How Can Mid-Size CPA Firms Use AI for Tax Prep Without Preparer Penalty Risk?

Most mid-size CPA firms hit the same wall in busy season. Forty preparers. Twelve hundred 1040s. Three hundred 1065s and 1120-S returns. Each return takes 30 to 90 minutes of pre-prep work before the actual return prep starts: pulling documents from the portal, sorting them by category, building the organizer review, drafting the missing-information email. By March 1 the staff is exhausted, partners are reviewing returns at midnight, and the firm is leaving 80 to 120 returns on the table because there is not enough capacity to take them on.

This is not a content problem. The preparers know the work. The partners know the work. It is a throughput problem: too much administrative assembly sits between the documents arriving and a credentialed preparer making real tax decisions. AI in tax prep is highest value exactly there, on the assembly side. It is lowest value on the final-position side, where a preparer's judgment under Circular 230 cannot be delegated.

The firms that figure this out get 25 to 40 percent throughput improvement in their first busy season running the workflow. Partners spend more time on tax positions. Preparers spend less time hunting for missing 1099s. The firms that do not figure it out either ban AI awkwardly and watch staff use it on personal accounts, or roll out a vendor tool nobody trained on and quietly stop using it by April 15.

This guide walks through the tax prep workflow that holds up under Circular 230, Section 7216, and the preparer penalty rules. It covers where AI actually adds value, where it cannot go, and the workflow split that keeps partners on the right side of the line.

Why this matters for mid-size CPA firms specifically

Mid-size CPA firms (10 to 50 preparers, $5M to $50M in revenue) are uniquely under-served by current AI tooling. The Big Four built internal tools they will not share. The InsurTech-style tax vendors (Blue J, Truewind) are real but expensive. The consumer AI tools were built for marketing, not tax. The result is a market gap exactly at the size where the firm is too big for partner-level workarounds and too small for in-house AI teams.

The firms that figure out the workflow get hours back per preparer per week, the ability to take on 15 to 25 percent more returns without proportional hiring, and a busy-season culture that does not burn out staff by March 15. They also get an audit trail that holds up under peer review and IRS inquiry. Firms that wait usually end up dealing with both the productivity problem and the staff retention problem in the same year.

What AI for tax prep actually does

The useful tools are general-purpose large language models (Claude, ChatGPT, Microsoft Copilot) running on Team or Enterprise tier with a Data Processing Addendum and training on inputs disabled. They take inputs (client documents, prior-year return summary, organizer responses, engagement notes) and produce structured output: missing-information lists, organizer reviews, draft client communications, document categorization.

Three things make these tools different from the AI features bundled into your tax software:

• They handle unstructured inputs. The first-pass document intake from a client portal is a mess: brokerage 1099s, K-1s, mortgage statements, charitable acknowledgments, all in one ZIP. The AI reads all of it and produces a structured summary that maps to the organizer.
• They produce consistent output across preparers. You define the firm's organizer review format once. The AI produces it the same way every time. This matters for partner review because consistency is reviewable.
• They iterate in plain English. "Add a section flagging any 1099-DIV with foreign tax paid over $300, since that triggers Form 1116 considerations" is a one-sentence edit, not a re-train.

Think of it as a senior staff accountant who has read every document in the file, never gets tired in busy season, and produces output in the firm's exact format every time, while a credentialed preparer makes every tax position decision.

Before you start

You need:

• A Claude Team, ChatGPT Team or Enterprise, or Microsoft Copilot for Microsoft 365 account with training-on-inputs disabled and a signed DPA. $25 to $60 per seat per month.
• About 60 minutes for your first session, mostly to build the firm's prompt template library.
• A Section 7216 written consent process. Your tax counsel or AICPA can provide the boilerplate. Non-negotiable before any client tax data leaves the firm's controlled environment.
• Read access to the firm's tax software (CCH ProSystem fx, Lacerte, UltraTax, Drake) and the document management system (CCH Axcess, GoFileRoom, SurePrep 1040SCAN, or whatever the firm runs).
• A partner and the firm's CCO, peer reviewer, or quality control partner on a 30-minute call to review the prompt templates and the workflow before any client return uses it.

