If you sold Bitcoin last year, drove for a rideshare app, or claimed a charitable deduction that dwarfed your W-2 income, the IRS has a faster way to check your math than it did even two years ago. A June 2025 Government Accountability Office review counted 126 active artificial-intelligence use cases already running inside the agency, spanning fraud detection, document processing, and compliance screening. (The figure has circulated publicly as “125 AI models”; the GAO tallied use cases, each of which can involve more than one algorithm.) No official source has published a model-by-model breakdown of what each one targets, but the direction is unmistakable: the IRS is cross-referencing more data, faster, across more categories of income and deductions than at any point in its history.
That matters right now because the 2026 filing season is the first in which several new reporting streams, including crypto broker filings and a lower threshold for gig-platform payment reports, are fully active at the same time the agency’s AI infrastructure has reached operational scale.
How the IRS built its AI infrastructure
Congress authorized roughly $80 billion in new IRS funding through the Inflation Reduction Act of 2022, with a significant share earmarked for technology modernization. Subsequent rescissions, including about $1.4 billion clawed back under the Fiscal Responsibility Act of 2023 and additional cuts in later spending bills, reduced the total by an estimated $20 billion or more. Even so, enough money reached the agency’s IT division to support a rapid buildout of data-matching and machine-learning tools.
The agency formalized its approach in IRM 10.24.1, an internal policy manual that requires every AI application to be logged in a use-case inventory, evaluated for its impact on the public, and flagged for additional review if it qualifies as “high-impact” under Office of Management and Budget guidance. The Treasury AI Use Case Inventory, most recently updated in early 2026, provides a public-facing catalog of these applications across the department, including IRS entries.
The GAO confirmed the 126-use-case count but also flagged gaps: the agency still needs to close workforce skill shortages and improve the quality of the data feeding its models before they can perform reliably at full scale. In practical terms, the plumbing is built; the agency is still tuning the pressure.
Crypto: new broker reporting meets the Form 1040 question
Every individual filer now encounters the digital assets question near the top of Form 1040, asking whether they received, sold, or otherwise disposed of digital assets during the tax year. That yes-or-no answer creates a self-attestation record the IRS can compare against a growing body of third-party data.
Under final regulations published in 2024 (TD 9989), custodial crypto platforms must report gross proceeds on the new Form 1099-DA beginning with tax-year 2025 transactions. Cost-basis reporting phases in for tax-year 2026 sales. Once those filings reach the IRS, automated systems can match reported proceeds against what appears on Schedule D and Form 8949. A gap between what Coinbase or Kraken reported and what a taxpayer disclosed is exactly the kind of structured discrepancy a matching engine is built to surface.
The practical risk is straightforward: if you sold crypto and did not report it, or reported a cost basis the platform cannot confirm, the odds of receiving a notice have increased. Taxpayers who transferred assets between wallets or used decentralized exchanges face murkier territory because those transactions may not yet generate third-party information returns. But the Form 1040 question still applies, and answering it incorrectly is a separate compliance problem.
Gig income: platform data meets Schedule C
The IRS Gig Economy Tax Center states plainly that all income from platform-based work is taxable regardless of whether the worker receives a Form 1099. The reporting threshold for Form 1099-K has been dropping in stages: $5,000 for tax-year 2024 transactions, $2,500 for 2025, and $600 for 2026. That final step pulls millions of part-time sellers and freelancers into the third-party data pool for the first time.
Picture a rideshare driver in Phoenix who logged 30 hours a week on two platforms last year and pulled in $14,000 between them. She reported $9,000 on Schedule C, forgetting that the second app also sent a 1099-K to the IRS. Under the old system, that $5,000 gap might have sat in a queue for months. Now, an AI-enabled matching engine can flag the discrepancy within days of processing, generating a CP2000 proposed-adjustment notice that lands in her mailbox proposing she owes roughly $1,200 in additional tax plus interest. She has 30 days to respond, and if she simply forgot the second platform, the fastest resolution is to agree and pay. Document matching itself is not new; the IRS has run these programs for decades. What has changed is the speed, volume, and pattern-recognition capability that machine-learning tools bring to the task.
Large deductions: pattern recognition at scale
No public document lists the specific algorithms assigned to scrutinize deductions, but the GAO’s description of “compliance” use cases strongly implies that models are scanning for patterns: charitable contributions disproportionate to income, repeated Schedule C losses, or credits that are frequently claimed in error, such as the Earned Income Tax Credit.
These models draw on historical audit data, adjustment outcomes, and current-year filing patterns to estimate which returns present the highest potential revenue impact if examined. For a taxpayer who donated $40,000 on a $70,000 salary and has the receipts to prove it, the flag may amount to nothing more than a letter requesting documentation. For someone who inflated deductions without support, the same flag could be the start of a longer process.
What the IRS still will not disclose
The biggest gap in the public record is performance data. Neither the GAO report nor the Treasury inventory discloses false-positive rates, audit conversion rates, or taxpayer appeal outcomes tied to AI-generated flags. Without those numbers, no one outside the agency can say how often automated matching produces accurate results versus how often it generates notices that burden compliant filers with weeks of paperwork.
The GAO flagged information quality as a concern but stopped short of publishing error-rate statistics or setting benchmarks for acceptable performance. The governance documents describe what the agency is required to do, including documenting training data and conducting impact assessments, but they do not show how consistently those requirements are met in practice.
From a taxpayer’s perspective, the experience has not visibly changed. A person who receives a notice still interacts with the agency through traditional channels: a letter in the mail, a request for documentation, or an adjustment to a refund. The IRS’s online account tools, including the refund status portal, can show processing updates and some correspondence, but they do not reveal whether an issue was initially flagged by AI or by an older rule-based filter. The technology behind the curtain is new; the letter on your kitchen counter looks the same as it always has.
What a CP2000 notice actually looks like
Because AI-driven matching feeds into the same notice pipeline the IRS has used for years, the downstream process is well documented. A CP2000 is not an audit. It is a proposed adjustment: the IRS believes there is a discrepancy between what you reported and what third parties reported, and it gives you 30 days to respond. You can agree and pay the difference, partially agree, or dispute the notice entirely with supporting documentation.
If you disagree and the IRS does not accept your explanation, you can request a meeting with a supervisor or file a formal protest with the IRS Independent Office of Appeals. The legal weight of a CP2000 is the same whether the mismatch was surfaced by a decades-old filter or a 2025-vintage neural network. AI does not change your rights; it changes how quickly the agency spots the discrepancy in the first place.
How to file with fewer surprises in 2026
The shift toward AI-driven matching means the IRS is less reliant on random selection and more focused on data-driven discrepancies. For the more than 150 million individuals who file each year, a few steps can reduce the chance of an unwelcome notice:
- Crypto sellers: Download transaction histories from every exchange and wallet you used. Reconcile cost basis before filing, and answer the Form 1040 digital-asset question accurately. If you used decentralized platforms, keep your own records; the IRS may not have third-party data yet, but the reporting obligation is yours.
- Gig workers: Track all platform income, even amounts below the 1099-K threshold. The IRS considers it taxable whether or not you receive a form, and matching systems may flag the gap.
- Large-deduction filers: Keep contemporaneous documentation for charitable gifts, business expenses, and any credit you claim. If your deductions look unusual relative to your income, expect the possibility of a verification request and have your records ready.
None of this guarantees you will avoid scrutiny. But filing accurately and keeping clean records is the most direct defense against a system that is getting faster at finding mismatches, even if no one outside the agency can yet measure how accurately it does so.
