What Should AI Handle vs. What CPAs Must Review

What Should AI Handle vs What Tax Professionals Should Review: The Question Behind the Question

Every CPA firm evaluating AI tax preparation software eventually runs into the same question, and it's rarely "which tool should we buy?" It's what should AI handle vs what tax professionals should review — a labor-division problem, not a shopping decision. Firms that skip past it end up with either an expensive tool nobody trusts or a review process so loose it creates real professional-responsibility exposure. This piece gives you a concrete rubric — anchored in Circular 230 and IRS due-diligence standards — for drawing that line task by task, plus a checklist you can adapt to your own return mix.

Why This Decision Matters More Than Which AI Tool You Pick

Firms shopping for tax prep software for professionals tend to compare feature lists: OCR accuracy, integrations, price per return. Those matter, but they're secondary. The primary question is structural — which parts of the preparation workflow can move to AI without moving the firm's risk profile with it?

Here's the fact that doesn't change no matter which platform you use: Circular 230 attaches its due-diligence duties to the practitioner, not the software. Section 10.22 requires diligence as to accuracy in preparing returns. Section 10.34 sets standards for the positions a preparer can take. Section 10.37 governs written advice. None of these sections carve out an exception for "the AI did it." If a return goes out with an error, the IRS and state boards look at the signing preparer's diligence, not at a vendor's marketing copy.

That's the framing behind every decision in this article: AI prepares and analyzes. The CPA reviews, decides, and signs. Everything below applies that sentence to specific tasks on a 1040, 1065, 1120, or 1120-S.

The Core Framework: 4 Questions to Score Any Tax-Prep Task

Rather than debate AI tax preparation in the abstract, score each task in your workflow against four questions. This works whether you're deciding about W-2 data entry or a shareholder basis calculation.

Q1 — Is the task deterministic, or does it require judgment? Transcribing a W-2 into Form 1040 wages is deterministic; there's one right answer, printed on the document. Deciding whether a shareholder's distribution should be recharacterized as compensation is judgment — it depends on facts and industry norms that don't reduce to a formula.

Q2 — Does the task carry professional-responsibility exposure under Circular 230? Sorting documents by client carries no independent liability. Tax positions, entity classification calls, and aggressive deduction claims directly implicate the "reasonable basis" and "substantial authority" standards under Section 10.34.

Q3 — Is the source data structured, or ambiguous? A 1099-DIV is a structured, government-formatted document with defined boxes. A client email explaining "I started renting my basement in March, and my brother helped me put in a bathroom" is narrative — it requires a human to translate facts into tax treatment.

Q4 — What's the downstream cost of an undetected error? A Schedule B total off by a rounding error is low-cost. Missing a passive-activity loss limitation, or misapplying a partner's outside basis, can mean penalties, amended returns, and malpractice exposure.

Scoring rubric: Assign 1–3 points per question (1 = leans toward automation, 3 = leans toward human judgment), then add the four scores.

Total score	Zone
4–6	Automate — AI handles it, spot-check on review
7–9	AI-Assist — AI drafts/flags, human confirms before it moves forward
10–12	Human-Only — AI should not make the call, even as a first pass

Run this exercise once across your firm's standard workflow — 1040 intake through diagnostics through final review — and you'll have a task map more useful than any generic "AI vs. human" debate.

What AI Should Handle (High-Confidence Automation Zone)

These tasks score low (4–6): structured data, deterministic rules, low judgment content, and low cost if a rare error slips through, because it's easy to spot-check against the source document.

Document intake and data extraction. Pulling wage and withholding figures off a W-2, dividend and interest amounts off 1099-DIV/1099-INT, gross proceeds and basis off 1099-B, and box-by-box K-1 data into structured format. This is the highest-volume, lowest-judgment task in the entire tax season.

Populating forms and schedules from extracted data. Mapping clean data onto Schedule B, Schedule D, and Form 8949 transaction-level import is mechanical — the IRS form structure defines exactly where each number goes.

Reconciling broker statements against prior-year basis. Matching this year's 1099-B against last year's carryforward basis schedule, flagging wash sales, and identifying missing cost-basis lots is tedious, rules-driven work AI handles faster and more consistently than a tired preparer at 9 p.m. in March.