One thing to settle before any tax return information goes into an AI tool: Circular 230, Section 7216, Section 6713, the preparer penalty rules, and AICPA SSTSs all shape what AI can and cannot do here. We have a dedicated section on this below. It is non-negotiable.

Material 1: First-pass document intake and categorization

The failure pattern: a staff accountant spends 20 to 40 minutes per return categorizing documents the client uploaded as a 30-file ZIP. They sort 1099s into a folder, K-1s into another, mortgage statements somewhere else, and build a checklist by hand. The work is mechanical, error-prone, and the same accountant could be doing actual return work instead.

What to ask the AI for instead:

Below is a list of file names and a one-paragraph description of each document the client [Client Last Name] uploaded for the 2025 tax year. Categorize each document into one of the firm's standard categories: W-2, 1099-INT, 1099-DIV, 1099-B, 1099-R, K-1 (1065), K-1 (1120-S), K-1 (1041), Mortgage Interest Statement, Property Tax Statement, Charitable Contribution Documentation, HSA Documentation, Education Documentation, Child Care Documentation, Foreign Account Documentation, Other (with explanation). Produce output as a table with three columns: file name, category, notes. Notes column should flag anything unusual (a 1099 from an account we did not see last year, a K-1 from a new entity, anything that looks like a draft or amended document). Do not invent categories. If a document does not match a category, mark it Other and explain in the notes.

The categorization is a low-risk first use because no tax position is being decided. The AI is doing administrative sort work. The preparer reviews the table, fixes any miscategorization, and the rest of the return prep starts from a clean, structured starting point instead of a raw ZIP.

For firms running SurePrep 1040SCAN, the AI step happens before the SurePrep auto-classification or as a parallel check. The two systems do similar work and the firms that run both end up with very few miscategorizations slipping into the return.

Material 2: Organizer review and missing-information list

The organizer review is the second most repetitive task in busy season. The preparer reads the prior-year return, reads the current-year organizer responses, reads the documents that came in, and builds a list of what is missing or unclear. That list becomes the missing-information email to the client. Most preparers do this in 25 to 45 minutes per return. Most of that time is mechanical comparison work.

Below are three inputs: (1) the firm's organizer responses from [Client Household] for tax year 2025, (2) a summary of documents received from the client, (3) a summary of the prior-year (2024) return as it sits in the firm's tax software. Build me an organizer review with two sections. Section A: Apparent items, things the firm has documentation for that map to organizer responses. Section B: Missing or unclear items, things where the organizer indicates an item but the documentation is missing, or where the prior year had an item the current year does not. For each Section B item, write one sentence framing it as a question for the client. Frame the questions politely and specifically. Do not characterize the client's responses as incomplete or careless. Where the prior year had an item that does not appear this year, ask whether the situation changed (sold the rental, closed the brokerage account, retired) rather than assuming the item was forgotten.

The preparer reads the output, edits anything that needs editing, and the missing-information email to the client is essentially drafted. The preparer adds the firm's standard sign-off and sends it. What used to take 30 minutes takes 8.

The Section A and Section B split is the move that makes the output useful. The preparer can scan Section A in 30 seconds (everything that lines up) and focus attention on Section B (where the work is). Without the split, the output is a paragraph that the preparer has to re-organize anyway.

For 1065 and 1120-S returns, expand the prompt to include a section comparing the K-1 detail to the prior year's K-1 detail and flagging changes in ownership percentage, capital account, or guaranteed payments that need explanation before the return goes any further.

Material 3: Prior-year comparison and significant-change flagging

The failure pattern: a return gets prepped from the current year forward without anyone looking at the prior year until partner review. The partner notices a $40,000 swing in dividend income or a missing rental property and the return goes back for rework. Three days lost. Two preparers' time burned.

The AI handles prior-year comparison fast.

