Accountants AI Risk Checklist: Error Points, Each With a Named Check

The main risks when an accounting firm uses AI are wrong coding, invented rules, client-data leakage, stale thresholds, unreviewed drafts and vendor changes. Each is caught by a named check with a named owner: a coding spot-check, a citation check, a data-handling rule, a dated-figures check, a two-person review and a monthly vendor review.

Generic advice such as "always review AI output" fails because nobody knows who reviews what, or when. Treat each risk as an error point in the workflow, give it one check you can name in a sentence, and put a person's name against it, plus a deputy so the check still runs in busy season and during holidays.

Follow me on Instagram@sagnikteaches

The map at a glance, then each point in detail

The table is the part to copy; the sections below explain how each check works and what a failure looks like.

Connect on LinkedInSagnik Bhattacharya

Error point 1: wrong coding in the books

AI categorisation of transactions is usually the first AI task a firm adopts, and the first place errors hide. The tool codes a supplier it has seen before the same way every time, including when it was wrong. A fuel purchase goes to travel, a plant hire invoice to repairs, a personal payment to a director's loan, and nobody notices until year-end.

Subscribe on YouTube@sagnikteaches

Named check: the ten-percent coding sample. For every batch, a person re-codes a random ten percent of lines, at least 20, without looking at the AI's code, then compares. Record the error rate. Illustrative rule: under 3% errors, release the batch; 3% to 8%, check all lines for the suppliers that failed; over 8%, stop and review the rules the tool is using.

An illustrative example: a batch of 400 bank lines for a plumbing firm. The checker samples 40 lines and finds 3 miscoded: a van-insurance payment in "office costs", a supplier of boiler parts in "sundries", and a customer refund as income. The error rate is 7.5%, so the checker pulls all lines for those suppliers and finds 11 more. All 14 are fixed, and a rule is added so the supplier goes to the right code. The check cost 20 minutes and prevented 14 wrong entries from reaching the accounts. For the setup of the tool itself, see how to categorise transactions with AI and check its work.

Error point 2: hallucinated rules and references

When a model is asked what the rule is, it may state one that does not exist or applies elsewhere, with a plausible reference attached. This is called a hallucination: fluent text that is not grounded in a source. A firm's exposure is advice built on a rule that was never there.

Named check: the source-open citation check. Any rule, threshold or reference that appears in advice must be opened at the primary source by a person, and the page or paragraph recorded. Asking the model whether its own citation is real is not a check. If the source cannot be found in five minutes, the statement is removed.

Illustrative failure: an assistant drafting a client note says "the annual limit for this relief is $12,000 under the relevant provision". The preparer opens the official guidance and finds the limit is different and applies only to a defined group of clients. The note is corrected before sending, and the near miss goes in the log. The same discipline for legal work is in how to stop AI inventing case law. For background on why this happens, read AI hallucinations explained for business owners.

Error point 3: client data leaving the firm

Leakage happens in ordinary ways: a junior pastes a client's bank statement into a free chatbot to save time; a browser extension reads page content; an "AI notetaker" joins a client call. The risk is not just training on your data; it is where the data lives, who can see it and how long it stays.

Named check: the approved-tools rule and settings audit. Publish a list of approved tools. Each has been through the privacy-terms check (training on content, retention, deletion, sub-processors, location of storage). Everything else is not for client data. Each quarter, the owner reviews the settings and the list of connected apps.

A few verifiable points help. Business plans of the major assistants, such as ChatGPT Business and Claude Team, do not train on business content by default, whereas consumer plans have a settings switch that individuals control. Both business plans have a two-seat minimum. The firm cannot enforce settings on personal accounts, which is why the rule is "approved tools only". Read the terms yourself on the vendor's pages, since they change, and use what to check in an AI tool's privacy policy and terms as the checklist.

Illustrative audit finding: the quarterly review finds a meeting-recorder extension installed by one preparer, set to join every video call. The practice manager removes it, adds "recording tools" to the approved-list process, and notifies affected clients as the firm's confidentiality terms require. See also how to stop AI tools training on your business data.

