AI for Doing Tax Accounting: What to Hand Over and What to Review

AI can reliably help a small tax practice extract fields from routine documents, flag missing records, assemble draft workpapers and draft plain-language client letters when the inputs and review rules are controlled. A qualified reviewer must verify source figures, tax treatment, calculations, assumptions and every client-facing conclusion before filing or sending advice.

The useful boundary is the point at which a copied fact becomes a professional judgement. Extracting a date from a document is one task; deciding whether an expense is allowable is another. Design the workflow so each AI output carries a source reference and an explicit status: extracted, checked, queried or approved.

Follow me on Instagram@sagnikteaches

Start with a task map, not a chatbot licence

List the actual steps in one recurring engagement: requesting records, receiving files, naming them, reading figures, reconciling totals, preparing workpapers, researching an unusual treatment, reviewing the draft and writing to the client. Mark the person who currently owns each step. A four-person practice may discover that most delay comes from asking for missing schedules, while the tax calculation itself is short. Automating the wrong step then creates a faster queue of incomplete files.

Connect on LinkedInSagnik Bhattacharya

Sort tasks into three groups. The first contains clerical transformations with an original document to compare, such as transcribing a gross figure. The second contains analysis that can be drafted but requires professional testing, such as a variance explanation. The third contains decisions and representations: choosing a tax treatment, approving a return, answering a client-specific question or signing a letter. AI can prepare material for the second and third groups, but it cannot own their conclusion.

Subscribe on YouTube@sagnikteaches

This distinction is consistent with professional guidance on AI in tax work, which stresses due care, confidential data and professional judgement. Two of its practical points translate directly into a small-practice rule: treat AI output as if a less experienced junior colleague had prepared it, and assume that pasting client data into a publicly available AI tool is likely to breach client confidentiality. It also says members remain accountable for the work whether or not AI was involved, and that cited legislation or case law should be confirmed to exist. Your firm's particular duties depend on its professional body, contracts and applicable rules, so a partner or qualified adviser should approve the policy before live client material enters a new system.

Do not label a process safe solely because the software is sold to accountants. The firm's procedure matters more than a feature list: who may upload, what the tool retains, which files it sees, how errors are captured and who approves the finished work. For a broader map of accounting risks, use the accountant AI error-check checklist; this tutorial concentrates on tax engagement hand-offs.

Extract documents into an evidence-backed queue

Choose one document family for the pilot, such as annual interest statements or invoices with predictable fields. Define the output columns before prompting: document ID, client ID, period, field name, extracted value, currency, page reference, confidence flag and reviewer note. A field without a page reference should be treated as missing. This makes spot checks possible and avoids forcing the reviewer to search a folder to understand where a figure came from.

A prompt for a supervised extraction test on synthetic or approved documents might read:

Extract these fields from each attached interest statement:
payer, account reference (last four characters only), period
start, period end, gross interest, tax withheld, currency.
Return one row per field with file name and page number.
If a field is absent or unreadable, write "not found".
Do not calculate totals, classify income or suggest a tax
treatment. Flag any date that could be read two ways.

An illustrative output for one statement: "Gross interest $2,418.70, file S-7, page 2; tax withheld: not found; period end 04/05, ambiguous format, flagged." The reviewer opens page 2, finds the withholding figure in a footnote box the model skipped, and confirms the period from the statement header. Two fixes follow: add "check footnotes and boxed summaries" to the instruction, and keep the ambiguous-date flag because it worked. What you are checking is not only whether each value is right but whether the model said "not found" instead of guessing.

On that statement the footnote box shows withholding of $241.87. An extraction that returns the $2,418.70 gross figure but omits withholding is incomplete even though its main number is right. The reviewer compares both figures to the original, checks that the period is the one requested and records the omitted field. A corrected row is then promoted to the workpaper; the first draft remains in the audit trail.

Another illustrative example: two scanned invoices both show $1,260, but one is a duplicate copy with a different filename. A simple extraction may produce two accurate rows and an inaccurate total of $2,520. The preparer needs a duplicate rule using supplier, invoice number, date and amount, followed by a visual check when a match appears. AI should flag probable duplicates, never silently delete one: repeated amounts can also be genuine instalments.

A third case is a statement where the date sits next to a reference number. The model reads 08/11 as the wrong period, and its output passes a numeric reasonableness test. Require the source crop or page link alongside the date and ask the reviewer to check any ambiguous format. If the statement period cannot be established, mark it queried and request confirmation from the client rather than choosing a convenient interpretation.

Keep extraction separate from classification. A tool that reads a payment of $900 accurately has not established whether it is income, a reimbursement, a capital item or a personal transfer. The label belongs in a review column. If the source and the client's description conflict, retain both and ask for evidence. The workflow should never let an AI-assigned category flow straight into a submitted figure.

Build workpapers that expose assumptions

A workpaper is a traceable record of how an amount was assembled and checked. AI is useful for creating a first draft of its narrative, matching headings to an approved template, carrying forward opening items and listing unexplained differences. It should display arithmetic inputs separately from explanatory prose. Reviewers need to see the actual formula or reconciliation, rather than accept a sentence claiming that figures agree.