Below are two summaries: the current-year (2025) draft return for [Client Household] and the prior-year (2024) filed return. Produce a comparison report with three sections. Section A: Items that changed materially, defined as a change of more than 25 percent in any line that totals over $5,000, or any new line item over $5,000, or any prior line item now zero that was over $5,000. Section B: Items that changed in expected ways tied to documented life events (the client mentioned a job change, a property sale, a retirement). Section C: Items that changed in unexpected ways with no documented explanation in the firm's notes. For each Section C item, write one sentence framing it as a question for the preparer or client. Do not project what the answer might be. Just flag the question.

The Section C items are the ones that matter. They are the items that, left unresolved, become the partner's questions during review. Surfacing them at the staff level means the partner spends review time on tax positions, not on "why did dividends drop $30K?"

For multi-state returns, add a section comparing state-by-state allocations to flag any new state filing requirement or any state with a meaningful change in apportionment. Multi-state issues are the ones most likely to slip through review and become preparer-penalty exposure later.

Material 4: Draft client letters and tax planning memos

Client cover letters and year-end planning memos are the highest-volume client communication a tax practice produces. Most firms have a stable set of templates that get manually personalized for each client. Most of that personalization is the same five sentences with the client's name and circumstances swapped in.

The AI handles this part well, with the constraint that any tax-position language has to come from the firm's approved phrasing, not from the AI's defaults.

Below are two inputs: (1) a summary of [Client Household]'s 2025 return as it sits in the firm's tax software, including key planning items, (2) the firm's standard cover letter template. Draft a personalized cover letter that follows the firm's template structure exactly. Personalize the planning paragraph with three to five items specific to this client's situation. Use only the firm's approved planning language, which I will paste below. Do not introduce planning ideas or tax positions that are not in the firm's approved language list. Do not characterize the return as straightforward, complex, or any other adjective. Do not project tax outcomes for future years. Sign with the partner's name as it appears on the engagement record.

The constraint that protects the firm: the AI cannot generate tax planning ideas. It can only restructure and personalize the firm's pre-approved language. If the client's situation calls for planning advice the firm has not yet templatized, the preparer adds that paragraph manually and the partner reviews it. The AI never invents tax advice.

For year-end planning memos, same pattern. The firm has 30 to 60 standard planning topics it knows how to address. The AI matches the client's situation to the relevant subset of those topics and assembles the memo. The partner reviews. Anything the partner wants to add, the partner adds.

Material 5: K-1 reconciliation and entity workpapers

K-1 reconciliation is the workflow inside the workflow. Each K-1 needs to flow correctly into each partner's 1040. When the same firm prepares both, it is a 30-minute manual exercise per K-1. When different firms prepare them, it is 60 minutes and the source of frequent errors.

Below are the K-1s issued by [Entity Name] for tax year 2025 and the prior-year K-1s. For each partner, produce a reconciliation worksheet with: (1) Income items with prior-year comparison and percentage change, (2) Deduction items, same comparison, (3) Capital account roll-forward, reconciled, (4) Guaranteed payments and self-employment income flagged separately, (5) Items the partner's 1040 preparer needs to consider: passive vs. active classification, basis limitations, at-risk limitations, Section 199A QBI eligibility. Do not give specific 1040 advice. Surface the items. Mark items that flow through automatically versus items that require the 1040 preparer's judgment.

The reconciliation is workpaper-quality output. It goes into the firm's workpaper file. The preparer's sign-off is what makes it a tax position. For 1041 trust returns with multiple beneficiaries, the same pattern applies with DNI tracking and simple-vs.-complex trust treatment.

Material 6: Year-end planning calls and engagement letter renewals

Year-end planning calls are the moment in the calendar where a tax practice either sells the next year of value or loses ground to the firm down the street. Prep is similar to advisory meeting prep: read the file, build the agenda, surface the questions, draft the talking points.

