A pre-season AI session for an accounting firm has four jobs and about seven usable hours to do them in: establish which tools client information may touch, teach a verification habit that survives a 70-hour week, convert two or three recurring tasks on the firm’s own documents, and set the rule for when someone stops and asks a person. Everything else on a vendor’s agenda is May work.
That subtraction is the decision. Run a long, catalog-shaped curriculum in November and February arrives with staff who have seen eleven topics and practiced none.
The number that should reframe the conversation
The profession already measured its own training habit, and the result is the strongest argument for a structured session that exists. In the 2026 tax software survey published by the Journal of Accountancy, 1,808 AICPA members who prepared 2025 returns for a fee answered questions about the software they use daily. Only 41.2% had received any training on that software. The same survey found 65% using AI for tax research and 32% for client communication, with the share reporting no plans to use AI down to 16% from 24.3% a year earlier.
Read those together. Most preparers were never trained on the compliance software the firm licenses, and most are now using general-purpose assistants on research and correspondence. The firm did not acquire an AI problem this year. It acquired a second untrained tool, and this one drafts prose that reads as authoritative whether or not it is correct. That is why a course link is not an answer, and why capability is something a firm builds rather than buys.
Run it in November, not January
A session has to land close enough to the season that the material is still warm, and far enough out that there is one low-stakes cycle to practice on. For most firms that is late October through early December: extensions are finished, planning work is in flight, and mistakes are still caught by a reviewer.
January fails both tests. Training in the second week of January asks people to learn a new working method in the week they are also absorbing software updates, new staff, and the first organizers. Nothing sticks, and the day is remembered as the one nobody billed. May through August is where the long curriculum belongs: advisory use cases, workflow redesign, anything that asks a firm to change a process rather than a habit.
One constraint worth planning around: whoever runs the session should work from the firm’s own documents, because a workshop built on a stranger’s engagement letters teaches tool mechanics and nothing else.
Block one: the client-data perimeter
This is the hour that cannot be cut, and the one every course catalog reduces to a slide reading “be careful with client data.” Two rulebooks apply at once, and they do not say the same thing.
The tax rulebook is criminal. IRC section 7216 makes it an offense for a return preparer to knowingly or recklessly disclose or use tax return information outside permitted purposes: a fine of up to $1,000, up to a year in prison, or both, per violation, plus a civil penalty of $250 per disclosure under section 6713, capped at $10,000 a calendar year. Treasury Regulation section 301.7216-1(b)(5) defines disclosure as making the information known to another person “in any manner whatever” — which a paste into a third party’s chat window may well satisfy. Consent, where that is the path, must be knowing, voluntary, written, obtained in advance, and for 1040-series returns in the language prescribed by Revenue Procedure 2013-14.
The professional rulebook offers a choice. Interpretation 1.700.040 of the AICPA Code of Professional Conduct lets a member either enter a confidentiality agreement with reasonable assurance the provider has appropriate procedures, or obtain the client’s specific consent — a choice, not a pair of requirements. The client should be told, preferably in writing, before confidential information reaches the provider, which is why the engagement letter is where most firms resolve it.
The output of the hour is a one-page list, posted where people work, in three columns: tools inside the perimeter and what may go into them, tools outside it and therefore prose-only, and the information that never leaves the firm’s own systems. A business-tier account with terms the firm has read sits in column one. A personal account someone signed up for in March does not.
Write the policy and skip the session and you get the gap described in our guide to the distance between an AI policy and safe practice: a policy nobody has practiced against is a document, not a control.
Block two: verification as a standards obligation
In June 2026 the IRS Office of Professional Responsibility published Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice. It created no AI-specific rules. It mapped AI use onto duties practitioners already carry under Circular 230: section 10.22 due diligence, 10.35 competence, 10.36 firm procedures and supervision, 10.37 written advice. AI may assist professional judgment and may not replace it, and due diligence cannot be delegated to an algorithm.
That turns verification from a soft recommendation into something a firm can be examined on, and it tells you what to teach: the failure modes, produced live. Ask an assistant a question with a confidently wrong answer available — a repealed provision, a threshold that changed, a cite that nearly exists — then have the room check it against the authority. People who have watched a fabricated citation arrive in clean prose stop treating formatting as a quality signal. People who were only warned do not.
