Quick answer (updated August 2026): AI fits a small CPA firm best in five back-office workflows: client communications, engagement letters, billing narratives, research summarization, and workpaper documentation. It does not belong anywhere near opinions or signed work. Before adopting anything, apply the data gate: IRC §7216 makes knowing or reckless unauthorized disclosure or use of tax return information a federal crime, and AICPA Rule 1.700.001 bars sharing confidential client information without specific consent. Practical translation: no client data in consumer AI tools, ever.

Most “AI for accountants” articles list tools. Almost none of them mention that CPAs work under two client-data rules with real penalties attached, one of them criminal. This guide covers the rules first, in plain English, then the workflows where the hours actually come back. It is informational, not legal advice; run your firm’s specific setup past counsel or your liability carrier.

Where do a small firm’s hours actually go?

Adoption is broad and shallow. The Federal Reserve Banks’ 2025 Small Business Credit Survey (6,525 employer firms) found 46% of small employer firms use AI, but just 7% of users have fully integrated it, and the top use is writing and marketing at 83%. CPA firms mirror that pattern: someone drafts an email with ChatGPT, and the firm’s real time sinks stay untouched.

Those sinks are rarely the tax engine. Your software already calculates. The hours go to the writing and explaining wrapped around the work: client status emails and document-chasing, engagement letters each season, billing narratives that justify WIP, summarizing research into client-ready memos, and workpaper documentation. Every one of these is drafting work, and drafting is where the evidence for AI is strongest. A randomized experiment published in Science (Noy and Zhang, 2023) found professionals completed mid-level writing tasks 40% faster with AI assistance, with 18% higher rated quality.

What stays off the table: audit opinions, tax positions, final review, and anything you sign. AI drafts; a CPA decides.

The data gate, part 1: what does IRC §7216 actually say?

Section 7216 of the Internal Revenue Code is a federal criminal statute aimed at tax return preparers. In plain English: if you prepare returns for compensation and you knowingly or recklessly disclose information a client gave you for return preparation, or use that information for any purpose other than preparing the return, you have committed a misdemeanor. Under the statute (26 U.S.C. §7216), conviction carries a fine of up to $1,000 per violation, up to a year in prison, or both, plus costs of prosecution. The fine rises to $100,000 where the identity-theft provision of §6713(b) applies.

A companion civil penalty, 26 U.S.C. §6713, adds $250 per improper disclosure or use, capped at $10,000 per calendar year ($1,000 and $50,000 in identity-theft cases).

Two things make this bite for AI. First, “tax return information” is broad: essentially anything the client furnished in connection with the return. Second, under the Treasury regulations summarized in the IRS’s Section 7216 FAQs, “use” includes any circumstance in which a preparer refers to or relies on tax return information to take or permit an action, and Regulation 301.7216-3 requires written taxpayer consent before disclosure or use unless a specific exception applies.

The data gate, part 2: the AICPA confidentiality rule

The second rule reaches further. AICPA Code of Professional Conduct Rule 1.700.001 states: “A member in public practice shall not disclose any confidential client information without the specific consent of the client.” The code defines confidential client information as any information obtained from the client that is not available to the public. That covers everything §7216 covers, plus bookkeeping data, payroll, CAS engagements, and advisory work with no tax return in sight.

The code’s third-party service provider guidance is the part that maps directly onto AI vendors. Before disclosing confidential client information to an outside provider, the code directs members either to obtain specific client consent or to have a contractual agreement binding the provider to confidentiality, with reasonable assurance the provider has procedures to prevent unauthorized release.

Read that against a consumer AI account and the problem is obvious. A free chatbot account gives you no contract, no confidentiality terms you negotiated, and no assurance about where the data goes. A business-tier subscription with signed terms is a different animal. That distinction is the whole basis of the tier rule below. State boards of accountancy layer their own confidentiality rules on top, so check yours.

What client data can touch which AI tool?

This is the same three-tier rule we apply across professional services firms, tuned for the two rules above.

Tier Tool class Client data allowed Why
1. Consumer Free or personal-plan chatbots (personal ChatGPT, consumer Claude) None. Ever. Consumer chats can be used for model training and have no negotiated confidentiality terms. Fails the AICPA third-party test outright.
2. Business, training off Business tiers with no-training defaults and signed terms (ChatGPT Team/Enterprise, Claude for Work) Non-return client data: engagement letters, communications, billing narratives OpenAI’s enterprise privacy page (updated January 2026) states business data is not used for training by default; Anthropic’s commercial terms (June 2025) bar training on customer content. The contract can satisfy the AICPA guidance. Confirm with counsel.
3. Specialist / zero-retention Tax-specific platforms and zero-data-retention deployments Tax return information, if at all §7216’s consent-or-exception requirement applies to return data no matter how good the tool’s security is. Tier controls disclosure risk; §7216 restricts use itself.

