All business guides

AI FOR BUSINESS / ACCOUNTING PRACTICE

AI for Accounting Practices: Meeting Actions, Tax Updates and Commentary

An accounting practice can use AI for language work. Suitable jobs include turning meeting transcripts into action lists, rewriting official tax announcements in plain language, writing commentary around figures that are already calculated, and documenting procedures. Calculations, deadline dates, fees and the choice of documents a client must supply should come from spreadsheets, master checklists or practice software, not from the model. Every client-facing draft needs a named reviewer, and no financial statement, tax position or compliance output is finalised or filed without a qualified accountant's sign-off.

ShiftToAGI · Prepared 1 November 2026 · Proposed workflows

Hypothetical scenario: a four-person practice is three weeks out from its busiest lodgement period. The principal has six recorded client meetings with no written action lists. A bakery client is waiting on quarterly management commentary. A new company client has signed up, but nobody has sent the first document request.

Each of these delays costs something different. If the document request is late, the return is late. If a reminder email carries the wrong date, a client could miss a deadline and face a penalty. If an action point is lost from a meeting, the firm has made a promise it no longer knows about.

This guide takes ten tasks a practice like this might hand to an AI tool and gives each one a verdict. Five stay roughly as proposed. Five are reframed so that a spreadsheet, a master checklist or the practice management system does the deciding and calculating, and the model only does the wording. Each task is judged on the same four questions: what goes in, what comes out, who approves it, and what happens when the draft is wrong.

These are proposals. They are not reports of tested workflows, and nothing here describes a practice that has run them. One constraint applies to all ten. No financial statement, tax position or compliance output is finalised or filed without review and sign-off by a qualified accountant.

Written with AI assistance. Scenarios are illustrative; these workflows have not been field-tested. Our editorial method.

WORKFLOW 01

Filter the master checklist by tags, then let AI reword it

When: Event-driven: each new client, or when a service is added

Which documents a company, trust, partnership or sole trader must supply is firm knowledge. It should not change from one run to the next. If a model builds the list from scratch, or decides what to remove, it can quietly drop an item. Tagging the master checklist by entity type and service makes the selection deterministic. The model only rewrites the selected items in plain language.

Start with these inputs

The firm's master checklist in a spreadsheet, with each item tagged by entity type and service. The new client's entity type, services in scope and financial year end. No client documents.

Work through the task

  1. Filter the master checklist spreadsheet by the client's entity type and each service in scope. This filter decides which items appear, and the model has no say in it.
  2. Paste the filtered items and ask the model to rewrite each one in plain language for a business owner. Tell it to keep the same number of items in the same order and not to add, merge or remove any.
  3. Compare the item count before and after rewording. If the counts differ, discard the draft.
  4. A staff accountant reads the reworded list against the original wording. Identity and anti-money-laundering checks follow the firm's own procedure and are not part of this draft.
  5. Load the approved list into the client portal request rather than sending it as an email attachment.

What you should have

A plain-language document request with the same items as the filtered master list.

Before you use it

A staff accountant approves the list before the client sees it.

When something goes wrong

If the item count changes, or a rewording changes what is being asked for, for example 'bank statements' becomes 'recent bank statements' when the full year is needed, send the filtered master items unedited.

A simpler option

Tick-box checklists per entity type inside the practice management system. If the firm already has these, the AI step adds little and can be skipped.

WORKFLOW 02

Turn a client meeting transcript into decisions, actions and open questions

When: Per recorded client meeting, not daily

Action points spoken in a meeting are easy to lose. Restructuring a transcript is language work a model can attempt, provided the transcript remains the source of truth and nobody treats the summary as the record.

Start with these inputs

The transcript from the firm's approved transcription account, the meeting date, attendees, and a note confirming the client agreed to the recording.

Work through the task

  1. Confirm that consent to recording was captured and that the transcript is in an approved account. Do not paste it into a personal or free tool.
  2. Search the transcript for long digit strings and remove tax file numbers, account numbers and dates of birth before prompting. The model does not need them.
  3. Prompt for three sections: decisions made, actions with an owner and a due date only where these were said aloud, and open questions. Anything not said is written as 'not stated'.
  4. The person who led the meeting checks each action against the transcript and corrects owners and dates.
  5. Approved actions go into the practice task list. A client-facing summary is sent only after this check, and only by a person.