Suppose an illustrative client ledger has receipts of $184,500, credit notes of $3,200 and a bank total of $181,300. AI might say the ledger and bank reconcile because $184,500 minus $3,200 equals $181,300. The arithmetic is correct, but it has not established that every receipt cleared the bank or that credit notes belong in that reconciliation. A reviewer traces the population, checks cut-off and reviews the underlying entries. The difference between a mathematical match and an accounting conclusion must be visible.

Ask the tool to produce an exception list rather than a green tick. A useful row reads: “Ledger receipt R-108, $1,950, dated after period end; bank line not found; confirm cut-off.” An unhelpful row reads: “Everything appears consistent.” The first gives the preparer a specific action. The second hides the uncertainty. When the model cannot locate a source, require “not found” rather than a plausible number.

For spreadsheet-heavy engagements, the accounting spreadsheet checks tutorial provides task-level controls. In tax work, add a column for the reviewer’s source of authority on each consequential treatment. A computed total can be recomputed; a treatment must be tested against facts and current guidance. If the facts change after the first draft, invalidate the earlier conclusion and rerun its review.

Worked papers should have a version date, a preparer, a reviewer and unresolved items. It is tempting to close an exception by asking AI for a more confident explanation. That is not evidence. Close it with a client document, a corrected entry, a verified source or an explicit professional judgement. Record why the chosen explanation displaced the alternative, particularly for material or unusual items.

A numbered hand-off for one small practice

Consider an illustrative four-person tax practice processing 30 similar engagements over a month. Each engagement produces roughly 12 source documents and 18 fields to check. That gives 540 candidate fields, before considering unusual items. The practice chooses a single document type that accounts for 120 of those fields, then runs a controlled pilot on 10 engagements. The numbers describe the test design, not a promised saving.

1. Prepare the sample. Use synthetic documents first, then approved live records under the firm's data policy. Include clear scans, faint scans, duplicates and a wrong-period document. Assign a reference ID to each source.
2. Define the output. The tool returns one row per field with file and page, extracted value, uncertainty and a blank reviewer decision. It is told not to infer missing values.
3. Review all pilot fields. A preparer compares every row with its source. A second person checks the exceptions and a small sample of clean rows. The firm records error types rather than one aggregate “accuracy” figure.
4. Prepare the workpaper. Only checked fields are imported. The AI drafts a variance note and a list of missing records. The qualified reviewer independently checks material totals and treatments.
5. Draft communication. AI converts approved points into a client request or explanation. A human verifies that it does not state an unapproved tax result and sends it through the normal channel.

Imagine the pilot returns 120 fields. Review finds two wrong periods, one duplicated statement and four omitted withholding entries. That is seven issues with different causes, not a single accuracy percentage. The team changes the prompt and intake rule, then repeats a fresh sample. If repeated faint scans still lead to omissions, route that document type for manual entry. A successful pilot is one where the review cost and error pattern are understood, not where an output merely looks polished.

A simple time ledger makes the decision honest. Record minutes spent on intake, AI setup, source checking, exception resolution, workpaper preparation and final review. If a manual extraction takes 6 minutes and AI extraction plus verification takes 8, the pilot has found a poor candidate. If a new client letter saves drafting time but increases correction time, compare the complete cycle. Do not remove final review from the comparison; it is part of the service.

Give a practice manager a stop rule before scale-up: any fabricated source reference, unflagged material omission or client data placed in an unapproved tool pauses the workflow. A typo in a draft internal note can be corrected without the same escalation. This is a risk-based distinction: the severity and detectability of the error determine what happens next.

Draft letters from approved facts only

Client letters are a good use of language generation once the facts are settled. Provide a short, approved fact sheet: purpose, confirmed figures, outstanding evidence, action requested and deadline chosen by the practice. Ask for a concise draft and an “unsupported claims” list. Never ask the model to decide the client's liability from a pile of raw files and then put that result into a friendly email.

An illustrative before-and-after shows the difference. Raw instruction: “Tell the client the return is ready and they owe $4,800.” Approved facts: “Draft calculation shows $4,800 subject to reviewer approval; one investment statement is missing.” Safe draft: “We have prepared a draft calculation, but need the outstanding investment statement before we can finalise it. Please send it by [date].” The first instruction would turn a provisional figure into a final assertion; the second preserves its status.

Another case is a client asking whether a new purchase can be claimed. A model may produce a smooth answer from an old rule or ignore how the item is used. The team should acknowledge the query, collect the purchase date, invoice and business use, and assign the technical answer to a qualified person. The plain-English tax explanation tutorial is useful after the conclusion has been checked; it is not a substitute for reaching that conclusion.

Review tone as well as figures. A letter that says “you failed to provide” may be needlessly accusatory when the file was never requested. A letter that says “we have verified” may overstate work that has only been drafted. The reviewer should check the verb behind every assurance: prepared, compared, confirmed, approved and filed describe different states. The client should understand which one applies.

