Where the profession actually is
Talk to accountants privately and you get two answers. First, quietly: "AI is doing more of my first-pass work than I let on publicly." Second, more carefully, in earshot of clients: "We’re evaluating it." Both are true. The profession is adopting AI tools faster than the official positioning suggests, and more cautiously than the vendor marketing suggests.
This guide is the version a working CPA would want: what’s actually being used, what’s still too risky, and the preprocessing steps that make the difference between AI that gives you useful answers and AI that gives you confident-but-wrong answers on a client file.
High-value use cases in public accounting
Financial statement analysis
The single most productive use of AI in accounting work today. Feed a client’s income statement, balance sheet, and cash flow statement to an AI with instructions like: "Compare the current period to prior year, flag any line item that moved by more than 15%, group the drivers by category, and suggest analytical procedures that would test the largest variances." The AI produces a structured variance analysis in minutes that would have taken a senior associate an afternoon.
The essential preprocessing step: convert PDF statements to clean markdown first. Bank PDFs and older accounting system exports routinely have merged cells, multi-row headers, and page-break artifacts that mangle direct AI extraction. Use the financial reports converter and verify the table structure before analysis. See financial statements to markdown for the detailed walkthrough.
Tax preparation — the data-entry half
A typical 1040 with rental, self-employment, and investment income involves pulling data from 15 or 20 source documents — W-2s, 1099s, K-1s, brokerage statements, property tax bills. That’s roughly 70% of the preparation time and roughly 0% of the professional judgment. AI is rapidly automating that half.
Current state at mid-size firms: prep software vendors (Intuit, CCH, Thomson Reuters) are shipping AI-assisted source document extraction. Smaller firms are building lighter-weight versions with general AI tools — upload source docs, ask for a structured extract, paste into the prep workflow. The second half of the return — the elections, the apportionment, the planning — stays with the preparer. Which is exactly where the value is.
Client onboarding and engagement letters
The new-client questionnaire, the initial information request list, the engagement letter from a template — all of this is high-volume, low-judgment work that AI writes well. Give it the client’s industry, entity type, and service scope; get a tailored first draft back. Review and send. The firm capacity freed up goes to actually advising clients.
Audit documentation cleanup
Memos, workpaper narratives, control descriptions — the prose wrapper around audit evidence — is a place AI can reliably add polish. Feed it the rough-draft bullet-point workpaper and ask for a cleaned-up narrative in your firm’s house style. It’ll produce something publishable in 30 seconds that would have taken 15 minutes.
This is cleanup, not substantive work. The AI is not making audit judgments; it’s helping you write down the ones you already made. Verify that the tightened narrative didn’t accidentally change the substance — AI will sometimes "improve" language by softening or hardening a conclusion. Read carefully.
Client-facing explanations
Explaining the depreciation schedule to a small-business owner. Walking a nonprofit board through an unfavorable variance. Turning "the cost basis allocation under §1060" into a paragraph a client can follow. AI writes plain-English finance exceptionally well. Draft with AI, fact-check the technical pieces, send. Client relationship quality goes up, your time on the "what does this mean?" call goes down.
Where the profession is still cautious
Anything with client-identifying information on consumer AI
The AICPA’s position, broadly: general-purpose consumer AI chats don’t meet the confidentiality standard for client data. A free ChatGPT upload of an unredacted financial statement isn’t that different from forwarding it to a random email address — same information leakage, same lack of vendor controls. Business and Enterprise plans with data processing agreements meet the bar for routine work; for highly sensitive data (valuations, forensic matters, anything with trade secrets), the industry is moving toward purpose-built tools or on-prem deployments.
Independence-sensitive engagements
If you’re auditing a client, you can’t build their internal AI tool for them and then audit what it produces — the independence rules still apply to AI advisory work. The quieter problem: if the client is using AI that your firm’s parent vendor also sells them, you may have sourcing disclosure obligations. Check with your risk management team; the guidance is evolving.
