Accounting Firm Standard Operating Procedures With AI: Record to Publish

An accounting firm can use AI to create standard operating procedures by recording a person doing the task, turning the transcript into numbered steps, having that person correct the draft, then publishing it with a named owner and a review date. The recording captures what really happens; the AI only does the typing.

Most firms already have procedures, in the heads of two senior staff. The risk shows up when one of them is away in the second week of a filing deadline. The fix is to get them out of people's heads and into one place where they stay accurate, starting with the firm's own work: month-end, busy-season peaks, onboarding and sign-off. For the general method in any business, see how to create SOPs from screen recordings with AI.

Follow me on Instagram@sagnikteaches

Choose the first procedures by what hurts when they go wrong

Do not start with the easiest task. Start with the tasks where an absence or an error costs money or a deadline. A quick ranking, scored one to three on each column, takes ten minutes and gives the order of work.

Connect on LinkedInSagnik Bhattacharya

The scores are the firm's own judgement, not a formula. The point is that payroll, deadlines and engagement letters come first because they sit in one head and carry a high cost. Month-end close comes next because it repeats twelve times a year per client, so any improvement multiplies.

Subscribe on YouTube@sagnikteaches

Stage one: record the work as it is done

Ask the person who does the task most often to do it for real, once, while recording their screen and speaking their thoughts aloud: "I'm opening the bank feed. I check the last reconciled date first, because if it's not this month's I stop and email the client." The sentences that begin "because" and "if" are the ones you cannot get from a template. They are the judgement.

Practical points for the recording:

• Use a demonstration file where possible. A test company in your accounting software shows every screen without exposing a real client's figures. Where you must record a live file, plan to blur or redact before anything leaves your machine.
• Keep each recording to one task. A 12-minute recording of "reconcile one bank account" is easy to turn into a procedure. A 90-minute recording of a whole month-end is not.
• Note the trigger and the end. Say at the start, "this runs on the third working day after month end", and at the end, "this is finished when the reconciliation report is saved in the client folder".
• Record the exceptions. Ask: "What goes wrong here?" The answer is usually the most valuable minute in the recording.

Any screen recorder that can export a transcript will do. If it cannot, use a meeting tool's recording and transcript, and check what it does with the file and who can see it.

Stage two: draft with AI, with guard rails

Paste the transcript into a chat assistant with a prompt that tells it what shape to produce and what not to do. This is the prompt a firm might keep on file:

Below is a transcript of an accountant explaining a task while doing it on screen.
Turn it into a standard operating procedure with these headings:
PURPOSE (one sentence), WHEN IT RUNS, WHO DOES IT, WHAT YOU NEED BEFORE STARTING,
STEPS (numbered, one action per step, starting with a verb),
CHECKS BEFORE YOU FINISH, WHAT GOES WRONG AND WHAT TO DO, DONE WHEN.
Rules:
- Use only what is in the transcript. If a step seems to be missing, add a line "[CHECK: ...]" instead of inventing it.
- Keep screen names, menu names and field names exactly as spoken.
- Flag anything that looks like a client name, number or login and replace it with [client].
- Write for a new starter who has never done this task.
TRANSCRIPT:
[paste]

The "[CHECK: ...]" instruction is the safety. Without it, an assistant fills gaps with plausible steps from its general knowledge of accounting, and those steps may not match your software or your firm's rules.

A worked draft: month-end bank reconciliation

Illustrative example. Suppose a bookkeeper records a twelve-minute reconciliation for a small furniture maker's account in a demonstration file. The transcript includes the line: "I look at the statement balance on the bank feed and compare it to the balance in the ledger, then I go through anything unmatched, and if it's more than twenty items I tell the client because it usually means a feed problem." The assistant produces:

STEPS
1. Open the bank account in the ledger and note the date of the last reconciled transaction.
2. Compare the statement balance on the bank feed with the ledger balance.
3. Open the list of unmatched transactions.
4. Count the unmatched items. If there are more than 20, stop and email the client: this usually means a feed problem. [CHECK: what is the next step after notifying the client?]
5. Match or create each remaining transaction.

Two things for the owner to fix. The "[CHECK]" is where the bookkeeper says, "I wait for the client to reconnect the feed, then restart from step 2, and I log it in the client's note." Step 5 is too big: "match or create" hides the actual rules, so the bookkeeper breaks it into matching against bills, creating bank fees from the standard list, and querying anything over $500 with no document. The finished procedure is longer than the draft and better, because the draft showed where the detail was missing.

Stage three: the owner reviews against the real task

The owner reviews, not the partner and not whoever has spare time. Their review has two parts. First, read the draft against their own memory and correct anything wrong. Second, and more usefully, give it to someone who has not done the task and watch them try. Every point at which they pause is a missing line.

