AI Training for Accountants: How to Keep Staff From Over-Relying on AI

Your staff accountants are already using AI to draft memos, summarize leases and answer research questions. The risk for the firm is a confident wrong answer that reaches a workpaper or a return with nobody checking it. AI training for accountants should start there: verification habits, clear task rules and a review layer that leaves evidence of supervision.

Why do staff accountants over-rely on AI?

Staff accountants over-rely on AI because its output looks finished. A tool that returns a clean reconciliation memo or a confident answer to a basis question invites acceptance, and a first- or second-year accountant has less experience to tell when something is off.

NIST names both halves of the problem in its Generative AI Profile (NIST AI 600-1). It describes confabulation as AI systems that "generate and confidently present erroneous or false content," and it calls excessive deference to automated systems automation bias. Staff accountants sit where those two risks meet: they are close enough to the work to use AI heavily and new enough that a reviewer may assume they checked it.

The consequence for the firm is concrete. An unverified AI number can end up in a workpaper, a return or client advice, and the partner who signs is accountable for it.

What do the professional rules expect when staff use AI?

The professional rules expect the same diligence and supervision for AI-assisted work as for any other work. None of the standards carve out an exception for AI. Three sets of standards matter most to a CPA firm.

• Circular 230. Under 31 CFR 10.22, a practitioner must exercise due diligence in preparing returns and in determining the correctness of representations to clients and to Treasury. Reliance on another person's work product is presumed diligent when the practitioner used reasonable care in engaging, supervising, training and evaluating that person. Section 10.37 requires written advice to rest on reasonable assumptions and a reasonable consideration of the relevant facts. A staff accountant's AI-assisted draft is work product the signer relies on, so the training record matters.
• SQMS No. 1. Firms that perform engagements under the auditing, attestation and review and compilation standards had to design and implement a quality management system under the AICPA's Statement on Quality Management Standards No. 1 by December 15, 2025. The first evaluation is due by December 15, 2026, and evaluations continue annually. Paragraph 33(f) covers the technological resources used to perform engagements, which includes AI tools.
• AICPA tax standards. The revised Statements on Standards for Tax Services, effective January 1, 2024, added standards on reliance on tools and on data protection.

The AICPA's October 2026 practice aid on AI for financial statement auditors is nonauthoritative, and it points the same way: firm policies for AI tools, team competence, and engagement partner oversight of the AI tools staff use.

What should AI training for accountants teach first?

A training program for staff accountants should teach verification habits before tool features. Most AI training starts with prompting. For staff accountants, the more valuable skill is knowing what a correct answer looks like and proving it.

Build the first module around four rules:

• Tie out every number. Any figure AI produces gets traced to the trial balance, a source document or a client-provided schedule before it enters a workpaper.
• Open the authority. When AI cites a Code section, regulation or ruling, staff open the source and confirm it says what the tool claims. An AI summary of a ruling is no substitute for reading it.
• Explain it back. Staff should be able to explain the reasoning in their own words with the tool closed. If they cannot, the work is not ready for review.
• Flag AI use. Staff note in the workpaper or review notes where AI was used and what they checked, so the reviewer knows where to look.

How do you decide which tasks staff can use AI for?

Decide by sorting tasks on two questions: how easy an error is to catch, and how sensitive the data is. A three-tier list works for most firms and fits on one page.

• Open use: drafting emails that contain no client data, summarizing public guidance, building spreadsheet formulas, outlining a research question.
• Use with review: first-pass variance analysis, summarizing leases or loan agreements, drafting review notes, tax research. A senior checks the output against source documents.
• Not permitted: uploading tax return information to any tool outside the firm-approved account, final conclusions on tax positions, and any tool the firm has not approved through its vendor review.

The data line matters as much as the accuracy line. Under Section 7216, a preparer who knowingly or recklessly discloses tax return information for a purpose other than preparing a return faces criminal penalties, and the regulations define disclosure broadly. The AICPA's third-party service provider interpretation (ET 1.700.040) calls for a confidentiality contract with reasonable assurance, or specific client consent, before confidential client information goes to an outside provider. Training should name the one approved tool and explain why personal accounts are off limits. Our guide to keeping client tax documents out of public AI covers that side in more depth.

How should reviewers check AI-assisted work?

Reviewers should check the staff accountant's verification as well as the answer itself. When a workpaper notes AI use, the reviewer confirms that tie-outs exist, that cited authorities were opened and that the conclusion matches the source documents.

Two practices help:

• Reperform with fresh eyes. Each month, a manager picks a few AI-assisted items and reperforms them from source documents before looking at the AI output.
• Track misses. When a reviewer catches an AI error, log the task type and the kind of mistake. The log shows which tasks need tighter rules and which staff need more coaching, and it becomes evidence of supervision.

This review layer is what supports the reliance standard in Circular 230 and the monitoring expectations in SQMS No. 1. It also produces records a peer reviewer can follow.

Does the FTC Safeguards Rule affect AI training?

Yes, for firms that prepare tax returns. The Safeguards Rule requires security awareness training that reflects the risks in the firm's risk assessment, so once AI tools are assessed, AI data handling belongs in that training. Tax preparers are financial institutions under 16 CFR 314.2(h)(2)(viii), and 16 CFR 314.4(e) requires security awareness training "updated as necessary to reflect risks identified by the risk assessment." The firm's written information security plan should also name approved AI tools and the rules for client data in them.

What should a firm do this quarter?

A firm can put the basics in place this quarter, starting with the staff who use AI most and the tasks where an error would reach a client.

• Pick one approved AI tool on a business account whose contract bars training on firm data, and block personal AI accounts on firm devices. Our guide to evaluating AI add-ons in tax software lists the vendor questions to ask.
• Write the three-tier task list and have each staff accountant acknowledge it.
• Run a one-hour verification workshop on a sanitized engagement file where the AI output contains planted errors, and have staff find them.
• Add an "AI used" field to workpaper templates and review checklists.
• Start the reviewer miss log and discuss it at the next quality management meeting.
• Add AI data handling to Safeguards Rule training and the WISP. Our FTC Safeguards Rule checklist for tax preparers covers the rest of the program.
• Record all of it in the firm's SQMS No. 1 documentation if the firm is subject to the standard. Our guide to documenting AI use under SQMS No. 1 walks through what to record.

How can s90 help?

s90 sets up the controls that make this training stick: an approved AI workspace, blocks on personal AI accounts and client data uploads, and the logs that reviewers and peer reviewers need. If you want a second look at your firm's AI rules before the next filing season, we are glad to talk.