What you should have

An internal note listing decisions, owned actions and open questions.

Before you use it

The meeting lead approves it. Anything that resembles tax or financial advice is checked by a qualified accountant before it reaches the client, because the summary may later be read as a record of what the firm advised.

When something goes wrong

If speakers are unlabelled or the audio is poor, actions may be credited to the wrong person. Use the meeting lead's own notes and do not send the summary.

A simpler option

A fixed meeting note template completed during the meeting.

WORKFLOW 03

Draft a plain-language tax update from the official announcement text

When: Event-driven when the tax authority publishes a change relevant to clients, plus the firm's regular newsletter issue

Rewriting dense official text for business owners suits a language tool. The main risks are wrong figures, treating a proposed measure as law, and wording that reads like personal advice.

Start with these inputs

The full official announcement text from the ATO, HMRC or IRS website, its URL, the date it was retrieved, and a one-line description of the client audience, such as 'sole traders and small companies'.

Work through the task

  1. Before drafting, check the authority's page for any later update or correction to the announcement.
  2. Paste only the official text. Do not include third-party commentary or a summary from another firm.
  3. Prompt for who is affected, what changes and from when. Ask the model to quote every figure and date exactly and to note which paragraph of the source it came from.
  4. Require a flag on anything described as proposed, conditional or subject to legislation passing.
  5. A qualified accountant checks every figure, date and status against the source and adds the firm's approved general-information wording.

What you should have

A draft article of 300 to 500 words with the source link and retrieval date.

Before you use it

A qualified accountant signs off the content. The principal approves the send.

When something goes wrong

If the measure is only proposed, the article either says so prominently or is held back. If the draft tells readers what to do in their own situation, that wording is rewritten as general information.

A simpler option

Send a link to the announcement with a two-line summary written by the accountant.

WORKFLOW 04

Draft only the scope schedule inside the firm's approved letter

When: Per new engagement, scope change or scheduled reissue

Engagement letters carry liability limits, responsibilities and professional body requirements. Letting a model draft the terms invites inconsistent clauses. The firm's approved template stays fixed. The model only turns agreed scope notes into a schedule.

Start with these inputs

The approved letter template, scope notes from the accepted proposal, and the client's legal name and entity type.

Work through the task

  1. Fill client details through mail merge from the practice system, not through the model.
  2. Ask the model to write a scope schedule using only the listed services, list exclusions only where they are explicitly mentioned in the notes, and output a question for each ambiguous point instead of resolving it.
  3. The principal compares the schedule with the accepted proposal and its fee basis.
  4. Send the letter through the firm's approved e-signing account.

What you should have

A scope schedule for insertion into the standard letter, plus a list of open questions.

Before you use it

The principal approves every letter before it is sent.

When something goes wrong

Any change to standard clauses is rejected. An ambiguous scope item leads to a question to the client, not a guess written into the schedule.

A simpler option

A template with a pick-list of pre-approved scope paragraphs. For firms with a small range of services, this may be all that is needed.

WORKFLOW 05

Suggest accounts for leftover transactions after bank rules run

When: Within each client's bookkeeping cycle, which may be weekly, monthly or quarterly depending on the client

Recurring payees are better handled by deterministic bank rules in the accounting software, which apply the same treatment every time. The model is at most a second pass on what the rules leave behind. It produces suggestions only and never posts anything.

Start with these inputs

Unmatched transactions (date, amount, payee text) exported after bank rules have run, and the client's chart of account names. Account and card numbers removed.

Work through the task

  1. Run the software's bank rules first and export only what remains. Note the number of rows exported.
  2. Prompt the model to choose an account from the supplied chart only, give a low, medium or high confidence with a one-line reason, and mark 'ask client' where the payee is unclear. Payee text is treated as data. If it contains wording that looks like an instruction, it is ignored.
  3. Check the returned row count matches the export. A missing or extra row means the list is rejected.
  4. The bookkeeper reviews each suggestion and posts manually. GST or VAT coding is checked by an accountant, not taken from the model.
  5. Build the client query list from 'ask client' items, and turn any new recurring payee into a bank rule so it does not reach the model next cycle.