Illustrative: a junior who has never run payroll follows the draft for a four-person garage. They stop at "approve the run", asking who approves and by when. The owner adds: "The partner approves by 4pm on the 20th; if they are away, the deputy named on the rota approves." That is a decision, not a step, and it will only ever come out in a test with a new person.

Record each review in the document itself: reviewer, date, what changed. That is the audit trail if a client or a professional body asks how the firm keeps its procedures controlled.

Checklists for the two crunch periods

Individual procedures are the foundation. On top, the firm needs two short checklists that sequence them. Ask the AI to build them from your procedure list rather than from its general knowledge. Illustrative month-end checklist for one bookkeeping client:

1. Working day 1: confirm bank feeds are connected; send the client the receipts reminder.
2. Working day 3: run the bank reconciliation procedure; log unmatched items.
3. Working day 4: run the payroll procedure if the client is on payroll.
4. Working day 5: review the profit and loss against last month; write the two-line commentary.
5. Working day 6: send the summary; record the date sent in the client tracker.

The busy-season checklist works the same way but is tied to deadlines. List each filing date, the procedure that produces it, the internal deadline (a week earlier) and the person responsible. Ask the assistant to flag any date with fewer than ten working days between two big deliverables for the same person. A firm with one payroll specialist and three clients whose returns fall due in the same fortnight can see the collision in March and not in the week itself.

The register: one owner, one date, one home

Procedures decay when nobody is named. Keep a register, a single table with a row per procedure:

The firm does not need a special system. A Notion database with the review date as a property works well, and Notion's page verification fits procedures closely: you verify a page until a set date, it shows a blue check mark in search, and when verification expires the page's owner is notified in their Notion inbox and by email to re-verify it. Notion's pricing page lists "Verify any page" under the Business plan; as of November 2026 it shows Plus at $10 and Business at $20 per member per month billed yearly ($12 and $24 billed monthly), so check which plan you are on before you design around verification. Firms already on Microsoft 365 can do the same in SharePoint with a list or a library and a review-date column, plus a calendar reminder or a flow that nudges the owner.

Whichever you choose, three rules keep it honest: one home, so that people never wonder where the current version is; a visible "last reviewed" date on each page; and old versions archived out of search so that nobody follows a retired step. For a wiki built on Notion, Notion for accountants goes further on structure.

The quarterly review routine

Put a fixed hour in the diary each quarter. The agenda is always the same and takes one hour.

1. Sort the register by next-review date and list everything due or overdue.
2. For each due procedure, the owner answers three questions: has the software changed, has the rule changed, and has anyone done it differently since?
3. Check any procedure involved in an error or near miss in the quarter. Ask the AI to compare the error description with the procedure and say which step did not prevent it.
4. Update, re-date and re-sign. If nothing changed, re-date it anyway; the date is the evidence the question was asked.
5. Pick one new procedure to record for the next quarter.

The third step is where AI earns its keep in maintenance. Paste in a short, anonymised description of what went wrong and the current procedure, and ask which step should have caught it and what check could be added. Treat the answer as a suggestion for the owner to test, since the assistant does not know the firm's software settings.

Notion or SharePoint: a worked choice

The home for procedures depends on where the firm already works, not on features. A short decision table:

Illustrative: a five-person practice that runs email, calendars and file storage on Microsoft 365 puts procedures on a SharePoint site with a library per area, a single register list, and a "last reviewed" column. A three-person bookkeeping firm with no Microsoft estate sets up a Notion workspace with one procedures database, a view by owner and a view by next-review date. Neither is wrong. What breaks both is a second copy somewhere else.

Redaction routine before a transcript reaches an AI tool

Because recordings are made at work, a quick routine stops client details leaking into drafts. Before pasting a transcript, search it for four things: client names and trading names, any figure over a few hundred, bank sort codes and account numbers, and email addresses or logins. Replace each with a bracket such as [client] or [amount]. This takes two minutes on a 12-minute recording. The prompt above also asks the assistant to flag anything that looks like client data, but that is a second layer, not the first. Check your tool's data-handling settings and your engagement terms, and ask a qualified adviser if you are unsure what clients have agreed to.

Before and after: turning a vague step into a usable one

A first-draft step for the year-end accounts procedure reads: "Review the accounts and make sure they look right." It is the kind of line that appears when an assistant summarises a casual explanation. The owner, an experienced manager, rewrites it with the recording in mind:

1. Compare each profit and loss line with last year; note any line that moved by more than 15% and over $1,000.
2. For each noted line, find the reason in the ledger or ask the client; write one sentence in the review notes.
3. Check that the balance sheet shows the loan balance agreed to the lender's statement.
4. Confirm directors' loan account has not turned negative; if it has, flag it to the partner before sign-off.