I have a year-end tax planning call with [Client Household] on [date]. Below are the firm's notes, the most recent return as filed, and any planning topics flagged in the engagement record. Build a one-page brief with: (1) Two sentences on the client's situation as of year-end, (2) Three planning topics worth addressing, drawn from the firm's approved planning topics list, (3) Two questions about life events or business changes the firm may not know about, (4) Three open items the firm needs from the client to start prep, (5) A short script for setting expectations on busy-season turnaround. Do not project specific tax outcomes. Use the firm's approved planning language.

Engagement letter renewals follow the same pattern. The AI takes the firm's standard letter, the client's prior-year scope, and any changes for the new year, and produces a personalized renewal the partner reviews and signs. The two together compress what used to be a two-hour exercise per client into 25 minutes of AI-assisted drafting and partner review.

The CPA-specific prompts that actually work

After watching mid-size firms run AI through a full busy season, the difference between useful output and generic output comes down to four prompt moves.

Specify the engagement type and the client situation. "1040 with K-1s from two passthrough entities, one rental property, and a Schedule C, MFJ, two dependents under 17, AGI roughly $480K" lands very differently than "a 1040." The AI calibrates the depth and the planning topics to the actual situation.

Specify the constraint that actually matters. For tax prep the constraints are: do not invent tax positions, do not project specific outcomes, do not characterize the return with adjectives, use only the firm's approved planning language. State them explicitly. Implicit constraints get violated. Explicit ones do not.

Specify the firm voice and the document format. Paste a sample of an existing client letter or planning memo in the firm's actual voice. The AI matches it. Without the sample you get the AI's generic accounting voice, which reads like a CPE handout.

Specify what stays static and what changes. For recurring documents (cover letters, planning memos, organizer review emails), tell the AI what is fixed (firm disclosures, engagement language, standard sign-off) and what is variable (this client's situation, this client's planning items). The structure stays the same across clients. The content scales.

The Circular 230, Section 7216, and preparer penalty non-negotiables

This section is short because the rules are simple, but it is the most important section in this guide.

Do not put any of the following into the consumer tier of any AI tool:

• Client SSNs, ITINs, or EINs
• Specific tax return information without Section 7216 written client consent
• Trust documents, estate plans, or insurance policy details with identifying information
• Bank account numbers or financial account numbers
• Any document where the client's name is paired with a tax outcome (refund amount, tax due, AGI, tax liability)
• Anything that could identify a client by name and tax situation

Use AI for templates, workflow structure, prompt patterns, and de-identified examples that you build on the Team or Enterprise tier with a Data Processing Addendum signed and training on inputs disabled. For client tax data, the firm needs Section 7216 written client consent before any tax return information goes into an AI tool, even an Enterprise tier with a DPA. Section 7216 is not waived by a DPA. It is a separate consent requirement under the Internal Revenue Code.

Circular 230 Section 10.22 (due diligence) and Section 10.34 (standards for advice) require that the credentialed preparer exercise judgment on the tax positions. AI cannot exercise that judgment. The preparer can use AI to assemble information, draft communications, and produce workpapers, but the tax positions and the signed return are the preparer's responsibility. The audit trail (prompt template, AI output, preparer review) is what proves the diligence if the firm gets a notice or a peer review inquiry.

If your firm has signed an Anthropic Business agreement, an OpenAI Enterprise agreement, or a Microsoft 365 Copilot agreement with a DPA, the rules can be different on the data-handling side. Section 7216 still applies separately. Ask your tax counsel and your malpractice carrier what is covered. Document the workflow in your firm's quality control policies. Do not assume.

When NOT to use AI in tax prep

AI is a generalist tool. It will not be the right answer for every part of a tax practice's workflow.

