Lesson objective

Distinguish rules-based automation, machine learning and generative AI in accounting work, and match the check you apply to the way each one fails.

Plain-language explanation

Why this matters now

"AI" now appears on almost every product an accounting professional touches: the bank feed that suggests a coding, the tax software that flags a missing form, the audit tool that scores journal entries, and the chat assistant that drafts a client email. The label is the same. The systems underneath are not, and they fail in different ways.

Two recent surveys show how quickly the question has arrived. Each needs its sponsor and sample stated before its numbers mean anything.

  • The Thomson Reuters Institute's 2026 AI in Professional Services Report (published 9 February 2026) gathered more than 1,500 respondents from legal, tax and accounting, corporate functions and government in 27 countries. It reports that organisation-wide generative AI use "almost doubled to 40% in 2026, compared to 22% in 2025." This is vendor research: the Institute is the research arm of a company that sells AI tax and accounting products. The figures combine every professional-services respondent, not tax and accounting firms alone, and the fieldwork dates and sampling method were not stated on the pages reviewed for this course.
  • The AICPA & CIMA Future-Ready Finance survey (announced 17 December 2025) asked 1,446 senior finance and accounting leaders and managers who are AICPA or CIMA members, with fieldwork in August and September 2025. 88% believe AI will be the most transformative technology trend in accounting and finance over the next 12 to 24 months. Only 8% feel their organisation is very well prepared to manage it, and 21% feel it is well prepared. 56% named generative AI as the most prominent skills gap. It is a member survey of mostly business and industry leaders, not a random sample of CPA firms, and it measures perceptions, not how well anyone actually uses AI.

Read only as far as their samples allow, both point to the same gap: use and expectation are rising faster than preparedness. The first step in closing that gap is knowing which kind of system you are looking at.

A question to sit with

Your bank-feed software suggests an account code, your tax software flags a missing schedule, and a chat assistant drafts a note to the client. All three are sold as "AI". Which of them could be wrong in a way you would not notice just by reading what it produced?

What goes wrong

Treating every "AI" label as one thing. A firm writes a single instruction, "check AI outputs", and applies it the same way to a fixed bank rule and to a generated draft. The bank rule is over-checked every month and the draft is under-checked every time.

Checking a generated output the way you check a rule. A rule that was tested once behaves the same way the next time it meets the same facts. A generative system produces new text every time it runs. Having checked yesterday's output tells you very little about today's.

Reading a confidence score as a verdict. Machine-learning suggestions often arrive with a score. The score is the system's own estimate of how strongly its learned patterns point to that answer. It is not proof that the coding is right for this transaction, and it is weakest exactly where the transaction is unusual.

Assuming fluent language means understanding. A well-written paragraph about a client's margin trend can rest on a misread figure or an invented reason. Fluency is a property of the text, not of the analysis behind it.

Letting the tool's label decide who owns the result. No product category moves responsibility. The professional who would have owned the work without the tool still owns it with the tool.

The practical answer

Sort every AI feature you meet into one of three families before you decide how to check it. Many products combine more than one family, so sort the feature, not the product.

Rules-based automation Machine learning Generative AI
How it works A person writes explicit if-then logic The system learns patterns from past examples, such as prior codings A model trained on very large amounts of text produces new text, summaries, tables or code
Typical accounting use A bank rule codes every payment to a named utility supplier to utilities expense Suggested account codes for bank-feed lines; flags on unusual journal entries Drafting a client email, summarising a lease, explaining a variance
How it fails Consistently: the rule is wrong, or a transaction falls outside it Quietly on unusual items; drifts when the business changes Plausibly: confident, well-formed content that is false
What a wrong output looks like The same wrong answer every time A reasonable-looking suggestion, often with a score A fluent, complete-looking answer
How you check it Test the rule, review exceptions, control changes to the rule Sample suggestions, look hardest at unusual and low-score items, watch for drift Verify every fact, figure and reference against a source you hold

Three questions turn the table into a habit.

  1. Which family produced this output? If you cannot tell, ask the vendor or your IT lead, and record the answer.
  2. How does that family fail? Match your check to the failure. A rule needs a test. A suggestion needs sampling. A generated draft needs line-by-line verification.
  3. Who owns the result? That depends on the work, not on the tool:
    • a licensed CPA for attest work (audits, reviews, compilations) and for any report that carries a CPA firm's name;
    • the signing tax return preparer for a return or tax advice;
    • the responsible accountant or controller, under management's responsibility, for a company's own books and financial statements.

What the evidence says changes in the work. Jung Ho Choi (Stanford Graduate School of Business) and Chloe L. Xie (MIT Sloan), in Human + AI in Accounting: Early Evidence from the Field, published in the Journal of Accounting Research in 2026, combined a survey of 277 professional accountants with field data from one AI-enabled accounting platform serving 79 small and medium-sized enterprises, covering more than 200,000 transaction-level records. The survey documents "substantial heterogeneity in adoption patterns, perceived benefits, and concerns." In the field data, adoption was associated with productivity gains and a shift of effort "away from routine data entry toward business communication and quality assurance tasks," with "more granular ledgers and faster month-end closing." Accountants "selectively intervene when AI confidence scores are low." In a framed field experiment, AI assistance improved classification accuracy on average, but "reliance on non-consensus AI recommendations can increase the risk of error."

Those findings are associations from one platform's small-business clients, not an estimate for every firm. Two points still carry across by reasonable inference. The work moves toward review, so review becomes the core skill. And the item that hurts is rarely the obvious error; it is the confident suggestion that departs from what the evidence supports.

