AI is reshaping public accounting faster than most firms anticipated, and the professionals adapting aren’t the ones treating it as a novelty. They’re the ones embedding AI tools into actual billable work, audit testing, tax research, client communication, document review, and seeing measurable changes in how long those tasks take.
The gap between “we have an AI subscription” and “we’ve changed how we work” is where most firms are stuck right now. Knowing where AI genuinely helps, where it falls short, and how to sequence adoption is what separates firms using AI as a marketing talking point from those using it to get work done differently.
How AI Is Being Used in Audit Work Right Now
Audit has probably seen the most concrete AI adoption of any service line in public accounting, and the reason is data volume. Traditional audit sampling means reviewing a subset of transactions and drawing conclusions about the whole. AI changes that math entirely.
Tools like MindBridge and Caseware IDEA can analyze entire transaction populations rather than samples, flagging anomalies based on patterns that would take a staff auditor weeks to find manually. MindBridge uses machine learning to score every transaction in a general ledger, surfacing the ones that deviate from expected behavior. A staff auditor working a receivables population of 40,000 transactions might sample 200. The same population run through MindBridge gets every entry scored for risk within minutes.
That doesn’t eliminate professional judgment. It concentrates it. The auditor’s job shifts from “find the anomalies” to “evaluate the anomalies the model found.” Whether that results in better audits depends entirely on whether the auditor doing the follow-up is actually digging into flagged items or just documenting that the tool ran.
Risk assessment and planning are also areas where machine learning is starting to get useful. Predictive models trained on financial ratios, industry benchmarks, and prior-year results can help engagement teams identify where material misstatement risk is highest before fieldwork begins. Some larger firms are building proprietary tools for this. Smaller firms are mostly waiting for vendor-built solutions to mature.
The Sampling Question Nobody Likes to Ask
There’s an open debate in the profession about what full-population testing means for sampling standards. If you’re testing everything, you’re not sampling. The standards haven’t caught up to the technology, and that creates real ambiguity for firms that want to claim AI-driven testing satisfies substantive procedures. This is worth watching as the AICPA issues updated guidance.
AI for Tax Work: Research, Drafting, and Review
Tax is where AI in public accounting has the widest range of applications and also where the failure modes are most dangerous.
Tax research has been genuinely transformed. Bloomberg Tax and Thomson Reuters Checkpoint have both integrated generative AI into their research platforms. Asking a plain-language question and getting a synthesized answer with cited primary sources is now a real workflow. Before these integrations, a tax research question might take two hours: finding relevant code sections, reading rulings, checking reg history, drafting a summary for the file. AI can compress that to 30 minutes when the question is well-scoped and the tool is used properly.
The word “properly” carries a lot of weight there. AI tax research tools hallucinate. They cite cases that don’t exist or misstate the holding of cases that do. The Thomson Reuters AI integration in Checkpoint includes source citations you can click through and verify, which helps. But a tax professional who doesn’t verify the underlying authority is taking on risk that no client engagement letter covers.
Memo drafting is a natural downstream application. Once the research is done, generating a first-draft memo structure using a tool like ChatGPT or Claude is genuinely useful, particularly for junior staff still learning how memos should be organized. The AI output needs review and almost always needs substantive editing, but starting from a draft is faster than starting from a blank page.
Tax return review is an area where accounting automation software is making incremental gains. Tools like SurePrep and CCH iQ read prior-year returns, flag year-over-year variances, check for common errors, and highlight items outside expected ranges. These aren’t replacing review partners. They are, however, catching things that get missed when a reviewer is moving through a stack of returns at 11pm.
Document Extraction and Client Data Management
This is the unglamorous part of AI adoption in public accounting, and possibly the part with the most immediate ROI.
Firms spend enormous amounts of time ingesting client documents: bank statements, trial balances, prior-year tax packages, W-2s, 1099s. Most of this lands in email attachments and client portals, often in formats that require manual data entry. AI-powered OCR and extraction tools can read these documents and populate structured data without manual keying.
Grooper, Docsumo, and the extraction capabilities built into Adobe Acrobat’s AI features are all being used for this. Tax preparation has seen particularly strong adoption. SurePrep’s TaxCaddy product automatically extracts data from uploaded source documents and maps it to return fields, with accuracy rates high enough that the technology has moved from “interesting pilot” to “standard workflow” at a lot of regional firms.
The limitation is unstructured or nonstandard documents. A bank statement from a major national bank extracts cleanly. A hand-annotated PDF of a small business’s bookkeeping notes does not. Setting realistic expectations about where extraction works and where human review is still required is part of implementing these tools responsibly.