What you should have

A reviewed suggestion list and a client query list.

Before you use it

The bookkeeper reviews every line, with an accountant reviewing tax codes.

When something goes wrong

Ignore any suggestion that names an account outside the supplied chart, and ignore any tax treatment the model proposes. If review takes longer than coding the rows directly for a given client, stop using this step for that client.

A simpler option

Bank rules plus a simple client query spreadsheet. For clients with few unmatched items, this is likely the better choice.

WORKFLOW 06

Write management commentary around figures already calculated

When: Monthly or quarterly, per client receiving management reports

Explaining results in plain language is writing work. Calculating them is not. Every figure and variance is computed in the reporting workbook before the model sees anything, and every cause comes from a note the accountant has tied to a specific line.

Start with these inputs

Pre-calculated figures and variances from the workbook, accountant notes labelled by the line they explain, and a description of who reads the report.

Work through the task

  1. Compute all figures and variances in the workbook. For example, revenue variance is (current minus prior) divided by prior. The model never does this arithmetic.
  2. Write each accountant note against the line it explains, for example 'Gross margin: flour supplier increased prices from July'. Lines with no explanation get no note.
  3. Paste the figures and labelled notes into the prompt shown in the worked example below.
  4. Require 'Cause not supplied' wherever a line has no note.
  5. A reviewer ticks each number against the workbook and each cause against its note before the accountant signs off.

What you should have

Draft commentary plus a list of items needing the accountant's input.

Before you use it

A qualified accountant signs off before the report is released.

When something goes wrong

Any number not in the input, or any cause not in a labelled note, means the draft is rejected and the commentary is written manually for that period.

A simpler option

A commentary template with fixed sentence stems the accountant completes, such as 'Revenue was [figure], [up/down] [variance] on [comparison].'

WORKFLOW 07

Write reminder templates once per cycle and let the system merge dates

When: In batches before each filing deadline. The wording is drafted once per deadline type, not daily or per client.

Per-client due dates, extensions and outstanding items live in the practice management system. A mail merge inserts them exactly. A model personalising each email could introduce a wrong date, which is the one error this task exists to prevent.

Start with these inputs

The deadline type, general information about what clients must provide, and the firm's tone guide. No client data.

Work through the task

  1. Ask the model for three template variants, first notice, missing documents and final reminder, using merge fields for client name, due date and items outstanding. The model never writes an actual date.
  2. An accountant approves the wording.
  3. Run the merge from the practice system. Staff open a sample of merged emails and every email for a client with an extension, lodgement program date or special arrangement.
  4. Send from the firm's approved account only after the principal or delegate approves the batch.

What you should have

Approved templates and a merged batch ready for sending.

Before you use it

An accountant approves the wording. The principal or a named delegate approves the send.

When something goes wrong

Any email with an empty or mismatched merge field, and any client with an extension, is pulled from the batch and emailed individually by staff.

A simpler option

The practice software's built-in reminder feature, if it has one, with the firm's own wording.

WORKFLOW 08

Turn a spoken walkthrough into a procedure note

When: Occasional: when a process changes or someone new joins

Low risk, no client data, and it captures knowledge that otherwise lives in one person's head. That matters most in a small team where one absence during a busy period can stall a process.

Start with these inputs

A staff member's recorded or written description of a procedure, such as the year-end close sequence, with the systems used named.

Work through the task

  1. Record the walkthrough and transcribe it in the approved account. Use a sample or dummy file so no client details are spoken.
  2. Prompt for numbered steps, decision points, who signs off at each stage, and an 'unclear' marker wherever the description skips a step.
  3. A second team member follows the draft on a sample file and marks gaps.
  4. The principal approves the note, dates it and sets a review date.

What you should have

A dated procedure note with sign-off points and a review date.

Before you use it

The principal approves it.