Running preliminary diagnostics and flagging missing information. Before a human opens the file, AI can flag "no Form 1095-A attached but Schedule 2 shows premium tax credit reconciliation is needed" or "Schedule C shows vehicle expenses but no mileage log referenced." This surfaces issues; it doesn't resolve them.

Drafting workpapers and book-to-tax adjustment schedules. For 1120, 1120-S, and 1065 returns, building the Schedule M-1/M-3 adjustment workpaper — depreciation differences, meals add-backs, tax-exempt interest — from trial balance data is largely mechanical once mapping rules are set.

Prior-year return comparison and variance detection. Comparing this year's draft against last year's filed return and flagging material swings — income jumped 40%, a dependent disappeared, a large Schedule E loss appeared — is pattern-matching AI performs well.

Repetitive data entry across Schedule C, E, and SE. Populating recurring rental properties, standard business expense categories, and self-employment tax calculations from structured client data is squarely automation territory.

Every one of these tasks shares a trait: source data is structured, mapping rules are published by the IRS, and a preparer can verify the output quickly against the source document. That verifiability is what makes automation safe.

What Tax Professionals Must Review (Mandatory Judgment Zone)

These score 10–12. AI should not make these calls autonomously, even in draft form, without a clearly framed human decision point before anything moves forward.

Reasonable compensation determinations for S corp shareholders. IRS factors here — training, experience, duties, time devoted, comparable industry pay — are facts-and-circumstances, not extractable from a document. AI can surface compensation benchmarks; the determination belongs to the preparer.

Partner and shareholder basis calculations. Outside basis tracking involves contributions, distributions, allocated income/loss, debt basis under Section 752, and at-risk limitations that compound year over year. A misclassified distribution can cascade through years of returns.

Ambiguous income characterization. Hobby versus business under the Section 183 factors, passive versus active participation under Section 469 — applying a multi-factor legal test to a client's actual conduct is not something a document-extraction engine can observe.

Nexus and multi-state allocation judgment calls. Whether a remote employee creates nexus in a new state, how to source service revenue, and how to apply each state's apportionment formula are legal-interpretation questions layered on ambiguous fact patterns.

Related-party transactions and step-transaction issues. Loans between a business and its owner, property transferred below fair market value, and multi-step transactions require a preparer to ask whether the form matches the substance.

Positions requiring Section 10.34 "reasonable basis" analysis. Any time a return takes a position not obviously supported by clear guidance, the preparer needs to document the authority relied on. That's a professional judgment call, full stop.

Final sign-off before the firm files. Regardless of how much of the return AI prepared, a human preparer with signing authority reviews the completed return before it goes out the door. This is non-negotiable under Circular 230.

The Gray Zone: AI-Assisted, Human-Confirmed Tasks

Not everything sorts cleanly into "automate" or "human-only." A meaningful chunk of the workflow scores in the middle (7–9) — AI does real work here, but a preparer has to close the loop before it counts as final.

Diagnostics that surface potential issues. AI flagging "this Schedule E shows a loss with no at-risk statement" is useful. Resolving whether the at-risk rules limit the loss requires a preparer to look at the client's investment and liability structure.

Suggested deductions or credits needing client verification. AI might flag a plausible home-office deduction based on Schedule C activity, or estimate vehicle business-use percentage from prior-year mileage patterns. Neither is safe to include until the preparer confirms the underlying facts — square footage, actual usage logs, exclusive-use requirements.

AI-drafted client questions for missing documentation. Drafting the "we're missing your mortgage interest statement" email saves time, but someone should scan the batch before it goes out, especially for sensitive or complex client situations.

Capital gains categorization when holding period or basis is unclear. AI can flag "this lot shows no reported basis" or "acquisition date suggests short-term but confirm inherited-property step-up," but a preparer needs to verify facts like inheritance dates or gifted-property basis carryover before locking in the categorization.

This gray zone is where most firms actually spend their AI-adoption energy, and it's worth treating as its own category rather than forcing it into "automate" or "don't automate."

Building a Human-in-the-Loop Review Process