Skip it for:

• Final tax positions and signed returns. The AI can surface the question. The credentialed preparer makes the call. The partner signs. Always. No exceptions, no shortcuts, no exceptions for "easy" returns.
• Direct client-facing AI interactions about specific returns. Letting the AI draft an email is fine. Letting the AI answer client questions about their return crosses preparer penalty, due diligence, and Section 7216 lines. Those tools require a different compliance posture entirely.
• Tax research where the AI cannot show its sources. AI tools that produce tax research without citation to actual statute, regulation, or guidance are unsafe to rely on. The AI can summarize Blue J or Checkpoint output. It cannot replace it.
• Anything that requires sign-off you do not have authority to give. Engagement scope changes, fee changes, scope of work for new services. The partners decide. The AI does not.

A simple rule: AI is an unfair advantage on the 80% of tax prep where the work is assembly. Trust the credentialed preparer's judgment and the partner's review for the 20% where the document or the position has Circular 230 weight.

The quick-start template

Here is the prompt scaffold that works across most CPA tax-prep use cases. Copy it, fill in the brackets, paste into Claude Team or ChatGPT Team after Section 7216 consent is in place.

Build me a [document categorization / organizer review / prior-year comparison / draft client letter / K-1 reconciliation / planning call brief] for [Client Household / Entity] for tax year [year].

Inputs: [paste de-identified or consent-covered tax data, prior-year summary, organizer responses, document list].

Output structure: [list 3 to 6 named sections with one sentence on each].

Voice and format: [paste a sample of how the firm writes the same document for another client].

Constraints: Do not invent tax positions. Do not project specific outcomes. Do not characterize the return with adjectives. Use only the firm's approved planning language, which I will paste below. Where the inputs are silent, mark blank rather than infer.

Output format: [plain text for workpaper, formatted for client communication, draft email, etc.].

For recurring documents, save the populated template once and reuse the structure. The client-specific inputs change. The structure does not.

Bigger wins beyond document intake

Once the busy-season prep workflow is running, the next layer of value shows up in places that affect the firm's economics, not just throughput.

Firm-wide template library with QC sign-off. Build a library of approved prompt templates: 1040 organizer review, K-1 reconciliation, 1120-S basis tracking, year-end planning brief, engagement letter renewal. Each gets QC sign-off once. Every preparer uses it. Compliance review happens at the template level.

New client intake automation. AI handles the first-pass document review, prior-firm return analysis, and planning topic identification faster than any process change. The first meeting goes from "tell me about your situation" to "here is what we noticed in your prior return."

Internal training acceleration. The firm's institutional knowledge lives in partner heads. AI can take the firm's training materials, planning memos, and historical communications and make them searchable. New staff ramp faster. Partners answer fewer repetitive questions.

Off-season advisory expansion. Firms that compress busy season find themselves with capacity in May and June that did not exist before. That capacity becomes the foundation for advisory services (CFO services, tax planning, succession) where margins are higher than 1040 prep.

The financial services AI consulting connection

This is one tool in one category of a CPA firm's workflow. The bigger question for the financial services niche is structural: client expectations are shifting toward year-round advisory relationships and away from once-a-year compliance work, while the regulatory frame around AI in regulated practices is tightening. Firms that figure out the workflow end up with preparers who get hours back, partners who can take on more advisory work, and a quality control posture that holds up under peer review and IRS inquiry.

If your firm is wrestling with the bigger AI question, the AI Consulting in Financial Services page covers the full scope: where AI actually fits in CPA firms, RIAs, wealth management firms, and insurance brokerages, what the common failure modes look like in regulated practices, and what an engagement looks like when it works.

For individual partners at mid-size firms, start with this guide. Pick one of the six workflows and build the prompt template this week. Run it on three returns. Compare the prep time before and after. The case for the rest of the workflow makes itself once the partners see the throughput change.

Closing

The goal is not for preparers to become prompt engineers. It is for preparers to spend their judgment hours on tax positions instead of administrative assembly. AI is the closest tool I have seen to that goal for CPA firms specifically. It rewards specificity, respects the regulatory frame, and gives back the hours that disappear into busy-season triage.

Pick one workflow. Build the template. Run it on three returns this month. See the throughput change.

If you want to talk about how AI fits into your firm at the practice level, the AI Consulting in Financial Services page lays out the full picture and how an engagement works.