Lesson objective

Apply a keep-or-cut test to a draft AI session for accountants, keeping only what the behaviour needs and moving everything else to a reference handout.

Plain-language explanation

Why this matters now

Every AI session grows. A history of AI, a diagram of how models work, a tour of the firm's tool, a survey of every rule that mentions technology, a slide of statistics, a case study from the news. Each piece is interesting, and each takes minutes away from the practice that produces the behaviour.

Designing backwards gives you a test for every slide: does the learner need this to perform the behaviour on Monday? If not, it moves to a handout or goes.

Some of what you might be tempted to cut is required understanding for part of your room, so the test needs care.

Professional guidance: IRS OPR Alert 2026-19, 24 June 2026, interpreting Circular 230 section 10.35 for practitioners before the IRS, says practitioners "must understand both the law and the technology used in their representation of clients before the IRS, including AI systems' operational mechanics, limitations, and risks." For that audience, how the tool works is not trivia. But the alert does not say how much mechanism a session must teach, and it creates no new rule.

Professional standard (international): IESBA Code, R320.11 and 320.11 A1, for accountants under the IESBA Code. The factors for deciding whether technology output is appropriate include the ability to understand and explain the technology, testing, prior experience, firm oversight, controls over use, and the appropriateness of inputs and user decisions. US CPAs are bound by the AICPA Code, not the IESBA Code; check the current IESBA Handbook for numbering.

Professional guidance (voluntary): COSO GenAI guidance, February 2026. It names risks including hallucination, prompt injection, opaque reasoning, model drift, rapid configuration changes, bias, vendor opacity and shadow AI. Not every session needs every risk; the capability type in scope decides which matter.

Professional guidance (staff observations, not binding): PCAOB staff Spotlight, July 2024. Firms interviewed said an engagement team member using a GenAI tool "is still responsible for the results and documentation of the work." A mid-2024 snapshot of mostly larger firms, not a rule.

Professional guidance (nonauthoritative): AICPA TQA 400.02, 2 March 2026. This programme read only its landing page. It belongs on a reference handout as something to read, not in a slide describing content you have not seen.

A question to sit with

Which slide in your current deck would you most hate to lose, and can you name the Monday behaviour it serves?

What goes wrong

1. The history lecture. Ten minutes on the origins of AI. It serves no behaviour.

2. The mechanism lecture pitched too deep. Twenty minutes on model architecture for a room that needs to know one thing: the tool can produce fluent, confident, wrong text, including citations that do not exist. Understanding limitations is required for some audiences; a lecture on internals is not the same thing.

3. The vendor tour. Every feature of the firm's tool, clicked through. Features change; the check does not. And a session that ranks or compares products teaches marketing, not practice.

4. The regulatory survey. Every rule, standard and guidance document that mentions technology, for a room bound by three of them. Each unrelated citation dilutes the ones that bind.

5. The statistics slide. Figures with no population, date or caveat, or figures that describe someone else's profession.

6. The cut that removes a caveat. The instructor cuts for time and removes "Texas only" or "guidance, not a rule". A shorter slide that overstates authority is worse than a longer one.

The practical answer

Best practice: apply a four-question keep-or-cut test to every slide and activity.

Question If yes If no
1. Does the learner need this to perform the behaviour on Monday? Keep, and make sure it is practised Go to question 2
2. Is it required understanding for this audience under a source that binds them, such as limitations and risks for practitioners before the IRS? Keep the smallest version that meets it Go to question 3
3. Is it a caveat that stops a rule being overapplied? Keep, and never shorten it away Go to question 4
4. Would a learner want to look it up later? Move to the reference handout Cut

Best practice: teach mechanism only as far as it explains the failure. For most accounting audiences, one explanation does the work: the tool predicts plausible text, so it can produce a citation that looks right and does not exist, or a calculation that reads well and is wrong. That serves the verification behaviour. For practitioners before the IRS, OPR's reading of section 10.35 expects understanding of operational mechanics, limitations and risks; give that audience the limitations and risks of the capability type in scope, and point to the handout for more.

Best practice: keep the rules that bind the room, and only those. A room of tax practitioners needs OPR Alert 2026-19 and SSTS section 1.4. A room of controllers in industry does not need either as binding authority. A room under the IESBA Code needs R320.11; a room of US CPAs does not, though it may be useful context if labelled as international.

Best practice: build a reference handout for everything moved out. The handout carries the fuller source list with each source's label, date and bound population, the capability types not covered, the risks not practised, and a line for TQA 400.02 as a nonauthoritative AICPA document to read. It lets you cut from the session without cutting from the learner's access.