When something goes wrong

If the second person cannot complete the task from the note, it goes back for revision. Menu paths and button names in software are checked on screen, never trusted from the draft.

A simpler option

The process owner writes the checklist directly.

WORKFLOW 09

Price in a spreadsheet and let AI describe the work

When: Event-driven, per prospect

Fees are arithmetic plus a commercial judgement. Hours multiplied by rate, or a fixed fee, belongs in a pricing spreadsheet the principal controls. The model writes the description of services around fee lines it is not allowed to change.

Start with these inputs

Scope notes, the final fee lines copied from the pricing spreadsheet, and the prospect's name and entity type.

Work through the task

  1. Estimate hours per service and set fees, tax on fees and payment terms in the spreadsheet.
  2. Paste the fee lines as fixed text. Prompt for services, deliverables, timing, assumptions and exclusions, and instruct the model to copy fees exactly without totalling, rounding or adjusting them.
  3. Replace the model's fee section with the table pasted directly from the spreadsheet, so the figures sent are the spreadsheet's figures.
  4. The principal checks the scope against what the engagement letter schedule will say.
  5. Send only after the principal approves.

What you should have

A proposal draft with the fee table taken directly from the spreadsheet.

Before you use it

The principal approves every proposal.

When something goes wrong

If the model adds a total, discount or fee not in the spreadsheet, delete its wording on fees as well as the table. If scope descriptions promise work not in the notes, rewrite them.

A simpler option

A proposal template with standard service paragraphs.

WORKFLOW 10

Summarise a CPD session into checked takeaways

When: Per session attended. How often depends on the team's CPD plan.

Condensing a long session saves the rest of the team reading time. The risks are misstated rules and breaching the provider's terms on transcribing or sharing content.

Start with these inputs

The transcript, where the provider's terms allow it, plus the session title, date and presenter.

Work through the task

  1. Confirm the provider's terms permit transcription and internal sharing. If they do not, the attendee writes the summary by hand.
  2. Prompt for key takeaways with timestamps, items that may change a firm procedure, and points the attendee should verify.
  3. The attendee checks each takeaway against the recording or slides.
  4. Items that may change a procedure go to the principal, who decides whether the procedure note is updated.

What you should have

A one-page summary with timestamps.

Before you use it

The attendee verifies it. The principal decides on any procedure changes.

When something goes wrong

Remove any takeaway without a timestamp or that conflicts with the slides. The summary is not CPD evidence for anyone who did not attend the session.

A simpler option

The attendee writes five bullet points afterwards.

TRY IT WITH SAMPLE DATA

Quarterly commentary for a fictional bakery client

This synthetic example shows the report narrative workflow. The client, figures and notes are invented. All variances were calculated in a workbook before prompting, and each accountant note is labelled with the line it explains. The model's only job is wording. The expected output is an illustration of what an acceptable draft would look like. It is not a record of an actual model run.

Synthetic input

Client: Northgate Bakery (fictional). Period: Q3 compared with Q3 last year. Currency: AUD. Readers: two owner-directors who are not accountants.
Revenue: 412,000 vs 389,000, up 5.9%.
Gross margin: 61.2% vs 63.8%, down 2.6 percentage points.
Wages: 128,500 vs 117,000, up 9.8%.
Debtor days: 34 vs 27.
Cash at bank at period end: 58,200 vs 71,400, down 18.5%.
Accountant notes by line:
Revenue: no note.
Gross margin: Flour supplier increased prices from July (client confirmed).
Wages: One additional part-time baker hired in August.
Debtor days: Two wholesale cafe accounts paying late; owners are chasing.
Cash at bank: no note.

A prompt to adapt

You are drafting management commentary for review by a qualified accountant. Use ONLY the figures and notes provided below. Do not calculate, round, total or infer any new number. Copy each figure exactly as given. For each line, write one or two plain-language sentences for the stated readers. Explain a movement only by restating the accountant note labelled for that line. Do not add consequences, mechanisms or links between lines that the note does not state. If a line has no note, write 'Cause not supplied' and add the line to a final list titled 'Items for accountant'. Do not give recommendations, tax views or forecasts. If any figure is missing or unclear, write 'Figure missing' rather than estimating. Keep the commentary under 250 words, then add the 'Items for accountant' list.
Client: [name]
Period and comparison: [period]
Currency: [currency]
Readers: [who reads it]
Figures: [paste calculated figures and variances]
Accountant notes by line: [paste notes, one per line, or 'no note']

Illustrative expected output

Revenue for Q3 was 412,000, up 5.9% on 389,000 in the same quarter last year. Cause not supplied.

