rose by three points". His recomputation gives 32.0% for Q2 2025 and 34.0% for Q2 2026, a two-point rise, so he marks it C and corrects it. The draft also claimed freight costs would fall next quarter. Nothing in the file says so, so he marks it R and deletes it. At stage 5 the engagement manager, Leah Pell, a licensed CPA, reads the controller's email and the trial balances before opening the draft. She recomputes the revenue change of $62,000 and the margins, agrees them, and adds a review note: the email also mentions a price increase from May that the draft omitted. Owen adds the point with the email as its source, and Leah clears the note. At stage 6 Leah approves the note for release. At stage 7 Owen completes the AI use record entry and Leah initials it.

Responsible-use warning

Never verify an AI draft by asking the same tool, or another tool, whether it is correct. A tool's confirmation is another output, not a check, and it can be confidently wrong about its own work. Verification means tracing each statement to a source document you hold and recomputing each figure with the ledger, the tax software or your own spreadsheet. If a statement has no source, it comes out, however plausible it sounds.

Human-review requirement

The named reviewer, a qualified professional, reads the key source documents before the draft, checks the preparer's verification grid against those sources, recomputes figures, and looks for what the draft left out. The owner of the output, who may be the same person, then decides whether it is released: the licensed CPA for attest work, the signing preparer for a return or tax advice, or the responsible controller for the company's books. For attest engagements the reviewer confirms required procedures were performed; AI analysis is not evidence.

Practical exercise

Use the fictional Kestrel Joinery figures in this lesson, or invent your own.

  1. Write a source index for a short fictional task, such as a variance note, listing at least four source documents with dates.
  2. Describe the verification you would perform on each input before a tool sees it.
  3. Write, or ask an approved tool to write using only your fictional packet, a five-sentence draft. Head it DRAFT FOR REVIEW.
  4. Build a verification grid with columns for draft statement, source, preparer mark and reviewer.
  5. Mark every statement S, C or R. Recompute every figure and show the working in the grid.
  6. Plant one deliberate error and one unsupported claim in a copy of the draft, and give the copy to a colleague acting as reviewer.
  7. Ask the reviewer to read the sources first and then the draft, and to record what they found and what they think is missing.
  8. Compare the reviewer's findings with your planted errors, and note anything the review missed or found that you had not planted.
  9. For each error the review missed, identify which stage of the workflow should have caught it first, and write one change to the grid or the source index that would make it easier to catch next time.
  10. Write the release decision you would expect the owner to record, naming the role of the owner and the date the deliverable would leave the firm.

Knowledge check

  1. Why is stage 2, verifying and preparing inputs, described as the cheapest control in the workflow?

    Answer guidance: Because an error in the input comes out of the tool as a confident, well-written error that is harder to spot later. Agreeing figures to the ledger or tax software and removing unneeded data before the tool sees anything prevents errors instead of hunting for them.

  2. What do the marks S, C and R mean on a verification grid?

    Answer guidance: S means supported: traced to a source and, for figures, recomputed. C means corrected: found wrong and fixed from the source. R means removed: no source exists, so the statement is deleted. An unmarked statement is not ready for review.

  3. Why should the reviewer read the key source documents before reading the draft?

    Answer guidance: So the draft cannot set the reviewer's expectations. Reading the sources first makes it easier to notice what the draft got wrong and what it left out, rather than reviewing for tone and flow.

  4. What does the Texas State Board of Public Accountancy's August 2026 article say about AI analyses in attest work, and whom does it address?

    Answer guidance: It states that AI-generated analyses do not constitute audit evidence and do not replace required procedures. It is a newsletter article, not a Board rule, addressed to Texas licensees; other states' boards must be checked separately.

Sources and further reading