On 24 March 2026 the Government Accountability Office published GAO-26-107522, an assessment of how the IRS uses artificial intelligence. One number in it does most of the work.
What the Inventory Actually Says
The IRS held 126 active AI use cases in its inventory as of June 2025. In August 2022 it reported 10.
Two details in the same report matter more than the headline count, and both cut against reading it as a finished capability.
Sixty-one percent of the use cases, 77 of the 126, were listed as still in development in June 2025. And 65 of the 126 were either too sensitive for public reporting or were research and development efforts exempt from it.
So the public inventory is roughly half the picture, and more than half of what is in it had not shipped. Anyone telling you the IRS has a finished AI audit machine is overstating the report. Anyone telling you this is early-stage experimentation is understating a twelvefold increase in three years.
The report itself is titled around gaps, not triumphs: skills gaps, information quality and strategic management. GAO noted that officials in the Research, Applied Analytics and Statistics group had lost 63 employees who had been working on AI full or part time.
Selection Got Cheaper. Examination Did Not.
Set that against staffing. The IRS workforce fell roughly 25% over the year to mid-2025.
Those two facts pull in opposite directions and the combination is the actual story. Building a model that ranks returns by risk is a fixed cost paid once. Examining a return is a variable cost paid in trained hours, and trained hours got scarcer.
The rational response to that constraint is exactly what the inventory shows: spend on selection, so that the examinations you can still afford land on the returns most likely to justify them.
For a firm, that reframes the practical question. It was never really about the odds. It is now about what happens in the weeks after a return is flagged, because a model-selected examination arrives already pointed at something specific.
The Bias Finding, Stated Plainly
There is a caveat here that deserves to be reported rather than skipped.
GAO has flagged the potential for unintentional algorithmic bias in return selection, on the straightforward mechanism that a model trained on historical selections inherits whatever was in them. Independent studies cited by GAO found Black taxpayers audited at three to five times the rate of others. GAO has called for better documentation and transparency around the IRS use of AI.
That is a serious finding about a public system, and it is a reason to want the documentation GAO asked for. It is not a reason to assume a particular client was flagged wrongly. Both things are true at once, and a firm advising a selected client has to hold them separately.
What Actually Protects a Position
If selection is model-driven and examinations are fewer and better aimed, the defensible position is the one that can be supported quickly.
The mechanics of that are unglamorous. When a notice arrives about a specific line, the firm needs the source documents behind it, the reason the position was taken, and the correspondence around it, in the time the notice allows rather than in the time it would take to reconstruct three years of filing decisions.
That is where most of the cost of an examination actually sits for a small firm. Not in the technical argument, which is usually straightforward, but in assembling the evidence for an argument nobody expected to have to make about a return that was filed correctly.
MetaWurks works on that half. It ingests a client's returns, statements, invoices, contracts and correspondence and lets an accountant query the whole set in plain English, so establishing what supports a particular figure is a question rather than an excavation across a portal, a drive and an inbox. Role based access controls decide who can open which client's records, audit logs record who opened what and when, and documents ingested into the platform are not used to train models or exposed to other users.
It has no view on how a return was selected, and no software does. What it changes is the speed of the answer once the letter is on the desk.
The IRS went from 10 use cases to 126 while losing a quarter of its people. Read the second half of that sentence as carefully as the first.
Join the Conversation
If a client received a notice next week questioning one figure on a return you filed in 2024, how long would it take to put the supporting documents in front of them?