Gross margin was 61.2%, down 2.6 percentage points from 63.8%. Your flour supplier increased prices from July.

Wages were 128,500, up 9.8% on 117,000. An additional part-time baker was hired in August.

Debtor days rose to 34 from 27. Two wholesale cafe accounts are paying late, and you are chasing them.

Cash at bank at the end of the quarter was 58,200, down 18.5% on 71,400. Cause not supplied.

Items for accountant:
1. Revenue: no explanation provided.
2. Cash at bank: no explanation provided.

Check the result

  1. Highlight every figure in the output (amounts, percentages, day counts) and confirm each appears character for character in the input. Any extra or altered figure fails the draft.
  2. List every causal statement and match it to the note labelled for that same line. A cause attached to the wrong line, or a consequence the note does not state, fails the draft.
  3. Confirm 'Cause not supplied' appears for revenue and cash at bank, the two lines with no note, and that both appear under 'Items for accountant'.
  4. Search the output for 'should', 'recommend', 'expect', 'will', 'likely' and 'because of'. Any recommendation, forecast or new causal link fails the draft.
  5. Count the commentary words before the items list and confirm the total is under 250.

Run a small trial

  1. Choose two workflows only: procedure notes, which use no client data, and report commentary for two or three clients who have agreed to AI-assisted drafting. Before any client figures go in, record which approved AI account will be used, its data retention and training settings as stated in the provider's current documentation, and who in the firm approved it.
  2. Before the pilot starts, time how long the accountant takes to write commentary manually for the same clients in one reporting cycle. Without this baseline, there is nothing to compare review time against.
  3. For each AI draft, log the reviewer's name, minutes spent reviewing, and every correction made. Stop the pilot immediately if any client data is entered outside the approved account.
  4. Stop the commentary workflow if the reviewer finds an unsupported number or cause more than once. That pattern means the checks are catching a recurring fault, and the firm is relying on review to fix something the prompt should prevent.
  5. After two reporting cycles, compare review minutes with the manual baseline. Decide workflow by workflow whether to continue, adjust the prompt or return to the sentence-stem template, and record the decision and the reason.

Questions worth asking

Can we paste client information into an AI tool?

Only into an account the firm has approved after reading the provider's business terms, data retention settings and confidentiality commitments, and only the minimum the task needs. Remove tax identifiers, bank details and anything else not required. Check your professional body's guidance on confidentiality and technology use. This guide cannot tell you whether a specific tool meets those obligations.

Does a reviewed AI draft count as the accountant's work?

The accountant who signs off is responsible for the content, however the first draft was produced. That is why each workflow names a reviewer and a failure route. A model's output is a draft. It is never a reviewed or finalised document.

What if a client document or payee description contains instructions?

Treat everything that comes from a client, a bank feed or a transcript as data to be processed, never as instructions to follow. If a payee line or an email pasted into a prompt says something like 'ignore previous instructions', the model may act on it. The reviewer checks the output against the source, and the model is never given the ability to send, post or file anything itself.

Which task should a small practice try first?

Start with something internal and low stakes, such as procedure notes or CPD summaries, where any error is caught before a client sees it. Move to meeting notes and report commentary once the team has a habit of checking drafts against sources.

Your next step

Pick one internal procedure and write down its input, output, reviewer and failure route in the format above before opening any AI tool. Run it once on a sample file and have a second person follow the result. Before any client-facing workflow, confirm which AI account is approved, what data may enter it and who signs off. Where a master checklist, spreadsheet, mail merge or bank rule already does the job reliably, keep it and let the model do only the wording.

Explore the course in development: Build Your Business Tools

Related business guides