AI Coalition Network recommendation: give practice at least half the session. If a sixty-minute session has fewer than thirty minutes in which learners do the behaviour on fictional material, cut content until it does. It is our recommendation for sessions in this programme, not a finding about how adults learn.

Remember this

  • Test every slide: does the learner need it to perform the behaviour on Monday?
  • Keep required understanding for the audience, such as limitations and risks for practitioners before the IRS, in its smallest useful form.
  • Never cut a caveat: the label, date and bound population stay.
  • Move everything a learner may want later to a reference handout.
  • Give practice at least half the session, as this programme's recommendation.

Accounting example

Olu Brennan, a fictional CPA and learning lead at the fictional Hawthorn Reach Advisory LLP, has a draft deck of 48 slides for a sixty-minute session on verifying AI-assisted tax research. Twenty slides cover the history and inner workings of AI, eleven tour the firm's tool, nine survey rules and guidance, and eight cover the verification behaviour itself. Twenty plus eleven plus nine plus eight makes forty-eight. He applies the four-question test. Of the twenty history and mechanism slides, he keeps three that explain why the tool can produce a confident citation that does not exist, because the room are practitioners before the IRS and OPR Alert 2026-19 reads section 10.35 as expecting them to understand AI limitations and risks. The eleven tool-tour slides go, replaced by a one-page quick reference from the firm. Of the nine rule slides, he keeps four: OPR Alert 2026-19 on section 10.22, SSTS section 1.4, and two slides on what varies by state. The rest, including a slide on the COSO capability types not in scope and a line on TQA 400.02, move to the reference handout with their labels and dates. He keeps all eight behaviour slides. Three plus zero plus four plus eight makes fifteen slides. He checks one shortened slide and restores the words "guidance, creates no new rule" that his cut had removed. The session now gives thirty-five minutes to practice on fictional research notes. He tells the commissioning partner that the handout carries everything cut, so nothing learners might want is lost.

Responsible-use warning

Never keep a slide because it is interesting if no behaviour needs it, and never cut a caveat to save space: a source's label, date and bound population stay on every slide that cites it. Never cut required understanding for an audience, such as limitations and risks for practitioners before the IRS, down to nothing. Never describe TQA 400.02 beyond its landing page, and never rank or compare AI products in a session.

Human-review requirement

The instructor owns every cut and every remaining slide. Before delivery, a second reviewer compares the shortened slides with the originals and confirms that no source label, date, jurisdiction or bound population was lost in the cut, and that the reference handout carries each moved source accurately. The commissioning partner confirms that what remains covers what the firm needs this audience to understand. The signing preparer or engagement owner remains responsible for real work.

Practical exercise

Cut a fictional draft session to its behaviour. Use a fictional firm and a fictional draft deck throughout.

  1. Write a fictional slide list of 30 to 50 slides for a sixty-minute AI session for accountants, grouped into categories such as history, mechanism, tool tour, rules and behaviour. Check that the category counts add up to the total.
  2. Write the one behaviour the session serves, as a statement with who, when, what, standard and record.
  3. Apply the four-question keep-or-cut test to each category, and record the result for each slide.
  4. For the mechanism slides you keep, write the one explanation of failure they must deliver.
  5. For each rule slide you keep, confirm it binds this audience, and write its label, date and bound population.
  6. Shorten one rule slide by at least half without losing its caveat, and show the before and after.
  7. List what moves to the reference handout, with each source's label and date.
  8. Recount the remaining slides, and check that practice on fictional material takes at least half of the sixty minutes.
  9. Ask a fictional colleague to compare your shortened slides with the originals, and record any label, date or bound population they find missing and how you restored it.

Knowledge check

  1. What are the four questions of the keep-or-cut test?

    Answer guidance: Does the learner need it for the behaviour; is it required understanding under a source that binds this audience; is it a caveat that stops overapplication; would a learner want to look it up later.

  2. Why can mechanism not be cut entirely for practitioners before the IRS?

    Answer guidance: OPR Alert 2026-19 reads Circular 230 section 10.35 as expecting them to understand AI systems' operational mechanics, limitations and risks.

  3. What must survive when a slide that cites a source is shortened?

    Answer guidance: The source's label, date and bound population, such as "Texas only" or "guidance, creates no new rule".

  4. Why should a vendor feature tour usually be cut?

    Answer guidance: Features change while the check does not, the tour serves no behaviour, and a session must not rank or compare products.

Sources and further reading