Two cautions. Consumer chat logs are also a legal-discovery surface: a May 2025 preservation order in the New York Times litigation required OpenAI to retain consumer output logs for a period (the going-forward hold was lifted in October 2025), while zero-retention API traffic was never covered, per OpenAI’s own account. And in United States v. Heppner (S.D.N.Y. February 2026), a federal court held a defendant’s consumer-tier chatbot conversations were neither privileged nor confidential, precisely because the consumer terms let the vendor retain the chats, train on them, and disclose them. Lawyers hit this same wall; the parallel analysis is in our ChatGPT-for-legal-work guide. And tier 3 is deliberately labeled “if at all”: many small firms simply keep return data out of general AI entirely and lose little, because the highest-return workflows below barely need it.

Workflow by workflow: where the hours come back

Each prescription below runs at tier 2 or lower. None requires tax return information in a prompt.

Client communications. Status updates, document-request chasers, and answers to routine questions (“what do you need from me?”, “when will my return be ready?”) are template work. Build a library of AI-drafted templates once, then personalize per send from your own knowledge, not from pasted return data.

Engagement letters. Start from your attorney-approved base letter. AI drafts the variations: entity type, scope, new services. The letter’s variables are engagement facts, not return data.

Billing narratives. Turning time entries into readable WIP descriptions is pure drafting. Reworded entries like “reviewed depreciation schedules and drafted client memo” carry no return figures.

Research summarization. Summarize authorities, IRS publications, and public guidance into plain-English client explainers. Public sources in, client-ready drafts out; the facts of the client’s situation get added by you, in your own systems.

Workpaper documentation. AI turns your bullet notes into clean narrative documentation of procedures performed. Keep identifying details generic in the prompt and complete the specifics in the workpaper itself.

What is that worth in hours?

For scale: the St. Louis Fed’s survey research (February 2025) found workers who use generative AI report saving an average of 2.2 hours in a 40-hour week, mostly without deliberate workflow design. A firm that targets specific drafting workflows should beat an incidental-use average; the table stays conservative anyway. The math below is an illustration with stated assumptions, not a promised result: a three-to-five person firm, hours rounded down.

Workflow Typical weekly time (assumed) With AI drafting Hours back
Client emails and follow-ups 5 hrs 3 hrs 2
Engagement letters (seasonal average) 1 hr 0.5 hr 0.5
Billing narratives 2 hrs 1 hr 1
Research summarization 2 hrs 1.5 hrs 0.5
Workpaper documentation 3 hrs 2 hrs 1
Total 13 hrs 8 hrs 5

Five hours a week, at a conservatively assumed $200 realized rate over 48 weeks, is roughly $48,000 of capacity a year, again as an illustration. That capacity either becomes billable work or becomes a busy season that ends before your staff does. Finding your firm’s actual numbers, workflow by workflow, is precisely what an AI opportunity assessment does. We sell no software and take no commissions, so the prescriptions have no thumb on the scale.

When should a firm roll this out?

Not in February. Busy season is the worst time to change how work gets done and the most dangerous time for the data gate, because tired staff under deadline pressure reach for whatever tool is fastest, including personal chatbot accounts the firm never approved. The U.S. Chamber of Commerce and Teneo’s 2025 survey of 3,870 small businesses found 58% now use generative AI, up from 23% two years earlier; assume some of your staff are already among them, on their own accounts.

The sequence that works: pick one workflow in May or June, pilot it with one or two people, and write the one-page data policy at the same time (which tools are approved, at which tier, with which data). Add workflows through the fall. Train everyone before January, including the explicit list of what never goes into a prompt. By the time the crunch hits, the approved path should also be the easiest path. That is what makes a policy hold at 9 p.m. on April 12th.

Bottom line

AI can conservatively return about 5 hours a week to a small CPA firm across client communications, engagement letters, billing narratives, research summaries, and workpaper documentation, all without tax return information in any prompt. The gate comes first: §7216 (criminal, up to $1,000 per violation and a year in prison) and AICPA Rule 1.700.001 mean no client data in consumer AI tools, business tiers with training off and signed terms for non-return data, and specialist or zero-retention deployments, with consent where required, for return data, if at all. Pilot off-season, write the policy before January. For the broader method, start with our small business AI guide.

Sources