Error point 4: stale thresholds and rates

Models learn from material that ends at a date, then answer as if no time has passed. Rates, limits, filing dates and allowances change, often yearly. A draft that quotes last year's figure looks right to everyone who does not check.

Named check: the dated-figures table. Keep one table, owned by a technical lead, of every figure the firm commonly quotes, with the tax year, the primary source and the date last verified. Any AI draft that contains a figure is checked against the table, never against the model's memory. Update the table at the start of each tax year and whenever a change is announced.

The table is deliberately blank on values here because yours must come from your own checks; copying a value from a tutorial defeats the purpose. A useful extra: paste the table into the prompt for any drafting task and tell the assistant to use only those figures. The check becomes a comparison of the draft with the table, which takes seconds.

Error point 5: drafts that nobody reviewed

The quietest risk is a fluent document that leaves the firm without a second pair of eyes: a client letter, a summary of a meeting, a computation note. Fluency makes review feel unnecessary, which is exactly when errors slip through.

Named check: the two-person release rule. No AI-assisted document goes to a client until a second person has read it against the source documents and initialled it. The person who prompted the tool is never the only reviewer. For documents with figures, the reviewer re-adds the key totals. Mark AI-assisted drafts in the file name or metadata so the reviewer knows to look harder.

Illustrative failure: a letter summarising a client's year contains the right narrative but swaps two years' profit figures. The preparer misses it because the paragraph reads smoothly; the reviewer, comparing the letter with the computation, catches it. Time cost of the review: about 10 minutes. For tasks with higher stakes see AI return review against software diagnostics.

A review checklist for AI-assisted documents

• Every figure traced to a source document or the dated-figures table.
• Every rule or reference opened at the primary source.
• Names, dates and years correct (models mix up clients if context carries over).
• Tone suits the client; no promises, no definitive legal or tax advice beyond your remit.
• Nothing in the text reveals another client's details.
• Reviewer's initials and date on the file copy.

Error point 6: vendor changes you did not choose

Tools change under you. A feature is retired, a plan is merged, a model is swapped, a data policy is updated. Any of these can break a workflow or alter your data position without any action on your side.

Recent examples show the pattern. Excel's =COPILOT() worksheet function, offered as a preview, stopped working on 14 September 2026: results already calculated stay in the workbook as cached values, but any cell that recalculates now returns a #NAME? error. OpenAI has scheduled custom GPTs to retire on 11 December 2026, and its migration route turns a GPT into a plugin, so any firm process built on one needs testing in its new form before then. Microsoft stopped selling Copilot Pro in October 2025 and moved subscribers towards Microsoft 365 Premium. A bookkeeping service that closes or changes hands can leave clients exporting years of records at short notice. None of these needed a mistake by the firm; each needed someone who noticed.

Named check: the monthly vendor-notice sweep. The owner spends 20 minutes a month reading release notes or email notices from each approved vendor and updates a one-line status per tool: unchanged, changed (what and when), or at risk. Any workflow that depends on a feature has a recorded fallback. Keep an exit plan for any tool holding client records: how to export data and how long it takes.

The one-page register

Put the map in a document the team can see. Add three columns that turn it from a list into a control: evidence (where the check's record lives), last run and next due. A register without dates cannot show whether checks are happening.

Sample log entries so the team knows what to write

People skip logs when they do not know what a good entry looks like. Illustrative entries, each one line:

• Coding, 3 Nov: batch of 400 lines, sample 40, 3 errors (7.5%), all lines for the three payees rechecked, 11 more fixed, rule added.
• Citation, 4 Nov: draft cited a relief limit; opened source, figure wrong, corrected before sending.
• Settings audit, 1 Oct: unapproved recording extension found and removed; two clients informed.
• Figures, 6 Apr: new tax-year table verified against the official pages, 14 figures updated.
• Release, 5 Nov: client letter had two years' profits swapped; caught at review.
• Vendor, 30 Sep: Excel COPILOT function retired; two workbooks showed #NAME? errors and were rebuilt with ordinary formulas by 10 Oct.

