---

The IRS's AI Infrastructure — What It Actually Does

The IRS has been building its AI infrastructure quietly for years. The March 2026 GAO report gave the clearest public picture yet. Here's what the AI is doing:

Audit case selection: AI systems analyze return data looking for statistical anomalies — patterns that, in historical data, have been associated with underreporting or fraud. The model doesn't know your client. It knows whether their deduction ratios, income year-over-year, and self-employment income reconciliation match the patterns it's learned to flag.

Known flag triggers (from public IRS guidance and GAO analysis):

• Year-over-year income changes without documentary explanation
• Deduction ratios that are extreme for a given income level and industry
• Round numbers across Schedule C, where precise actual figures would be expected
• Mismatches between third-party reporting (1099s, K-1s, W-2s) and what the return shows
• Underreported self-employment income relative to social media, property records, or industry benchmarks

Voice bots for taxpayer correspondence: Over 4.8 million calls are handled by AI systems. If your client calls the IRS about a notice, there's a meaningful chance they're talking to an AI before a human.

Fraud detection: AI cross-references return data against third-party sources, identity databases, and prior return history to flag potential identity theft and refund fraud.

---

The Bias Problem — And Why It Matters for Your Clients

The finding buried in GAO-26-107522 is the one CPA firm owners need to understand.

The GAO found that the IRS has expanded AI deployment rapidly but lacks the workforce and oversight infrastructure to identify and correct unintentional biases in its systems. Translation: some taxpayer categories may be getting flagged at disproportionate rates because of how the training data was built — not because of actual compliance risk.

The GAO also found that IRS has been cutting skilled AI personnel even as AI use expands. The agency that's building AI audit infrastructure is simultaneously reducing its capacity to supervise and improve those systems.

This creates a gap. And it's a gap that professional advisors — CPA firm owners who understand both the AI systems and the client's situation — are positioned to step into.

The client who gets an audit notice and has clear, organized documentation explaining every anomaly on their return is in a very different position than the client who relied on a consumer AI tool to file and has no workpapers.

---

What Changes in How You Prepare Returns

You've always prepared accurate returns. The question now is whether your returns are prepared to communicate with an AI that can't read context — only patterns.

1. Documentation explains what AI can only flag.

When a return has an anomaly — a large casualty loss, a major income change, a home office deduction in a high-income year — that anomaly gets flagged by the model regardless of whether it's legitimate. The difference between a client who gets audited and a client who gets the audit closed quickly is documentation.

Build documentation into your workflow, not just your audit defense file. If the deduction is real, the explanation should travel with the return.

2. Precision beats estimation.

AI audit models have been trained on data that associates round numbers with estimation, and estimation with underreporting. Where actual figures are available, use them. A Schedule C expense of $4,287 looks different to the model than $4,000 — not because of the dollar difference, but because precision signals documentation.

This doesn't mean manufacturing false precision. It means not defaulting to estimates when actual records exist.

3. Reconcile everything before you file.

The IRS AI systems automatically compare third-party reporting — 1099s, K-1s, 1098s, broker statements — against the return. A mismatch doesn't trigger an audit notice from a human reviewer who might call to ask. It triggers an automated flag.

Your reconciliation checklist matters more than ever. Every third-party document your client receives should reconcile to the filed return.

---

The Client Who Used ChatGPT to File

Every CPA firm has at least one client who prepared their own return this year with AI assistance. The risk isn't that the IRS knows they used ChatGPT. The risk is what consumer AI tools produce.

Consumer AI tax tools are generating returns that pattern-match to the errors the IRS audit selection model was trained on: inconsistent figures across related schedules, round-number deductions without documentation, income items that don't reconcile to known third-party forms.

Your value in this environment is not knowing more tax law than an AI. It's documentation quality, reconciliation accuracy, and the contextual judgment to explain anomalies before an algorithm flags them. That's the irreplaceable advisory layer — and the AI-selected audit environment makes it more defensible to clients, not less.

---

The Opportunity in the Interpretation Gap

Here's the honest counterintuitive angle on all of this: the IRS expanding AI without sufficient oversight infrastructure creates an interpretation gap that professional advisors can fill.

Your clients have IRS correspondence they don't understand, AI-generated notices they can't interpret, and audit selection they can't predict. You can. You know how the systems work, you understand what triggers flags, and you can build return preparation practices that minimize unnecessary audit risk.

The accounting firm that positions itself as the interpreter between clients and an AI-driven tax enforcement infrastructure has a concrete value proposition that didn't exist three years ago.

That's not a defensive posture. That's a service line.

---

Three Things to Do Before April 15

1. Review your reconciliation checklist. Ensure every return your team prepares includes full reconciliation of third-party documents before filing. Make it a non-negotiable step, not a catch-all at the end.

2. Build a "documentation explains the anomaly" habit. For any client return with a significant deduction change, income change, or unusual item — add a brief workpaper note. It costs five minutes. It changes the audit outcome.

3. Have the conversation with clients who self-prepared. The AI-prepared return that's coming through your door for an extension still benefits from a documentation review. Make that a billable service, not a favor.

The IRS is using AI to decide who gets audited. Your preparation work is now partly a communication with that system.

---

Related: IRS AI Audit-Proofing: How to Build the Advisory Service — The complete guide to pricing, positioning, and delivering AI audit-proofing as a CPA firm service line.

Sources: GAO-26-107522 | FedScoop — IRS ramps up AI use amid staff cuts | EisnerAmper — AI at the IRS | Capitol Technology University — Audited by an Algorithm