Apply that idea to a single feature at a time. For each AI feature in your work, write one line: what the system does, what a person does, and who decides.

Feature System does Person does Who decides
Bank rule Codes matching payments Reviews the exceptions report monthly Controller
Suggested coding Proposes an account with a score Accepts, changes or rejects each line Staff accountant, reviewed by the controller
Drafted client email Produces a draft Verifies every fact and figure, edits, sends The engagement owner

Remember this

  • "AI" covers three families (rules-based automation, machine learning and generative AI) that fail in different ways.
  • Match the check to the failure: test rules, sample suggestions, verify generated content line by line.
  • A confidence score is the system's own estimate, not proof that the answer is right for this transaction.
  • A survey figure means nothing without its sponsor, sample and date.
  • The tool never changes who owns the work: the CPA, the signing preparer or the responsible accountant or controller.

Accounting example

Harrow & Pell CPAs is a fictional eleven-person firm with a client accounting services team. Senior accountant Maya Okafor looks after the books of Linden Row Bakery, an invented three-shop business. Before the March close, her manager asks her to list every AI feature touching the file and say how each is checked. Maya finds three. The first is a set of bank rules written by a former staff member. One rule codes any payee containing the word "Power" to utilities expense. She tests it against the month and finds it has coded a $3,480.00 payment to Power Oven Services, an equipment repairer, to utilities. The rule is rules-based automation: it fails consistently, so she corrects the rule and adds it to the exceptions review. The second is the ledger's suggested coding, a machine-learning feature that shows a score. It has proposed "Repairs and maintenance" with a high score for a $9,200.00 payment that was in fact a new proofing cabinet, which the bakery's policy capitalises. The pattern looked like past repair payments to the same supplier, so the score was high and the answer was wrong. Maya changes the coding and asks that single payments over the bakery's capitalisation threshold be routed to a person regardless of score. The third is a generative assistant that drafts the monthly commentary. Its draft says gross margin improved "because flour prices fell in March." Nothing in the file mentions flour prices; the bakery's supplier invoices show flour costs were flat. The sentence was generated because it sounded like a reason. Maya deletes it and writes the real driver from the sales mix report. Her one-page note lists each feature, its family, its check and its owner: the controller for the rules and coding, and Maya, reviewed by the engagement manager, for the commentary.

Responsible-use warning

Do not let a high confidence score or a polished paragraph shorten your review. A score is the system's own estimate based on patterns it has seen before, and it is least reliable on the unusual transaction that matters most. A fluent generated explanation can contain a reason nobody supplied and no document supports. Neither is evidence. Before you accept an AI output, name the kind of system that produced it and apply the check that matches its failure: test the rule, sample the suggestion, or verify every fact and figure in the draft.

Human-review requirement

The owner of the work decides whether each AI feature may be used and how it is checked, and that owner depends on the work. For a company's own books, the responsible accountant or controller, under management's responsibility, owns the coding rules, the review of suggestions and the exceptions. For a tax return or tax advice, the signing preparer owns every figure. For attest work or any report carrying a CPA firm's name, a licensed CPA owns the result. The reviewer confirms the family of each feature was identified, that the check matched its failure mode, and that no output was accepted on the strength of a score or fluent wording alone.

Practical exercise

Map the AI features in one real workflow you know, using fictional or anonymised examples only. Do not paste client data anywhere for this exercise.

  1. Choose one recurring task, such as a monthly close for a small client, a batch of bank-feed coding, or a set of client emails.
  2. List every feature in the tools you use for that task that is described as AI, smart, automatic or suggested.
  3. For each feature, decide its family: rules-based automation, machine learning or generative AI. If you cannot tell, write "unknown" and the name of the person you will ask.
  4. For each feature, write in one sentence how that family fails.
  5. For each feature, write the check that matches the failure: a rule test, a sample of suggestions, or line-by-line verification of generated content.
  6. Complete a four-column table: feature, what the system does, what a person does, who decides.
  7. Name the owner of the output using the three categories in this lesson: licensed CPA, signing preparer, or responsible accountant or controller.
  8. Circle any feature where the current check does not match the family, and write one change you would propose to your manager.
  9. Keep the table with your workbook notes; you will reuse it when you build a risk matrix later in this Level.

Knowledge check

  1. Why should a bank rule and a generated client email not be checked in the same way?

    Answer guidance: A bank rule is explicit logic that fails consistently, so it is tested and its exceptions reviewed. A generative system produces new content each time and fails plausibly, so every fact, figure and reference in each output must be verified against a source.

  2. What does a confidence score on a suggested account code tell you, and what does it not tell you?

    Answer guidance: It is the system's own estimate of how strongly its learned patterns point to that answer. It does not prove the coding is right for this transaction, and it is least reliable on unusual items that do not resemble past data.

  3. What did Choi and Xie find about relying on AI recommendations in their framed field experiment, and what limits their findings?

    Answer guidance: AI assistance improved classification accuracy on average, but reliance on non-consensus AI recommendations could increase the risk of error. The findings are associations from one platform serving 79 small and medium-sized enterprises, not an estimate for all firms.

  4. Before quoting the Thomson Reuters Institute figure that organisation-wide generative AI use rose to 40%, what must you state?

    Answer guidance: That it is vendor research from the research arm of a company selling AI tax and accounting products, that it covers more than 1,500 respondents across all professional services in 27 countries rather than accounting firms alone, and that fieldwork dates and sampling method were not stated.

Sources and further reading