Professional judgment calls
Is this a contingent liability that needs to be accrued? Is this lease an operating lease or finance lease? Is this revenue recognizable now or deferred? AI will answer these questions. The answer will often sound reasonable and will sometimes be wrong in ways a partner would spot in two seconds. Use AI to structure the analysis, pull out the relevant authoritative standards, and draft the memo. Don’t let it conclude.
The math problem
Everyone discovers this eventually. You paste a trial balance into ChatGPT, ask "what are total assets?" and get an answer that’s close but not right. This is because the underlying language model is guessing at sums, not computing them.
The fix is structural. For arithmetic, use the tool that can actually do arithmetic — Excel, a database, or AI’s code interpreter mode (ChatGPT’s Advanced Data Analysis, Claude’s code execution). When you need an AI to work with numbers reliably:
- Upload a spreadsheet, not a PDF, when possible.
- Ask the AI to write Python (or use code interpreter) to compute the answer, not to compute it directly in chat.
- Verify the computed result by eye against the source data.
This is the difference between "AI got the wrong number" and "AI wrote a three-line script that got the right number." The latter is reproducible. The former is expensive.
A realistic audit workflow: analytical procedures
Concrete example. Manager running analytical procedures on a $50M manufacturing client.
- Prep. Export the current year and prior year trial balance from the client’s accounting system. Convert any PDF comparatives to markdown tables.
- Variance calculation. Use Excel or Code Interpreter (not chat-based AI) to compute absolute and percentage variances by account.
- Analysis. Feed the variance table into Claude (firm deployment) with the prompt: "Flag all variances exceeding materiality of $125,000 or 10%, whichever is lower. For each flagged variance, suggest two analytical procedures and identify potential drivers based on typical manufacturing industry patterns."
- Review. Manager reviews AI output, adjusts for client-specific knowledge (e.g., the AI doesn’t know the plant closure in August), and drafts the memo.
- Document. Include prompt used and AI output as a workpaper exhibit. Firm’s AI policy requires this, and it’s good practice regardless.
Time saved: roughly half the manager-hours on the analytical procedures step. Quality maintained because the AI is doing structured flagging, not substantive conclusion-making.
Privacy and the three-tier approach
Think about client data in three tiers and match the AI tool to the tier.
| Data tier | Examples | Appropriate AI tool |
|---|---|---|
| Public | SEC filings, IRS pubs, case law, your own CPE | Any AI, including free consumer |
| Internal | Firm procedures, templates, training | Business plans with DPA |
| Client-confidential | Financial statements, tax returns, M&A data rooms | Enterprise plans with zero retention, or purpose-built accounting AI |
A common mistake: using a "good enough" tier-appropriate tool for most work, and then absent-mindedly pasting client data into the free consumer version when in a hurry. Firms with clear policy and muscle memory around which browser tab holds which tool avoid this. The ones that don’t often find out the hard way.
Getting started: a 30-day plan for a small firm
- Week 1. Pick one enterprise AI platform (ChatGPT Team, Claude for Work, or an accounting-specific vendor). Sign the DPA. Get credentials for partners.
- Week 2. Update engagement letters to disclose AI use. Document an internal AI policy — which data tier, which tool, what gets logged.
- Week 3. Train staff on three initial workflows: financial statement analysis, engagement letter drafting, client communication cleanup. Skip tax prep for now (too many failure modes to start with).
- Week 4. Measure. Which workflows are saving time? Which aren’t? What’s producing clean output vs needing heavy rework?
After 30 days you’ll have an informed view of what’s working in your practice. That beats a year of theoretical debate about whether AI matters.
Where to read further
The blog post on markdown saving on AI API costs covers the economics of the preprocessing step, which matters more for high-volume workflows. For the hands-on side of getting financial PDFs ready for AI, the financial statements walkthrough is the most directly applicable. And the small business markdown guide covers the bookkeeping side from the business-owner perspective — useful for understanding what your clients are already starting to experiment with on their own.