Then make the standard concrete: every number traced to a source document, every authority opened and read, every piece of client-facing language reviewed by the person whose name goes on it. Section 10.36 makes that the firm’s obligation, so the session ends with the review step written into the workpaper checklist.
Block three: three workflows, converted in the room
Generic agendas list use cases. A pre-season session converts a few of them end to end, on the firm’s own material, and leaves the prompts saved where colleagues can find them. Pick from the work the profession has already ranked. The Thomson Reuters Institute’s 2026 AI in Professional Services Report found organization-wide adoption nearly doubling to 40%, roughly a third of tax firms using generative AI, and tax research leading the use-case ranking at 69%, ahead of bookkeeping at 60% and document summarization at 59%. Data quality is its largest reported barrier for the second year running — the quiet argument for starting with text rather than numbers.
| Candidate workflow | Why it fits | What to produce in the room |
|---|---|---|
| First-pass research memo | High volume, text-based, reviewed anyway | A template plus the verification checklist that follows it |
| Client correspondence and organizer chasing | Repetitive, low judgment, measurable time sink | Three drafts for the most common January emails |
| Document summarization for review | Cuts reading load without touching numbers | A prompt that cites page or section for every claim |
| Workpaper and review-note drafting | Captures reviewer intent in plain language | A worked example from last season’s notes |
Three is the number: two is thin, four stops fitting in the practice time. Which three depends on the seat — a tax senior, an audit associate, a bookkeeper, and a firm administrator do not share a curriculum, the argument our role-based enablement playbook sets out in full and the format trade-off we walk through for a team of sellers.
Note what is absent: anything that produces a filed number without a human computing it independently.
Block four: the escalation rule
The last twenty minutes are the cheapest insurance in the agenda, and almost nobody runs them. Write down the three or four situations where someone stops and asks a person: the engagement letter restricts third-party processing; the question involves a position the firm has not taken before; the output contradicts the source document; a client asks whether AI touched their return.
That last one is routine now — the Thomson Reuters data has 74% of corporate tax clients saying their outside firms should use AI. The firm needs one agreed answer on scope, review, and data handling, delivered the same way by whoever picks up the phone. Drafting it in November costs half an hour. Drafting it on March 10 costs more.
What to leave out until May
The subtraction list is where a firm buys back time to practice. Defer:
- Model architecture and AI history. It opens most catalog courses and changes nobody’s Monday.
- Agentic workflows and automation builds. About 14% of tax firms report using agentic AI, and the ones doing it well built on an established base. Staff who are not yet fluent cannot supervise an autonomous process.
- Tool selection and platform comparison. Teach the one or two assistants the firm already pays for. A comparison exercise produces opinions, not skills.
- Advisory service-line design. Real opportunity, wrong season.
- Firm-wide process redesign. The wrong risk in the quarter before work moves fastest.
The test for any agenda item: if it will not change what someone does in the first week of February, it is not November material.
What CPE a hands-on session actually earns
Settle this before the invoice. Credit depends on the sponsor, not on the quality of the teaching. Under NASBA’s Statement on Standards for Continuing Professional Education Programs, the sponsor is responsible for meeting the standards, issuing certificates, and keeping records, and sponsors are approved per delivery method — group live, group internet based, self-study, nano learning, blended. All self-study and nano-learning credit must come from a sponsor on the National Registry of CPE Sponsors, and state boards add their own rules. A hands-on session from a consultancy that is not a registered sponsor generally earns no CPE.
Split the purchase. Satisfy the credit requirement with institute or state-society courses built for exactly that — the AICPA’s Core Concepts of Artificial Intelligence for Accounting Professionals is a 2.5-credit self-study course at $95 for nonmembers and $79 for members, Surgent’s two-credit Introduction to AI in Accounting runs $109 to $119, and the AICPA and CIMA’s AI Accelerator Skills Program, announced in April 2026, bundles three tiers for up to 42 accredited credits. Then buy the capability separately. Prices verified October 2026; catalogs move quarterly, so confirm before budgeting.
What it costs against the alternatives
Three options, three different things purchased.
| Option | Market range | What it buys | CPE |
|---|---|---|---|
| Self-serve courses | $75 to $400 per person | Credit hours and vocabulary; transfer to the firm’s work unmeasured | Yes, from Registry sponsors |
| Hands-on firm session | $2,000 to $15,000 | Practiced capability on named workflows, plus a written perimeter | Usually none |
| Custom automation project | $25,000 to $150,000 | A built process, once people are fluent enough to supervise it | None |
Run the arithmetic against what it replaces rather than against zero: a ten-person tax group reclaiming half an hour a day each through February and March is a simple calculation at the firm’s own realization rate. We work that through for a comparable team in our breakdown of what AI training for a sales team costs. An in-house finance function faces a different version of the decision — different calendar, rulebook, and buyer — which our guide to training options for finance and accounting teams covers.