After a quarter, the log shows you which error point fails most. If coding errors dominate, the tool's rules need work. If citation errors dominate, the team is asking models for rules from memory and needs the habit of pasting sources first.

A worked example: a six-person practice sets this up in a week

An illustration with invented figures. A six-person practice (two partners, a manager, a senior bookkeeper and two preparers) already uses an AI categorisation feature in its bookkeeping software and a business-plan assistant for drafting letters. Nobody has written down who checks what. The practice manager decides to build the register in a week.

1. Monday, 60 minutes. List every place AI touches client work: bank-line coding, letter drafting, meeting notes, research questions. The list has seven items.
2. Tuesday, 60 minutes. Map each to one of the six error points. The meeting-notes tool turns out to be an unapproved recorder, which becomes the first leakage finding.
3. Wednesday, 90 minutes. Build the dated-figures table with the technical lead: 22 figures the firm quotes most, each opened at the official page.
4. Thursday, 45 minutes. Agree owners and deputies, and set the first sampling run. The senior bookkeeper samples 30 lines from last week's batch as a trial and finds 2 errors (6.7%).
5. Friday, 30 minutes. Hold a short team session, walk through the six sample log entries, and publish the register on the shared drive.

Total set-up: about five hours of the practice manager's and technical lead's time, spread over a week. Running cost afterwards is roughly the 10% coding sample, a few minutes per advice item for citations, 20 minutes a month for the vendor sweep and an hour a quarter for the settings audit. For the firm in this illustration that is perhaps three hours a week across the team, which is the price of being able to say how AI is controlled.

Using the dated-figures table inside a prompt

The table is most useful when the model is told to use it. A prompt that drafts a client reminder letter might read:

Draft a short letter to a sole-trader client reminding them of
their filing deadline and the home-office flat rate. Use ONLY the
figures in the table below. If a figure you need is not in the
table, write [FIGURE NEEDED] instead of supplying one. Do not
state any rule or reference not listed in the table.

[DATED FIGURES TABLE: figure, value, tax year, source, verified]

A realistic output: a four-paragraph letter with the deadline and flat rate correctly taken from the table, and a placeholder, [FIGURE NEEDED], where the model wanted to mention a late-filing penalty amount that was not in the table. That placeholder is the system working. The reviewer either adds a verified figure from the source or deletes the sentence. Without the instruction, the model would probably have written a number that looked right.

What to do the day a check fails

A failed check is information, not a disaster, provided you respond the same way each time. A short incident routine keeps it calm.

1. Contain. Stop using the tool or the prompt for that task until the cause is known. If a document has already gone to a client, tell the owner at once.
2. Correct. Fix the wrong entry, figure or sentence at source. If a client received wrong information, send a short correction in plain words, reviewed by a partner.
3. Find the cause. Was the prompt vague, the source missing, the check skipped or the tool changed? One sentence in the log.
4. Change something. Update the prompt, the table, the rule or the owner. A log entry without a change is incomplete.
5. Tell the team. Two lines at the next meeting. People repeat what nobody mentions.

Escalate to a partner anything involving client data leaving the firm, since confidentiality duties and notification rules may apply, and take advice from a qualified adviser on what you must do.

Making the checks stick with a small team

The most common reason checks fail is time pressure at deadlines, so build them into the workflow, not on top of it. Put the citation column in the working-paper template so the preparer sees it. Add the release initials to the document checklist. Set a calendar reminder for the monthly sweep on the first Monday. Make the dated-figures table the default source in your shared prompt library. Reward the person who catches an error in the team meeting; if catching errors feels like an accusation, people stop logging them.

New starters should meet the register in their first week. Give them the six error points on one page, the approved-tools list and two real log entries, and ask them to run one coding sample under supervision. It takes an hour and sets the habit before bad ones form. A wider introduction plan is in what an AI implementation looks like in a small accounting firm.

