The IRS Is Using AI to Screen Returns — Here's What That Means If You're Also Using AI to Prepare Them

A flagged return doesn't automatically mean an audit. It means the IRS AI has scored the return above a threshold that warrants closer review — either by another automated layer or, eventually, a human examiner. Most flagged returns never become full audits. But a flagged return does mean delays, potential follow-up requests, and a higher probability of landing in front of a human who is looking for something wrong.

The IRS systems doing this work include:

• Discriminant Function: The oldest and most widely deployed. A machine learning model that scores every return for audit potential by comparing income, deductions, and credits against statistical norms for similar filers. Returns that deviate significantly from the expected distribution get higher scores.
• Individual Taxpayer Model: Flags the top three specific line items most likely to need adjustment on each return. This is surgical — it doesn't just flag the return as suspicious, it identifies which lines look off.
• Large Partnership Compliance Model: Targets complex pass-through entities. In 2021 it selected 82 high-risk partnership returns for review; human reviewers had previously identified only single digits.

The IRS also deployed voice and chat automation that handled over 4.8 million calls and 450,000 inquiries this filing season — largely a product of the 25% workforce reduction that shifted service delivery toward automated systems.

---

The Five Flags Most Likely to Trigger a Review

These are the patterns the IRS algorithms are built to find:

1. W-2 and 1099 mismatches

The IRS receives copies of every W-2 and 1099 directly from employers and payers. If your client's return reports a different amount than what the payer reported, the system flags it immediately. This is the most straightforward catch — and the one AI preparation tools are most likely to create if they OCR a document incorrectly or use prior-year figures as a starting point.

Before you submit: Manually compare every income line on the return against the physical or digital document the client provided. Don't trust the AI's read if the document quality was poor.

2. Deduction ratios that exceed statistical norms

The Discriminant Function scores returns against comparable filers. A client with $180,000 in business income who claims $90,000 in home office, vehicle, and travel deductions looks statistically unusual — regardless of whether every deduction is legitimate. "Extreme deduction ratios" is a specific flag the system is built to catch.

Before you submit: Flag any return where total deductions exceed 40% of gross income. Review each item. If the deductions are legitimate, make sure the documentation exists. Don't just log it in the file — confirm it's retrievable.

3. Round numbers throughout

Round numbers suggest estimates. The IRS AI is trained to notice when a return has suspiciously clean figures — $12,000 in charitable contributions, $5,000 in vehicle expenses, $8,000 in office supplies. Round numbers across multiple categories read as estimated rather than actual, which increases scrutiny.

Before you submit: Scan for round numbers, especially in deduction-heavy schedules. Replace estimates with actuals where you can. Where you can't, flag the methodology.

4. Year-over-year income swings

A client who reported $95,000 last year and $190,000 this year isn't automatically suspicious — businesses grow. But a sharp income swing without a corresponding change in the business profile is a pattern the system is designed to catch. It also works in reverse: a significant income drop in a profitable-looking business is a flag for underreporting.

Before you submit: Pull last year's return for any client with a significant income change. If there's an explanation (new contract, business sale, one-time event), note it in the file. You'll need it if the return comes back with questions.

5. Self-employment income below what the business activity suggests

The IRS uses third-party data — bank records in some cases, industry benchmarks in others — to cross-reference self-employment income. A sole proprietor in a cash-heavy business who reports thin income against visible overhead raises flags. This is the highest-scrutiny category for the IRS AI, particularly for small businesses.

Before you submit: Cross-reference reported self-employment income against bank deposits for your highest-risk clients. If there's a gap, get the explanation documented now.

---

The Critical Reminder: The AI Didn't Sign the Return

This is worth saying plainly before you finish today's work.

Your AI preparation tool — whatever it is — does not sign tax returns. You do. When the IRS AI flags a return and a human examiner opens the file, they are looking at work product that carries your signature and your professional license. The software vendor has no liability. The AI has no liability.

"The AI did it" is not a defense. It never has been. It's just new.

What this means practically: using AI to prepare returns is not a problem. Using AI and skipping the human QC step is. The tools are fast and generally accurate — but accuracy isn't guaranteed, and the CPA's review is the only check between a clean AI output and a signed return.

---

What to Do Before 6:00 PM Today

If you have returns going out today or tomorrow, run this checklist before you submit:

• Match every W-2 and 1099 against the original document, not just the prior-year import
• Flag any deduction over 40% of gross income and confirm the documentation exists
• Scan for round numbers in deduction schedules and replace estimates with actuals where possible
• Pull last year's return for any client with an income swing greater than 30%
• Cross-reference self-employment income against bank activity for your highest-risk clients
• Log your human review step in the client file — date, preparer, what was checked

That's it. Not a massive process change — a one-hour triage against the patterns the IRS AI is specifically built to catch.

---

The Larger Point

The IRS doubling its AI use cases in a single year is a structural shift, not an anomaly. For accounting firms adopting AI tools on the preparation side, this creates a new professional obligation: know how the screener thinks, not just how the tool prepares.

You are now operating in an environment where both sides of the filing have algorithmic attention. The firms that adapt their QC protocols to match that reality will have fewer surprises. The ones that treat AI preparation as set-it-and-forget-it will not.

---

The Crossing Report covers AI adoption for professional services firm owners. Subscribe here for weekly insights on what's changing, what matters, and what to do next.