Dev.to AI 🤖 Ai 👁 0 📖 5 min read

What an AI accounting agent should do in a CPA firm, and what stays with your people

Every tax season the same hours disappear. Partners and seniors spend them chasing clients for missing documents, rebuilding status lists out of email threads, and writing the same follow-up note for the fourth time. Non

Every tax season the same hours disappear. Partners and seniors spend them chasing clients for missing documents, rebuilding status lists out of email threads, and writing the same follow-up note for the fourth time. None of that is review work, and none of it is client work. It comes straight out of the hours you'd rather spend on the return in front of you, and in March you can't buy them back.

That's the gap an AI accounting agent is meant to fill. Not the judgment and not the signature. The chasing, the assembling and the sorting.

What the agent actually does

An AI accounting agent is software that handles one defined job, such as chasing missing documents or preparing review files, inside systems the firm has authorized. That definition matters more than any feature list. It's one job, in systems you chose, with a clear split between preparing the work and checking it.

Take document chase. The agent builds a received-versus-requested list for each client and drafts the follow-up email. A staff accountant reads it and releases it. The agent never decides that a client has been chased enough, or that a vague reply counts as a document.

1040 intake works the same way. The agent compares what came in against last year's return and flags what's missing or changed. The preparer works from that list, and the signing CPA signs. For review prep, the agent puts together the workpaper index, the tie-outs and the open items, and the engagement partner signs off. When an IRS notice arrives, the agent matches it to the client, the year and the return, and the preparer takes it from there.

In every one of those, the agent does the assembly and a named person owns what goes out.

Where the rules still put a person

Software doesn't move the duties that sit on the firm. Circular 230 section 10.22 requires due diligence in preparing and approving returns and other papers filed with the IRS. A tool can make that diligence quicker to carry out. It doesn't carry it out for you, and it doesn't shift the obligation to whoever sold you the tool.

Then there's disclosure. Section 7216 of the Internal Revenue Code makes it a crime for a preparer to knowingly or recklessly disclose or use tax return information, except where the law allows it. Under 26 CFR 301.7216-3, you need the taxpayer's written consent unless another rule already permits the use. In practice, you settle the consent question before return data ever reaches an agent. Not halfway through the pilot, and not by assuming the vendor's terms cover it.

Security is the third piece. IRS Publication 4557 says that under the FTC Safeguards Rule, tax preparers have to create and carry out a written security plan. The rule itself, at 16 CFR 314.4, calls for encrypting customer information in transit over external networks and at rest (or approved compensating controls where that isn't feasible), and for logging what authorized users do. So when you look at any agent, ask where the data sits, whether it's encrypted both ways, and whether you can see who did what and when.

Oversight should be written down too. The NIST AI Risk Management Framework (subcategory MAP 3.5) says processes for human oversight should be defined, assessed and documented. For a firm, defined means you can name the reviewer for each job. Assessed means you check whether the reviews are catching problems. Documented means there's a record you could hand to a peer reviewer.

That last one is where most firms get loose. Everyone says a person reviews the output. Fewer firms can say who reviewed which draft, and what exactly they approved. If the agent records each draft, who approved it and the version they approved, you get that record without anyone keeping a spreadsheet.

A 30-day pilot

Pick one job. Document chase is the usual starting point because the output is easy to judge: either the list was right and the email was usable, or it wasn't.

Week one is mostly decisions. Choose the job. Name the reviewer and the signer as people, not roles. Write down the systems the agent may read, and nothing past that. Settle 7216 consent. Then set aside a handful of finished files the firm considers right, so you have something concrete to compare against.

Week two is quiet. Run the agent on old files while staff work as usual, then compare. Did it find the same missing items your people found? Did it miss anything, or flag things that weren't really missing? You're testing your inputs as much as the tool. A sloppy request list produces a sloppy chase no matter what software runs it.

Weeks three and four use live work, and every draft gets reviewed before it goes out. For each one, note whether it was accepted as written, edited before release, or redone from scratch. Time the review as well. If checking a draft takes as long as writing it, the agent isn't helping yet.

On day 30 you decide. If most drafts go out with light edits, add a second job. If the agent keeps missing the same kind of document, fix the request list before you add anything. If your people keep redoing the work, change the setup and run another month, or stop. Stopping counts as an answer. Finding out in 30 days that a job doesn't fit beats finding out in April.

Why we hold the sends

At Precision AI OS we build an Agent OS for firms like yours. It gives the firm one shared memory and an agent for each area of work, and by default it holds every send and every change in your apps until someone on your team approves it. Each approval is recorded and tied to exactly what was approved. No rule can approve payments, amounts or bank details, email to someone who isn't already on the thread, or deleting files and records. Those are the places where a mistake is expensive and hard to undo, so a person on your team decides.

That's the standard I'd hold any tool to this season, ours included. The agent should make the preparation faster and the record better. The consent, the judgment and the signature stay with your people.

More on how we set this up: https://getprecision.ai/?utm_source=devto&utm_medium=social&utm_campaign=ai-accounting-agent

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.