How to decide how strict to be

Match the control to the cost of an error. A wrong category on an internal management report is annoying; a wrong threshold in a letter to a client is a professional-standards matter. For low-stakes work, keep the checks light: a coding sample of one in ten, a quick read. For anything that leaves the firm or touches a filing, use the full checks. Write down which documents fall into which tier, so staff do not decide under deadline pressure. The same thinking helps when you cost the rollout; what AI costs a small accounting firm counts checking time as a real line item, which it is.

Also remember that these checks cover errors, not professional obligations. Your regulator, your insurer and your engagement letters set the rules on confidentiality, competence and disclosure; confirm with them and with a qualified adviser what applies to you. If you want help deciding which of the six matter most in your workflow, that is what my AI implementation consultation covers; see the contact page.

Questions firm owners ask about AI risk controls

How often should each check run?

Match the frequency to the risk. Coding spot-checks run on every batch, citation and threshold checks run on every piece of advice that relies on them, data-handling reviews run quarterly, and the vendor-change check runs monthly. A schedule you can keep is better than a perfect one you drop.

Who should own the AI risk register?

One named person, usually a partner or the practice manager, owns the register. Each check also has its own owner who performs it. If everyone owns it, nobody does. Write the names on the register so absences are covered by a named deputy.

Do small firms really need a written register?

Yes, a short one. A single page with six rows takes an hour to create, and it is what you show a client, an insurer or a reviewer who asks how you control AI. Without it, the controls exist only in people's heads.

Should staff be allowed to use free AI tools on client work?

Not for client data. Free consumer plans may use chats for training unless a setting is changed, and the firm cannot enforce settings on personal accounts. Provide a paid business tool whose terms you have read, and make the rule explicit in your AI policy.

Further reads

• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven AI mistakes that put client trust at risk.
• Can AI Do My Bookkeeping? What Still Needs an Accountant — What still needs an accountant when AI does bookkeeping.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — A readiness checklist for professional firms.
• Rolling Out AI in a Bookkeeping Practice Without Losing Control — Rolling out AI in a bookkeeping practice with control.
• How Small Accounting Firms Use AI Day to Day: Real Examples — Real examples of AI in daily accounting work.
• Questions to Ask an AI Vendor Before You Sign Anything — Questions to put to a vendor before signing.
• Dext vs AutoEntry for a Small Practice: Costs, Accuracy Claims and How to Check Them — Run a 30-document accuracy test on Dext and AutoEntry, then compare cost per client with cost per credit at your own volumes.
• Virtual CFO Pricing: Build a Monthly Package With AI-Prepared Reports — How a firm can price a monthly virtual CFO package: three tiers, the reports in each, review time, tool costs and add-on rules.
• AI Accounting Workflow: Client Upload to Reviewed Books, Step by Step — Five stages from client upload to partner review, each with an owner, tool and check, shown through one worked month for a small contractor client.
• AI Accounting in New Zealand: GST Returns, IRD and Xero Fit — Use AI to check a GST return before filing: IRD dates and thresholds, supplier document checks, a box-by-box worked review and what each Xero plan includes.
• AI Accounting in Singapore: GST, ACRA Filings and Software Choices — How to use an AI assistant for GST F5 coding checks, a filing calendar and software questions, while a named person files with IRAS and ACRA.
• 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.
• AI Chatbot for Accountants: Website FAQs Without Giving Tax Advice — Scope a website chatbot to fees, documents and dates, with refusal wording, a hand-off flow and a weekly log review for a small accountancy practice.
• AI Marketing for Accountants: LinkedIn, Newsletters and Local Search — A five-hour monthly plan for an accounting practice: newsletter, LinkedIn and Business Profile posts, with rewrites that keep tax claims out.
• Notion for Accountants: Firm Wiki, Client Trackers and Notion AI — A firm wiki, a client tracker database and a costed verdict on Notion AI, with the templates written out field by field.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business pricing page and OpenAI Data Controls FAQ; Claude pricing page and Anthropic privacy articles on model training; Microsoft Support article on the COPILOT function; OpenAI help article Custom GPT retirement and migration FAQ; Microsoft 365 Premium announcement (1 October 2025); all checked 4 November 2026. Example figures are illustrative.