Lesson objective

Explain how volume, handoffs, shared inputs and fading verification change the risk of an AI-assisted accounting task once a team runs it across many clients, and convert an error rate into an expected monthly count.

Plain-language explanation

Why this matters now

Level 2 taught one careful person to run one AI-assisted task well: triage it, prompt it, verify it against the source, record it and have it signed. Level 3 is about what happens when that task becomes something a team does every day for many clients.

A bank reconciliation narrative drafted by a senior who checks every reconciling item is a task. The same narrative produced for 120 client entities each month by nine people, and reviewed by two managers in the middle of the close, is a system. The tool can be identical. The risk is not.

Evidence about how accounting firms actually use these tools is still thin, so read it carefully.

Professional guidance (staff observations, not binding): in July 2024 PCAOB staff published a Spotlight on their outreach about generative AI in audits and financial reporting. It states that it "represents the views of PCAOB staff and not necessarily those of the Board. It is not a rule, policy, or statement of the Board." The outreach was limited and covered mainly larger firms. Staff observed that GenAI use in audits by the firms interviewed "appears to be focused primarily on administrative and research activities." Firms said an engagement team member using a GenAI tool "is still responsible for the results and documentation of the work." Treat it as a mid-2024 snapshot of mostly large firms, not a picture of practice in 2026.

Research (observational): Jung Ho Choi and Chloe L. Xie, Human + AI in Accounting: Early Evidence from the Field, Journal of Accounting Research (2026), combined a survey of 277 professional accountants with field data from an AI-enabled accounting platform serving 79 small and medium-sized businesses, covering more than 200,000 transaction-level records. They report that GenAI adoption is "associated with significant productivity gains and systematic reallocation of effort away from routine data entry toward business communication and quality assurance tasks." These are associations from one platform's clients, not a controlled estimate for every firm. The direction still matters for operations: as more is produced, checking becomes a larger share of the job.

This programme does not have a reliable general error rate for AI on client tax or audit work, and it will not invent one. Every rate in this module is either illustrative or one your own firm measures.

A question to sit with

A senior who checks every AI-drafted reconciliation note finds a problem about once a month and fixes it. Next year the firm runs the same method across 120 entities with nine preparers. How many problems a month should the firm now expect, who will find them, and how would anyone know if they stopped being found?

What goes wrong

Volume turns a rate into a count. A method that needs a correction in a small share of outputs feels safe when it runs a few times a month. Run thousands of times a month, the same share is a steady stream of wrong outputs. Nothing about the method changed except how often it runs.

Handoffs strip away context. The person who built the method knew which entities have unusual intercompany accounts and which bank feeds drop the last day of the month. The ninth preparer to use it does not. Each handoff, from builder to team, preparer to reviewer, reviewer to the client's controller, passes on the output and loses some of the knowledge needed to judge it.

Shared inputs make errors travel together. Nine people each working their own way make different mistakes, and colleagues notice different things. Nine people using one shared prompt, one account-mapping table or one template make the same mistake in every entity at once. Consistent outputs then look like quality when they are really a common cause.

Checking fades when it finds nothing. A reviewer who checks two hundred notes and finds two problems learns, reasonably, that checking rarely pays. They check faster, then spot-check, then skim. The step is still in the procedure and no longer happening. Lesson 2 looks at why this happens to experts as well as beginners.

Silent errors raise no flag. Most operations count exceptions: items the tool could not match, fields it left blank. A wrong answer delivered confidently is not an exception. It passes through every counter the operation has, which is why an exception rate is not an error rate.

Ownership spreads until nobody holds it. The innovation lead built it, the vendor configured it, staff ran it and a manager signed a summary. Ask who owns the method and each points to another. Lesson 4 deals with naming the owner.

The practical answer

Start by separating the risks that scale multiplies from the risks it creates. They need different controls.

Risk Present in one careful task? What scale does Control that addresses it
An invented figure or reference in an output Yes Multiplies it Verification against source, as in Level 2
Client data sent to an unapproved tool Yes Multiplies it Tool approval and access control
Preparers running different versions of the method Barely Creates it A written procedure with one owner (Module 4)
Verification quietly stops Barely Creates it Sampling and quality assurance (Module 13)
One defect in a shared input reaches every client No Creates it Version control and staged rollout
Nobody owns the operation No Creates it A named owner (Lesson 4)

Rows three to six are what this level means by operations. A better prompt fixes none of them.

Do the frequency arithmetic before rollout. Take the correction rate your firm has actually measured, multiply it by the planned volume, and write the answer as a sentence.

If that sentence is unacceptable, the question is not whether the tool is good. It is which control changes the number: a narrower task, a heavier review gate, a smaller rollout or a different step in the workflow.

Ask four questions of any task you plan to scale. These are this programme's recommendation.

Question Why it matters If the answer is worrying
How often will it run? Frequency converts a rate into a count State the expected monthly count of wrong outputs
Who will run it? More people means more variation Write the procedure down before widening use
How far is the reviewer from the source? Review needs evidence, not summaries Keep a sampled path back to source documents
What do all the runs share? Shared inputs cause correlated failure Put every shared prompt, table and template under version control

Treat anything shared as a controlled asset. A prompt used by nine preparers is not a personal note, and an account-mapping table used for 120 entities is not a spreadsheet on someone's desktop. Best practice: give every shared prompt, template, mapping table and tool configuration a version number, an owner, a change record and a way to go back to the last good version.

