How Do Accounting Firms Use AI? The Six Jobs It Actually Does, With 2026 Survey Data

Accounting firms use AI most heavily for tax research, where 60 percent of US tax professionals now reach for it at least weekly, up from 33 percent a year earlier. After research, the leading uses are advisory projects (44 percent), tax planning (40 percent), compliance research (39 percent), document analysis (36 percent) and drafting (35 percent). Those figures come from the second annual Blue J and CPA.com study of more than 1,000 US practitioners, released June 8, 2026.

That is the measured answer. The more useful one is that individual accountants have moved much faster than the firms they work in, and the gap between those two facts explains most of what is happening in the profession right now.

The adoption gap nobody plans for

Put the two big 2026 surveys side by side and the picture stops being a simple upward line.

The AICPA's PCPS CPA Firm Top Issues Survey was fielded April 20 to May 22, 2026 with 629 respondents across practice types and firm sizes. Technology adoption ranked No. 1 as a current issue only at firms with 500 or more professionals. At firms of 101 to 500, the top three current issues were all technology-related. At every size below that, from solo practitioners through firms of 100 professionals, technology and AI did not appear in the top five current issues at all.

Yet in the same survey, managing change due to technology and AI ranked No. 1 across the board for projected impact over the next five years. Every firm size agrees it matters enormously and most sizes are not treating it as a present priority.

Meanwhile the Blue J and CPA.com data says individual practitioners are already using these tools weekly. So the realistic state of a typical fifteen-person US firm in 2026 is not "we have not started". It is "several people started months ago, nobody wrote anything down, and no partner knows which client data has been pasted into which chat window". That is a governance problem wearing a technology costume, and it is the one worth solving first.

The six jobs AI is actually doing in firms

1. Tax research

The clear leader, and the reason is structural. Tax research is a well-defined question against a large corpus, which is the exact shape of problem language models handle best. Purpose-built research assistants cite primary authority and show their working, which matters because an uncited answer to a tax question is worthless. General chat tools are a poor substitute here: they will produce a confident answer with a fabricated citation, and checking the citation costs more than the research saved.

2. Advisory and tax planning

Forty-four and 40 percent respectively, and the highest-value use on the list. What AI does well here is generate the scenarios: run the numbers three ways, surface the options a busy preparer would not stop to list, and draft the explanation in language a client will read. What it does not do is decide which option fits the client sitting in front of you. That judgment is the product being sold.

3. Document analysis and cleanup

Thirty-six percent, and the least glamorous entry on the list. Firms use AI to read what clients send and turn it into something workable: pulling figures out of statements, matching supporting documents to the request that asked for them, flagging the month where three transactions do not reconcile. The bottleneck it addresses is real. Client documents arrive as photographs of paper, PDFs exported from a bank portal, and spreadsheets built by someone with a personal system.

Much of this is unglamorous format conversion rather than intelligence. A bookkeeper who can turn a client's PDF bank statement into a clean spreadsheet in a few seconds has removed the step that used to consume the first hour of every cleanup engagement, and no amount of downstream AI helps until the data is in a shape something can read.

4. Drafting client communication

Thirty-five percent, and the use that spreads fastest because it needs no procurement. Explaining a notice, summarizing what a return means, writing the fourth polite request for a missing 1099. This work is high volume, low judgment and genuinely tedious, which is a good profile for automation. It is also where firms most often let a draft go out without a human reading it, and where that habit does the most damage the first time a figure is wrong.

5. Meeting notes and engagement records

Less measured in the surveys and very common in practice. A client call generates commitments on both sides, and the record of what was agreed is usually a partner's memory plus three lines in a file. Automatic notes turn that into something a manager can pick up in November when the client asks why a position was taken.

6. The administrative tail

The part that never shows up as a line item and reliably costs the most. Scheduling that takes four emails. Chasing the same missing document five times. Filing what arrives into the right client folder with a name someone else can find. Practice management software tells you these items are outstanding. Somebody on staff still does them, and in a small firm that somebody is often a chargeable person.

What firms are doing with the time

Eighty-four percent of respondents in the Blue J and CPA.com study agreed AI saves time, and the study asked the better follow-up question: where does that time go. Fifty percent said improving client response and delivery timelines, 47 percent said staff work-life balance, and 46 percent said delivering higher-quality advice.

The commercial consequence is in the same report. Sixty-nine percent expect to move toward value-based, hybrid or fixed-fee billing. That follows logically: once the hours attached to research and drafting fall away, billing by the hour means billing less for the same delivered value. Firms that hold onto hourly pricing while automating the hours are quietly cutting their own revenue.

Where firms are getting it wrong

Four patterns come up repeatedly.

Buying AI before defining the process. If jobs live in a spreadsheet and deadlines live in one partner's head, AI produces faster chaos rather than less of it. The system of record comes first. Our comparison of AI tools for accounting firms prices the practice management platforms that hold that system, from $40 to $74 per user per month at entry tier, and says plainly which firms should buy one before considering anything else.

Pasting client data into consumer tools. The consumer tier and the business tier of the same vendor often have different terms about training on your inputs. Accounting firms hold Social Security numbers, bank details and unreleased financials. This one is worth checking this week rather than next quarter.

No named reviewer. Any output containing a figure that reaches a client should have passed a named human. The failure mode is not a dramatic hallucination, it is a plausible number in a paragraph nobody read closely because the paragraph was well written.

Evaluating in February. Firms shop for software when the pain peaks, which is the worst month to change how work moves. Run trials in the trough, take one service line, and migrate it completely rather than half-migrating everything.

A sensible order for a firm under twenty people

1. Write the policy first. Which tools are approved by name, what client data may go into each, who reviews output containing figures, what gets logged, and who to ask about anything not covered. One page beats a document nobody opens. Keep it wherever your other procedures live, which is what an SOP template is for.
2. Get a system of record. Clients, jobs, deadlines, documents, in one place with recurring work templated.
3. Automate the highest-volume, lowest-judgment task you have. Usually document chasing or first-draft client correspondence.
4. Then add AI against a specific technical bottleneck, most often tax research.
5. Measure two numbers: elapsed days from a client's last document to delivery, and admin hours per job. Everything demos well; only your own jobs show you what moved.

Will AI replace accountants?

Nothing in the survey data supports it. What the data shows is a change in what firms sell rather than how many people sell it: research and drafting stop consuming the hours, hourly billing stops making sense, and advisory work is what remains. The judgment, the signature and the professional liability stay with a licensed human, and the firms treating AI as a way to sell more advisory rather than fewer staff are the ones the numbers favor.

The realistic risk to a small firm this year is not being replaced. It is having six people quietly using six different tools on client data under no policy at all.