Lesson objective

Recognise when adding AI to a public company's financial reporting process may be a change in internal control that management must evaluate, and record step-level volumes, times and exceptions on the map so later improvement claims can be tested.

Plain-language explanation

Why this matters now

A mapped workflow is the right moment to ask two questions that are easy to skip. First, for a public company, does the proposed change touch internal control over financial reporting in a way management must evaluate? Second, what does each step cost today, so anyone can later check whether AI changed it?

Law / rule: SEC Rule 13a-15 (17 CFR 240.13a-15) binds issuers with a class of securities registered under section 12 of the Securities Exchange Act. Among its requirements:

  • 13a-15(a): every such issuer "must maintain disclosure controls and procedures."
  • 13a-15(b): management, with the principal executive and financial officers, must evaluate the effectiveness of disclosure controls and procedures as of the end of each fiscal quarter.
  • 13a-15(c): management's annual ICFR evaluation must be based on "a suitable, recognized control framework."
  • 13a-15(d): management must evaluate any change in ICFR during each fiscal quarter that "has materially affected, or is reasonably likely to materially affect" ICFR.
  • 13a-15(f) defines ICFR as a process designed by or under the supervision of the principal executive and financial officers to provide reasonable assurance regarding the reliability of financial reporting.

The rule does not mention AI. Whether introducing an AI tool into a reporting process is a change requiring evaluation under 13a-15(d) is a matter of applying the rule to the facts, not something the rule states. That judgement belongs to management, with its advisers.

Professional guidance (SEC interpretive guidance): the SEC's 2007 guidance says ICFR "cannot provide absolute assurance due to its inherent limitations" and involves human diligence subject to lapses and failures. That is a useful reminder when a vendor or an internal champion describes an AI tool as removing error.

A question to sit with

A public company's controller says the new AI tool "just helps the team prepare reconciliations faster; the controls haven't changed." Looking at the map, what would you check before accepting that sentence?

What goes wrong

Assuming a tool is not a control change because the checklist did not change. If a tool now prepares what a person used to prepare, or routes items differently, the design of the control may have changed even if its name has not.

Deciding too late. The change goes live mid-quarter, and the question of evaluation comes up only when quarter-end certifications are being prepared.

Only thinking about ICFR for public companies. Rule 13a-15 binds issuers. Private companies and CPA firms are not bound by it, but the same map questions help them judge their own control changes, and lenders or auditors may ask.

Claiming improvement with no step-level data. Module 1 set operation-level baseline measures. Without step-level timings and exception counts, nobody can tell whether a faster close came from the AI step, a different client mix, or a senior who worked weekends.

Borrowing published figures as your result. Research (observational): Choi and Xie (Journal of Accounting Research, 2026), using data from one AI-enabled platform serving 79 small and medium-sized businesses, report GenAI use "associated with significant productivity gains" and with "more granular ledgers and faster month-end closing." Those are associations from one platform. They are not evidence about your process.

The practical answer

Screen each mapped AI use for ICFR relevance. For a public company, AI Coalition Network recommendation: add a column to the placement table from Lesson 2 and answer these questions for each use. The answers go to the people who make the 13a-15(d) judgement; they do not make it.

Screening question If yes
Does the use sit at a step that produces or records a figure in the financial statements or disclosures? Flag as ICFR-relevant
Does it change who or what performs a control, or what a reviewer sees? Flag as a possible change in control design
Does it add, remove or bypass a review point, or change a routing threshold? Flag as a possible change in control design
Does it depend on IT general controls that are new or not yet evaluated? Flag the IT general controls for evaluation
Will it go live during a fiscal quarter? Record the date so the quarterly evaluation can consider it

A flag means "send to management's ICFR evaluation process," not "this is a material change." Materiality and the evaluation are management's.

Remember what the framework rule asks. 13a-15(c) requires a suitable, recognised control framework; many issuers use COSO 2013. COSO's 2026 GenAI publication adapts that framework's principles to GenAI and lists "External auditors assessing GenAI-related controls" among its intended users. It is voluntary and not authoritative for SEC or PCAOB purposes, but it gives management and auditors a shared vocabulary.

Time each step while you map. AI Coalition Network recommendation: add volume and time to every step row during the walk-through and exception trace, over at least one full cycle.

Step-level measure How to collect it Why it matters
Items per cycle Count from the system or the checklist Shows where volume sits
Minutes per item, typical Observation or time entries for a sample of items Shows where effort sits
Exceptions per cycle Count items that left the normal path Shows where the process strains
Minutes per exception Trace a sample of exceptions Exceptions often cost far more than normal items
Corrections found at the next review point Count from review notes, using a fixed definition Shows where errors originate
Waiting time before the step Time an item sits before work starts Distinguishes slow work from slow handoffs

Do the step arithmetic. Multiply items by minutes for each step and add exception time. The result shows where the hours go, which is often not where people expect.

Freeze definitions and record limits. Use the same definition of an exception and a correction before and after any change. Note what the cycle did not represent, such as a quarter-end or a new client.

