The IRS Reviewing This Year's Returns Isn't the One From Three Years Ago

Something changed at the IRS, and most of your clients have no idea it happened.

The agency was running 126 active AI projects as of mid-2025, up from 10 in August 2022, according to Accounting Today's read of the GAO findings. Twenty-seven of those projects point straight at compliance and enforcement. The Discriminant Information Function, the scoring model that decided for decades which returns earned a second look, is giving way to machine learning. The Small Business/Self-Employed Division built its own return-selection model aimed squarely at self-employed filers.

All of this arrived while the agency shed roughly a quarter of its workforce between January and May of 2025. Revenue agents fell 31 percent. Read those two facts together and you have the whole story: more automated selection, fewer humans to reach when a notice lands wrong.

Your clients will assume the return they file this year gets the same treatment the one they filed three years ago got. It does not. That gap between what they expect and what is actually happening is where you earn your keep now.

What the models are looking for

The GAO analysis, as reported, names the patterns the new models reward with attention. Returns showing low taxable income against visible signs of wealth or asset accumulation. Income or balance sheet numbers that do not reconcile across multiple years. Round numbers sitting in deduction schedules, the kind that signal an estimate rather than a record. Partnership allocations that make no economic sense.

The agency also codified how it uses these tools. IRM 10.24.1, effective February 10, 2026, governs AI in audit selection and exam support. And a new reporting line, Form 1099-DA for digital assets, took effect for the 2025 tax year, which means the IRS now receives its own basis data on client crypto activity and can match it against what the return claims.

None of this is exotic. It is your ordinary small business and self-employed clients, the three categories the models target most: large partnerships, the self-employed, and small business owners. The people who fill your appointment book are the people the system is built to notice.

The asymmetry that will hurt

Here is the part that changes how you should run the practice. The flags arrive fast. The resolution does not.

An automated notice can hit a client's mailbox within weeks. But reaching a live agent, getting a substantive answer, and working a notice through to a close now takes far longer than it did two years ago, because the people who used to answer are gone. So the machine moves quickly on the front end and the humans move slowly on the back end, and your client sits in the gap paying interest and losing sleep.

You cannot fix the IRS's staffing. You can change where your value sits. It moves upstream, to the work you do before anything is filed.

Where we would start this quarter

We are not tax advisors, and this is not tax advice. What we think about is the systems your clients are being measured by. From that seat, four things stand out.

Document contemporaneously, not defensively. Build the record while the engagement is live, not after a notice arrives. If a deduction is real, the support should exist the day the return is filed, not be assembled under pressure eight months later. The models reward returns that look estimated. Precision is your defense.

Run a pre-filing pass against the known flags. The red-flag list above is public. Before a return goes out, look at it the way the model will. Round numbers, unreconciled multi-year figures, income that looks thin against a client's lifestyle. Not to change the truth, but to attach the explanation before the machine asks for it.

Make 1099-DA basis reconciliation standard for any client who touched digital assets. The IRS now has its own copy of that data. A mismatch is a clean, automatable flag. Reconcile it every time.

Reposition from seasonal preparer to year-round risk partner. This is the real opportunity buried in a story that otherwise sounds like bad news. When enforcement runs on continuous data matching, seasonal service stops being enough. Clients need someone watching the whole year. That is a better business for you, and a better deal for them.

The honest read

The tax gap the agency is chasing is enormous, around $688 billion for 2021 by the GAO's count. That pressure is not going away, and neither is the automation. But automation is only as good as the data it runs on, and the data it runs on is the return you help prepare.

That is where you have real influence, and it is yours to use. A well-documented, reconciled, economically coherent return is not just less likely to get flagged. It is faster to defend if it is. You are not competing with the model. You are deciding what it gets to see.

We would rather your clients learn this from you, in a planning conversation this fall, than learn it from a notice next spring.

If you are trying to figure out where AI actually changes the ground under your firm, and where it is just noise, that is the kind of question we work through with people.