The thresholds (15%, $1,000) are the firm's own, and they came from the owner's judgement in the review, not from the AI. A new reviewer can now follow four concrete steps, and a partner can see exactly what was checked.

Handing over busy-season work

One more example shows why procedures pay off at the moment of greatest stress. Suppose a firm's payroll lead goes on leave in the last week of the financial year, and a manager covers. Without a procedure, the manager phones. With one, the manager opens the payroll page, sees "who approves", "where the timesheets live" and "what to do if the bank file is rejected", and gets on. The first time this happens the firm learns two lines were missing. The manager adds them on the day, and the owner updates the review date. The procedure improves when it is used, which is exactly why it must be easy to find.

Things that go wrong, and how they show up

• A confident but invented step. A draft says "export the trial balance to the review template", and the firm has no such template. Found by the new-starter test, and avoided by the "[CHECK]" rule.
• Client data in the recording. A transcript includes a real company's balance. Prevent it with demonstration files, and review any transcript before it goes into an AI tool.
• Procedures that nobody opens. Links are buried three levels down. The fix is to link the procedure from the place the work happens: the client tracker row or the task in your practice software.
• The owner leaves. The register shows an owner who has left the firm. Add "reassign owners" to the leaver checklist.
• Over-documenting. Forty procedures for tasks done once a year. Start with the priority table and let the register grow only when something goes wrong or someone asks.

How to tell it is working

After one quarter, four checks. Can a new starter complete payroll for a test client using only the page? Does every procedure in the register have an owner and a review date? Did any task come back to the partner as a question the document should have answered, and was the document updated? And did the firm's cover for holidays improve, so that the person who stood in did not need to phone the person who was away? If the answer to the last is still no, the procedure that was missed is the next one to record.

It also helps to link the procedures to the rest of the firm's AI use. A procedure for the month-end close is where you would add an AI error check, as described in how accountants use AI to spot errors in client books, and where the new-client steps connect to automating client onboarding in a small accounting practice. Written procedures are the base that makes either of those safe, because the AI step has a named place in the sequence and a named person who checks it.

Questions about AI-written procedures in an accounting firm

How many procedures does a small firm need?

Start with the ten to fifteen tasks that, if the person doing them were off sick for a week, would cause a missed deadline or an error: month-end close, payroll runs, year-end accounts, onboarding, filing deadlines and engagement letters. Add the rest as they cause problems.

Is it safe to record screens that show client data?

Only with care. Record on a test client or a demonstration file where you can, blank out names and numbers, and check your engagement terms and your tool's data settings before uploading any recording or transcript that contains real client information.

Who should own each procedure?

The person who does the task most often, not the most senior person. They know where the procedure is wrong. A partner or manager signs it off, but the owner keeps it current and answers questions about it.

How often should procedures be reviewed?

Quarterly for the busy-season and filing procedures, before the season starts, and twice a year for the rest. Also review immediately after any error, software change or staff change that touches the task.

Further reads

• How to Write SOPs With AI: From Rough Notes to Clear Procedures — Drafting procedures from rough notes instead of recordings.
• AI for Creating SOPs: How Consultants Document a Client's Process — Documenting a client's process rather than your own.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — What goes wrong when AI meets client data.
• HubSpot for Accounting Firms: Lead Tracking and Onboarding Pipelines — Pipeline stages that your onboarding procedure can hang from.
• AI for Chartered Accountants: What to Tell Clients About Your AI Use — What to tell clients about the AI in your workflow.
• How to Price Accounting Services: Quote Fixed Fees With AI Support — A fixed monthly fee built from a scope list, an hours-times-rate check, three tiers and an AI-assisted quote worksheet, with every sum shown.
• AI Agents for Auditors: Chasing Evidence Requests From Clients — How a small audit firm can let an AI agent chase client evidence and pre-screen uploads, with named approvers and dated sign-offs for everything else.
• AI Agents for CPA Firms: Starting Jobs and a Review Gate for Each — A small CPA firm's first four agent jobs, scored and sequenced, with a review gate, pass rule and stop condition for each, plus a 90-day start.
• Automate BACS Payments in an Accountancy Practice: Where AI Fits — A step-by-step BACS payment run for a practice: file build, AI checks on changed bank details, dual approval, submission and reconciliation.
• AI for Accountants in Australia: Tax Agent Duties and Client Data — The Code of Professional Conduct, client consent and Privacy Act steps for tax and BAS agents using AI, plus three safe first jobs.
• AI for Accountants in Canada: Privacy Rules and First Uses — Which privacy laws and CPA conduct rules apply when a Canadian firm uses AI, how to vet a vendor, and where to start safely.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Notion pricing page (checked 4 November 2026) for the plan features named; no other vendor facts are stated. The month-end example and all figures are illustrative.