How are accounting firms and CPA firms using AI?

The short answer: accounting firms are getting real value from AI on the practice-management side, not on the return. The jobs that work are prospect and referral research, chasing missing client documents, drafting routine client follow-up, triaging the busy season inbox, keeping job statuses current and producing the weekly workflow report. The jobs that do not work are preparing or signing returns, giving tax advice, and anything that involves feeding client tax return information into a general AI tool. That last one is not a best-practice suggestion. It is a criminal statute, and almost nobody selling AI to accountants mentions it.

Start with the rule, because it decides everything else

IRC 7216 makes it a misdemeanor for a return preparer to knowingly or recklessly disclose information furnished for return preparation, or to use that information for any purpose other than preparing the return. The penalty is a fine of up to $1,000, up to a year in prison, or both, rising to $100,000 where the identity theft provision of IRC 6713(b) applies. IRC 6713 adds a civil penalty of $250 for each disclosure or use, capped at $10,000 in a calendar year.

Now apply that to the most obvious AI idea a firm has. You export your client list from the tax software, hand it to a tool, and ask it to write a personalized message to every client who might need advisory work. That is a use of tax return information for a purpose other than preparing the return. Treasury regulations do permit it with taxpayer consent, but the consent has to be knowing, voluntary, written, signed and obtained before the use, in the specific form the regulations prescribe. Nobody's marketing page is handling that for you.

This is why the firms doing AI well look boring from the outside. They are not running clever things on client data. They are running unglamorous things on data that has nothing to do with a return.

How are small accounting firms using AI, specifically?

Document chasing is the near-universal first job, and for good reason. At any moment a small firm has a third of its open jobs blocked on something a client has not sent: a missing 1099, a brokerage statement, an unsigned engagement letter, a K-1 the client insists was mailed. Historically somebody works that list when a gap appears in the day, which in March means never. The work is unambiguous, the message is low-risk, and the result is visible inside a week.

Prospect and referral research is the other common starting point, and it has the advantage of touching no client data at all. Before a call with a new inquiry or a referring attorney, an agent pulls together what the business does, its size, its likely entity structure, recent local news and a specific reason to talk. Partners who used to do that in the car park now walk in prepared.

Then there is the inbox. February through April, the shared inbox becomes the actual bottleneck in most practices. Sorting what is a document delivery, what is a scheduling request, what is a question that needs a preparer and what can wait until May is genuinely useful work that requires no judgment about tax. Drafting the routine replies saves more hours than any other single thing on this list.

What accounting firms are not using AI for

Preparing returns. Anyone who prepares or assists in preparing a federal return for compensation needs a PTIN, a human signs the return, and the professional standards for the positions taken on it attach to that person. Software has done the computation for forty years. What it does not do is take responsibility, and the responsibility is what the client is paying for.

Answering tax questions. Under Circular 230, the Treasury regulations at 31 CFR Part 10, practice before the IRS is limited to attorneys, CPAs and enrolled agents in good standing. Practice is defined broadly and expressly includes giving oral or written tax advice. So when a client emails asking how to treat the sale of a rental, the fastest thing in the world is to let software answer it. Do not. That reply is tax advice, and it needs a credentialed person's name on it.

The behavior you actually want from an agent is to recognize a technical question, route it to the preparer who owns that client with the file history attached, and draft a holding reply that confirms receipt without saying anything substantive. That is useful and it is safe.

Where the line sits, job by job

We keep the full version of this map, with the workflow around each job, on the AI agent for accounting firms page.

Is AI for bookkeeping the same thing? No

These get searched as if they were one product and they are two. Bookkeeping automation works inside the ledger: categorizing transactions, matching them to source documents, reconciling accounts, closing a period. That is a data pipeline problem and it belongs in the accounting platform or a tool purpose-built for it. If your bottleneck is that someone is retyping figures from a stack of client invoices, the fix is a tool that will pull the line items straight out of the documents, not a work agent.

A work agent sits outside the ledger and runs the human workflow around it: getting documents in, keeping clients moving, researching the next ten prospects, keeping the practice system honest, producing the report. Most firms want both eventually. It is worth being clear which problem you are buying for, because the tool that is good at one is generally not the tool for the other.

Will AI replace accountants or bookkeepers?

Accountants, no, and the regulatory structure is a large part of why. Signing a return and giving tax advice are activities the rules attach to a licensed person. Beyond the rules, the value a good accountant adds sits in knowing the client's business well enough to notice the thing the client did not think to mention, and that is not a document-processing task.

Bookkeepers, partly, and it is already happening to the transaction-coding portion of the work. That part was always going to be automated. What is not being automated is the judgment around the ledger: catching a categorization that is wrong because you know how the business operates, spotting the deposit that should not be revenue, and being accountable for numbers someone will rely on. The bookkeepers under real pressure are the ones whose service was only data entry.

What percentage of accounting firms use AI?

Be skeptical of the confident numbers here, including the ones in vendor surveys. Adoption figures for AI in accounting swing wildly depending on who is asked and what counts as use, and a firm whose staff quietly paste things into a chatbot will answer that question differently depending on who is in the room. The more useful question for your own practice is narrower: which specific recurring job would you hand over first, and would you be comfortable if a client learned exactly how it was handled?

If the answer to the second half is no, that is the signal to change the job, not to hide it. Most of the discomfort traces back to client data going somewhere it should not, which loops straight back to 7216.

How to start without creating a problem

Pick one narrow, repetitive, non-return job and get the format right before adding a second. Document chasing or prospect research, nothing else, for two weeks. Write down your conventions the way you would brief a new hire: how many attempts before escalation, which clients get a phone call instead of an email, how job notes should be structured, and what never gets sent without a partner reading it.

Then write the boundary down too, in a page anyone in the firm can read: no return information leaves your consented systems, no client-facing message that answers a technical question goes out without a credentialed review, and nothing touches the ledger. That single page will do more for your risk posture than any vendor's security page. Have your counsel look at it before you formalize anything, because this article is general information rather than legal or tax advice.

From there the additions are obvious. The same agent that chases documents can keep your practice records current, run the prospect research and assemble the same workflow report every Monday. The conventions you set in week one carry across all of it, which is the actual reason this compounds.