Keep tax decisions and filings under qualified control

Tax planning and unusual treatments require current authority and an accurate understanding of the client's facts. AI can propose questions for a research memo: Which period applies? Which document proves ownership? What alternative treatment may apply? What facts would change the answer? A qualified professional then obtains the primary source, reads it in context and records the conclusion. Citations produced by a chatbot should be treated as search leads until opened and verified.

For example, a client reports a $14,000 disposal and asks if the whole amount is taxable. The model does not know the original cost, transaction costs, holding period or relevant rules. It can produce a missing-information list; it cannot responsibly decide the treatment. A reviewer may use the list to ask for purchase records and an explanation of the asset, then calculate using verified evidence. The figures are illustrative and do not imply a tax outcome.

Keep the filing act separate from preparation. A final return or signed advice carries the practice's responsibility. Require a named approver to compare the final output with the approved workpaper, confirm unresolved queries are cleared and authorise the submission. If a software integration can submit automatically, disable that path until the firm has a documented approval control and a way to retrieve what was actually filed.

Also avoid treating one qualified reviewer as a rubber stamp. Review depth should rise with uncertainty, value, novelty and potential client harm. A familiar interest amount can be traced quickly; a new relief claim or disputed classification needs more evidence and possibly a second technical opinion. Staff should know how to escalate rather than ask AI to generate an apparently settled answer.

Data controls before the first real record

Tax files can include addresses, identifiers, income, family details and business records. Before adopting a tool, document what categories it receives, who can access them, whether a vendor uses them for model improvement, where outputs are retained and how deletion works. Check the practice's confidentiality duties, engagement language, client instructions and relevant data-protection requirements. A provider's general statement that it is “secure” is not enough to establish fit for a particular client file.

Use role-based access. The person drafting a client letter may need the approved fact sheet, not the entire document folder. A pilot operator may need synthetic files, not live client names. Keep the prompt, input version, output, review note and final approved document together under the practice's normal retention process. Do not create an untracked second copy of client records in personal accounts or email attachments.

A filled-in example of a review record could read: “Engagement T-041; source S-7 page 2; field gross interest $2,418.70; AI value $2,418.70; checked by preparer on 12 November; exception: withholding omitted; reviewer requested corrected extraction; final amount approved after source comparison.” The point is the chain from source to decision. Replace the example identifiers and dates with your actual audit convention.

Before the practice expands the pilot, review the contract and the technical settings with someone responsible for information security and professional compliance. If terms cannot be established, use a tool already approved for that data or keep the task manual. A data-safe workflow may be slightly slower than a public chatbot demonstration, but it is one the practice can explain to a client or reviewer.

Check success by error type and review time

After two or three pilot cycles, compare like with like: the same document types, similar complexity and the same review standard. Track omitted fields, wrong values, wrong period, duplicate documents, invented references, unsupported treatment and letter corrections separately. A fall in total minutes is useful only if the serious errors are found before release. Count near misses, because they reveal where the control worked and where it was lucky.

Set a sampling plan for any later routine use, but keep full review for high-risk outputs. For example, a firm could check every extracted field for a new document type, then consider a smaller secondary sample once the first reviewer and exception rules have proved reliable. It should still trace every material figure used in a final return and review every client-facing conclusion. The exact thresholds are a firm policy decision based on its work and obligations.

If the practice uses AI to send record requests, log whether the request was correct, whether it reached the right client and whether the response contained the missing evidence. A request that saves drafting time but repeatedly asks for documents already present may damage trust. The small-practice automation tutorial deals with workflow triggers; the tax reviewer remains responsible for deciding when an engagement is actually complete.

The practical test is simple: could another qualified person reconstruct the final figure and understand who approved the judgement? If the answer depends on trusting an opaque AI summary, redesign the hand-off. AI earns its place by making the source trail easier to inspect and the open questions easier to see.

Questions before putting AI into tax work

Can AI prepare a return without a reviewer?

It can populate a draft from records, but the preparer should reconcile source totals, investigate missing documents and confirm the applicable rules before a qualified reviewer approves the return. A plausible calculation is not evidence that the facts or treatment are right. The firm must also follow its engagement terms and professional obligations.

Should we upload full client records to a general chatbot?

Only after the firm has checked the tool's data handling, access, retention and contractual terms against its client confidentiality obligations. For an initial test, use synthetic documents with the same layout problems. If a vendor cannot answer who may access the files and how they are deleted, leave identifiable client records out.

What is a useful first pilot for a small practice?

Try extracting a narrow set of figures from one recurring document type into a review queue. Keep a source reference for each field, check every pilot result against the original and log corrections. Once the error pattern is understood, consider workpaper drafting; do not begin with automatic tax conclusions or filing.

Further reads

• AI Readiness Checklist for Accountants, Solicitors, Consultants — Assess the practice's starting controls.
• AI for Chartered Accountants: What to Tell Clients About Your AI Use — Plan the client explanation of AI use.
• AI Engagement Letters for Accountants: Faster Drafts, Clear Scope — Set scope and review expectations in client terms.
• How Accountants Can Use ChatGPT Effectively: Review Notes and Memos — Develop review notes and technical memos.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ICAEW, Topical guidance covering the application of PCRT to the ethical use of artificial intelligence tools (updated 19 January 2026). Checked November 2026.