Widen in stages and watch between them. A defect found while five entities use a method costs five reworks. The same defect found at 120 entities costs 120 reworks, plus the work of finding out which of the 120 were affected and which version each used.

Map before you manage. Professional guidance (voluntary framework): NIST's AI Risk Management Framework (AI RMF 1.0, January 2023) says that, after instituting its GOVERN outcomes, "most users of the AI RMF would start with the MAP function" and continue to MEASURE or MANAGE. It is voluntary and binds no firm, CPA or preparer, and NIST has said it is revising 1.0. The point for operations is simple: understand how the work actually runs before deciding how to control it. Module 2 builds that map.

Remember this

  • Scale multiplies some risks and creates others: version drift, fading verification, correlated error and unowned operations.
  • Convert your measured correction rate into an expected monthly count, and write it down before rollout.
  • An exception rate is not an error rate; confidently wrong output raises no flag.
  • Every shared prompt, mapping table and template needs a version, an owner and a way back.
  • If honest review cannot keep pace, change the operation rather than the sign-off record.

Accounting example

Fenwick Lowe CPAs is a fictional 60-person firm with a client accounting services group. Senior accountant Marisol Tan builds a method for AI-drafted explanations of bank reconciling items. For three months she runs it on 6 clients, traces every explanation to the bank statement and ledger, and corrects the prompt twice. Her manager, Dev Kapoor, signs each reconciliation. The method works, so the group rolls it out to all 120 clients with nine preparers. Six months later, two things have happened that nobody decided. First, the prompt has forked: most preparers use the version Marisol emailed in week two, while three use a copy a colleague edited to handle credit card clearing accounts. Nobody recorded which clients used which. Second, the tracing has thinned. The procedure still says to agree every explanation to source, but in busy weeks most preparers check the larger items and skim the rest. In month seven a reviewer notices that one explanation describes a deposit in transit that cleared two months earlier. Checking further, Dev finds the edited prompt tells the tool to carry forward last month's explanations where the amount matches. Any recurring amount has kept its old explanation. The defect sits in the reconciliations of every client that used the edited version, about 40 clients, across four closes. Dev asks two questions: which clients used the edited prompt, and in which months. Nobody can answer either from the records. The team rebuilds the answer from email dates and preparer memory over two days. Marisol's method was sound. The operation around it had no version control, no sampling of the tracing step and no named owner.

Responsible-use warning

Do not treat a successful pilot as proof a method is safe to scale. A pilot run by the person who built it, on clients they know, with checking they do themselves, tests the method under the one condition that disappears at scale. Version drift, fading verification, a shared defect reaching every client and unowned outcomes are hidden in a pilot because one careful operator suppresses all four. Ask what breaks at nine preparers and 120 clients before rollout.

Human-review requirement

The qualified professional who owns each output still owns it at scale: the licensed CPA for attest work, the signing tax return preparer for a return, or the responsible controller for a company's own books. At the operations level that person also judges two further things: whether the review actually performed still matches the review the procedure describes, and whether the volume is within what the named reviewers can genuinely examine. If it is not, their job is to narrow the scope, add reviewers or pause the rollout, not to sign a record that overstates what was checked.

Practical exercise

Use a fictional or anonymised method only.

  1. Choose one AI-assisted task you or your team perform, such as reconciling-item explanations, variance commentary or document extraction. Describe the steps as they are actually done, in no more than one page.
  2. Record the current volume in outputs per month, and the volume if your whole team or practice adopted it.
  3. From your last 20 reviewed outputs, count how many needed a correction. Calculate the rate, multiply it by the projected volume, and write the result as one sentence.
  4. List every shared input: prompts, templates, mapping tables, checklists, tool settings. Beside each, write who owns it and how a change would be recorded today.
  5. Identify the steps that exist only to catch errors. For each, write what evidence, other than asking people, would show that the step had stopped happening.
  6. Name the shared input whose defect would reach the most clients at once, and estimate how many clients that would be.
  7. Using the risk table in this lesson, mark each risk in your method as multiplied or created by scale.
  8. Write three questions you would need answered before widening use, and note which later module of this level addresses each.

Knowledge check

  1. Which risks does scale multiply, and which does it create?

    Answer guidance: It multiplies risks already present in one careful task, such as an invented figure or client data sent to the wrong tool. It creates risks that barely exist in one task: preparers running different versions, verification quietly stopping, one shared defect reaching every client, and an operation nobody owns.

  2. A pilot found 3 corrections in 200 outputs, and the firm plans 3,600 outputs a month. What should the rollout memo say?

    Answer guidance: The rate is 1.5%, and 3,600 × 0.015 = 54, so the memo should say the firm expects about 54 outputs a month needing correction and that the review design must be able to find them.

  3. Why is an exception rate not an error rate?

    Answer guidance: Exceptions are items the tool itself could not handle, such as unmatched or blank items. A confidently wrong output raises no exception, so it passes through every exception counter. Only checking outputs against source reveals it.

  4. What did the 2024 PCAOB staff Spotlight observe about GenAI use, and what are its limits?

    Answer guidance: Staff observed that use by the firms interviewed appeared focused primarily on administrative and research activities, and firms said the team member using the tool remains responsible for the results and documentation. It is staff views from limited outreach to mainly larger firms in 2024, not a Board rule and not a picture of 2026 practice.

Sources and further reading