AI-Powered Workflows for Client Communication and Advisory
Advisory work creates different AI use cases than compliance work, and the distinction matters. Compliance has right answers. Advisory has better and worse answers depending on client circumstances, and the quality of AI output in that context depends heavily on what context you give it.
Summarizing financial statements and drafting client-ready narratives is something GPT-4 and similar models do well. A partner who has reviewed a client’s year-end financials can prompt an AI model with the key figures, trends, and concerns, then get a first-draft narrative for the management letter. This doesn’t generate the analysis. It generates the words around the analysis the professional already did.
Meeting preparation is another underused application. Uploading a client’s prior financial statements or prior-year return and asking an AI to generate discussion questions or flag areas of concern can cut prep time significantly. Some firms are experimenting with AI tools that integrate with their CRM and automatically pull client context before scheduled calls.
Client email drafting is low-stakes enough that AI can handle it almost unassisted. Explaining a tax position, summarizing an audit finding in plain language, responding to a client question about estimated payments, all tasks where an AI draft with light editing produces a professional result.
How to Choose AI Tools for Accounting Firms
The market for AI in audit and tax is crowded and moving fast, which means the due diligence burden on firms is real.
Start with integration. The most useful AI tools plug into software the firm already uses. A standalone AI research tool that requires copy-pasting text out of your tax research platform creates friction that most staff will eventually work around. Thomson Reuters and Bloomberg have understood this and built AI into their existing products. Microsoft Copilot is being integrated into the Microsoft 365 ecosystem, which matters if the firm runs on Outlook, Word, and Excel.
Data security is non-negotiable. Public accounting firms handle highly sensitive client data, and any AI vendor’s security posture needs to meet the same standards applied to cloud storage or tax software. Where is client data stored when it’s processed by the AI? Is it used to train the model? What are the vendor’s breach notification obligations? Several major firms have restricted or outright banned consumer AI tools like the free tier of ChatGPT for exactly these reasons.
Feature hype is a real problem in this space. Vendors claim capabilities that work in demos and fail in production. Piloting a tool with a limited engagement team on a specific task before rolling it out firm-wide is basic vendor evaluation discipline. AI tools that automate document extraction should be tested on actual client document types, not clean sample documents.
Cost-per-user pricing also needs scrutiny. A firm that pays for 50 seats of an AI research tool and has 10 active users is wasting money. Adoption is a real implementation variable, and tools with the steepest learning curves tend to have the lowest utilization rates six months in.
Managing Quality Control in AI-Assisted Accounting Work
Quality control gets more complicated with AI, not less. That runs against the way these tools are usually marketed.
The risk in AI-assisted work isn’t that the AI makes a mistake. It’s that the human reviewer trusts the AI output more than they should and doesn’t catch the mistake. This is called automation bias, and it’s well-documented in research contexts. Auditors and tax professionals aren’t immune to it.
Firms need to be explicit about what level of human review is required for each AI-assisted task. Extracting data from a W-2? Spot-check the output. Generating a tax research memo? Read the cited authority yourself, not just the AI summary. Scoring transactions for anomalies? The human review of flagged items needs to be substantive, not just a sign-off.
Some firms are building AI-specific review steps into their quality control checklists, and that’s the right call. Treating AI output like any other work product that requires review, rather than treating it as final output that just needs formatting, is what prevents the failure modes.
Liability follows the professional, not the tool. No engagement letter shifts responsibility to an AI vendor if a return is wrong or an audit procedure is inadequate.
Building AI Into Your Day-to-Day Accounting Practice
Individual practitioners don’t need a firm-wide rollout to start benefiting from AI tools. Some of the most effective AI adoption in public accounting is happening at the individual level, one practitioner changing how they handle a specific task.
Start narrow. Pick one task that’s time-consuming and repetitive: drafting client response emails, summarizing prior-year returns before a meeting, generating first-draft engagement letters. Use an AI tool on that one task for 30 days. Assess whether it actually saves time or just shifts time from doing the task to editing the output.
The practitioners getting the most out of AI are the ones who’ve gotten specific about prompting. Vague prompts produce vague output. A tax manager who prompts “summarize this balance sheet” gets less useful output than one who prompts “summarize this balance sheet for a manufacturing company client meeting, highlight year-over-year changes over 15%, and flag anything that might indicate cash flow risk.” It takes practice.
Continuing education in this area is genuinely useful. The AICPA has published resources on AI adoption in public accounting, and state CPA societies are increasingly offering CPE courses covering specific tools and workflows.
The standards around AI use in audit and assurance engagements are actively being developed. Staying current on AICPA guidance and PCAOB positions on AI-assisted procedures isn’t optional for practitioners who want to use these tools in client work without creating documentation and methodology gaps. That guidance will shape which AI-assisted workflows are defensible, and the details matter considerably more than the general direction of travel.