What to do in the next two weeks
- Settle the perimeter on paper. List every assistant in use, personal accounts included, and mark each in or out against interpretation 1.700.040 and the section 7216 consent position.
- Check the engagement letter for third-party processing and AI assistance. If neither is addressed, that is a conversation with whoever maintains the template.
- Pick the three workflows, from work people already repeat, with last season’s examples ready.
- Book the courses for credit and the session for the method before Thanksgiving.
Frequently asked questions
What should AI training for accountants cover before busy season?
Four blocks: which tools client information may touch and under what consent, a verification habit practiced against deliberately wrong answers, two or three recurring workflows converted on the firm’s own documents, and a written rule for when someone asks a person. Model theory and tool comparisons can wait.
When is the best time to run an AI session for a tax team?
Late October through early December. Extensions are done, planning work gives people real material to practice on at low volume, and it is still fresh in February. January training competes with software updates, new staff, and the first organizers.
Does a hands-on AI workshop count for CPE?
Usually not. Credit depends on the provider being a NASBA Registry sponsor approved for the relevant delivery method, with state board rules on top, and all self-study and nano-learning credit must come from a Registry sponsor. Satisfy credit through institute or state-society courses and buy the working session separately.
Can we put client tax return information into ChatGPT or Claude?
Not without resolving two questions first. Treasury Regulation 301.7216-1(b)(5) treats making tax return information known to another person “in any manner whatever” as a disclosure, and section 301.7216-3 consent must be written, specific, and in the prescribed format for 1040-series returns. Interpretation 1.700.040 separately requires either a confidentiality agreement with the provider or the client’s consent. Business-tier accounts with reviewed terms are where most firms land; personal accounts are not.
What did the IRS say about AI for tax practitioners?
In June 2026 the Office of Professional Responsibility published Alert 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice. It added no new rules and applied existing Circular 230 duties — 10.22 due diligence, 10.35 competence, 10.36 firm procedures, 10.37 written advice — holding that AI may assist professional judgment but not replace it.
How much does AI training for an accounting firm cost?
Institute and marketplace courses run roughly $75 to $400 per person — the AICPA’s Core Concepts course is $95 for nonmembers, Surgent’s introduction $109 to $119. A delivered hands-on session runs roughly $2,000 to $15,000 at market rates, and custom automation is a separate $25,000 to $150,000 decision.
Should partners and managers attend the same session as staff?
They need a different one. Staff are learning to produce with an assistant; partners and managers are learning to review work an assistant touched, answer the client question about AI use, and own the section 10.36 supervisory obligation. A combined session teaches production and skips review.
How many people should be in the room?
Small enough that everyone works on their own material and gets unstuck, which in practice means 8 to 15. Beyond that it becomes a demonstration, and recorded courses do those more cheaply.
What if we already sent everyone a course link last year?
Check what changed in the work. Completion data tells you who clicked, not who can do the job, and one graded exercise on a real deliverable settles that faster than any dashboard.
Key takeaways
- Four jobs: the client-data perimeter, verification practiced rather than described, two or three workflows converted on the firm’s own documents, and a written escalation rule.
- Only 41.2% of the 1,808 preparers in the Journal of Accountancy’s 2026 survey had any training on the tax software they use daily, while 65% already use AI for research. The deficit is older than the technology.
- IRC section 7216 carries criminal exposure up to $1,000 and a year per violation, and Treasury Regulation 301.7216-1(b)(5) reaches disclosure “in any manner whatever.”
- IRS OPR Alert 2026-19 maps AI use onto Circular 230 sections 10.22, 10.35, 10.36, and 10.37, and 10.36 makes supervision the firm’s obligation.
- CPE and capability are two purchases: Registry-sponsored courses earn credit, a working session on real documents changes the work. Schedule it by Thanksgiving.
If you want a view on which of your tools sit inside the perimeter and which three workflows are worth converting before January, book a free AI readiness call. We will say what a session should contain for your seats.
Arthur Wandzel