Track return on the tool honestly. The Thomson Reuters Institute report (vendor research, February 2026, all professional-services respondents combined) found only 18% of respondents knew their organisation was tracking AI return on investment in some manner. Step-level data is how a team becomes able to track it.

Remember this

  • Rule 13a-15(d) requires issuers' management to evaluate quarterly any ICFR change that has materially affected, or is reasonably likely to materially affect, ICFR; the rule does not mention AI.
  • Screen each mapped AI use for ICFR relevance and send flags to management's evaluation; the screen does not decide materiality.
  • ICFR cannot provide absolute assurance, with or without AI.
  • Record items, minutes, exceptions and corrections for every step during mapping.
  • Exceptions often hold the hours that AI does not remove; do the arithmetic before promising savings.

Accounting example

Corvane Systems Inc. is a fictional SEC-registered software company. Its controller, Leah Morrow, plans to introduce an AI tool in the second fiscal quarter to prepare first drafts of balance sheet account reconciliations, with senior accountants reviewing. Using the map, the team screens the proposal. The tool sits at steps producing reconciliations for accounts in the financial statements: ICFR-relevant. It changes who prepares the reconciliation and what the reviewer sees: possible change in control design. It introduces a routing rule sending reconciliations with unreconciled differences under $500 straight to review without preparer investigation: possible change in control design. The tool's settings access has not been evaluated: IT general controls flagged. Go-live is planned for day 20 of the quarter. Leah sends the screen to the chief accounting officer and the SOX team, who own the 13a-15(d) evaluation. Separately, the team times the current process over one close. Step 4, prepare reconciliations: 80 accounts at 35 minutes each, 2,800 minutes. Step 4 exceptions: 12 accounts with unexplained differences at 90 minutes each, 1,080 minutes. Step 4 total: 3,880 minutes, about 64.7 hours. Step 7, review: 80 accounts at 15 minutes, 1,200 minutes, or 20 hours. Reviewers record 9 accounts needing a required change, 9 ÷ 80 = 11.25%. The vendor suggests the tool will "halve reconciliation time." Leah's memo notes that halving normal preparation saves 1,400 minutes, about 23.3 hours, but only if exception time and review time do not rise, and that the same measures will be taken after go-live using the same definitions.

Responsible-use warning

Do not accept "the controls haven't changed" because the checklist looks the same. If a tool now performs part of a control, changes what a reviewer sees, or routes items differently, the control's design may have changed. For an issuer, send that to management's quarterly evaluation early, not at certification time. And do not promise savings from the normal path alone: exceptions and review often hold the hours a tool does not touch.

Human-review requirement

For an issuer, management, with the principal executive and financial officers, owns the evaluation of ICFR changes under Rule 13a-15(d) and decides whether a change is material. The controller or SOX lead confirms the screen was completed for every AI use and sent before go-live. For any organisation, the responsible professional checks that step timings were collected with fixed definitions over a representative cycle, that the arithmetic is right, and that limits are recorded beside the numbers.

Practical exercise

Use a fictional company and the map from earlier lessons.

  1. State whether your fictional organisation is an issuer bound by Rule 13a-15, a private company, or a CPA firm, and what that means for the ICFR screen.
  2. Apply the five screening questions from this lesson to each AI use on your map, and record yes or no with one sentence each.
  3. For each flagged use, write the date it would go live and which fiscal quarter that falls in.
  4. Write a two-sentence note sending the flags to the people who own the ICFR evaluation, making clear the screen does not decide materiality.
  5. For at least four steps on your map, record items per cycle, minutes per item, exceptions per cycle and minutes per exception, using fictional figures.
  6. Calculate total minutes for each step, showing the multiplication and addition.
  7. Take a vendor claim, such as "cuts preparation time in half," and recalculate the saving for one step assuming exception time stays the same.
  8. Write the definitions of an exception and a correction that you will keep unchanged after any redesign.
  9. List two limits of your one-cycle timing that should be recorded beside the numbers.

Knowledge check

  1. What does Rule 13a-15(d) require, and does it mention AI?

    Answer guidance: It requires management of an issuer to evaluate any change in ICFR during each fiscal quarter that has materially affected, or is reasonably likely to materially affect, ICFR. It does not mention AI; whether an AI tool is such a change is a matter of applying the rule to the facts.

  2. What does a flag from the ICFR screen mean?

    Answer guidance: It means the AI use should go to management's ICFR evaluation process. It does not mean the change is material; that judgement belongs to management.

  3. A step has 1,200 normal items at 0.5 minutes and 60 exceptions at 6 minutes. If AI halves normal time only, what is the saving?

    Answer guidance: Before: 600 + 360 = 960 minutes. After: 300 + 360 = 660 minutes. The saving is 300 minutes, or 5 hours.

  4. Why can Choi and Xie's findings not stand in for step-level data?

    Answer guidance: They report associations from one platform serving small and medium-sized businesses. They cannot show where time goes in a particular process or whether a particular change helped, which only that process's own measurements can show.

Sources and further reading