You have read this story. A tech giant lets people go, a filing mentions artificial intelligence — and the headline finishes the thought. In June, CNBC reported that Oracle "reduced its workforce by 13% over the past 12 months and cited AI deployment in an annual filing," in a story that also ran through Meta's layoffs and Microsoft's buyouts. By September, one site had boiled it down to "Oracle admitted AI eliminated 21,000 jobs."
You probably believed it, and you weren't wrong to take it seriously. Oracle's paperwork does tie AI to its job cuts. It just never says what that headline says (no filing I read puts a number on AI's share of the cuts).
The quarterly report Oracle filed on September 11, 2026 describes a restructuring plan to "further improve operational efficiencies, including through the adoption and integration of artificial intelligence technologies across certain functions and other operational activities." The termination email that followed, in the lines the San Francisco Chronicle quoted, cited "a broader organizational change," not AI. And further on, the same report shows quarterly capital spending going from $8.5 billion a year earlier to $28.5 billion, "primarily due to the expansion of our data centers."
Three stories get told about AI and lost jobs. The machine does the work. The machine is the excuse. And the third, which none of the layoff forms below asks about: the machine eats the budget.
The restructuring note, then the cash pages
Oracle calls it the 2026 Restructuring Plan (the "2026" is its fiscal year, which ended May 31, 2026). "Layoff plan" is my shorthand, and a fair one: the costs recorded so far "primarily related to employee severance costs." The plan was expected to cost up to $2.1 billion; after the quarter closed, management added approximately $700 million "to reflect additional actions that we expect to take." Up to about $2.8 billion, then — about $2 billion of it already booked; the supplement, as Sanchit Vir Gogia of Greyhound Research told CIO, is "an estimate, not a bill." How many people does it cover? Oracle hasn't said, and Gogia was blunt: "There is no solid headcount for the September round."
The AI language isn't new in September. It arrived in June, when the annual report added a sentence to a risk factor on workforce restructurings: "In addition, the adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce." Last year's version of the same risk factor has no such sentence, and the new one names no roles and gives no number. Before June, the plan's own description had no AI in it either: a year earlier it was "due to our acquisitions and certain other operational activities", and by February it had grown to up to $2.1 billion without a word about AI. The plan reached most of its size first; the AI clause came afterward.
What the annual reports do give is a net change: approximately 162,000 full-time employees on May 31, 2025, approximately 141,000 a year later. That is about 21,000 fewer people (roughly 13%), with every hire, resignation and retirement folded in alongside the layoffs.
Now the cash. The jump to $28.5 billion follows a fiscal year in which capital spending went from $21.2 billion to $55.7 billion, and Oracle expects this year's to be "higher than fiscal 2026." Free cash flow for the quarter, by Oracle's own non-GAAP measure, was negative $5.4 billion. In the same three months, Oracle sold approximately 141 million new shares for net proceeds of $19.9 billion.
The expense lines show the tilt in miniature. Against the same quarter a year earlier, cloud and software expenses rose $2.8 billion, "primarily due to higher infrastructure expenses" (a different $2.8 billion from the plan's; pure coincidence), while sales and marketing fell $252 million, "primarily due to a decrease in employee-related expenses." The filing puts them side by side. It never says one paid for the other — and neither will I.
What it does say, in words Oracle has used for years (last year's annual report said the same of its 2024 plan), is that certain of the plan's savings were "offset by investments in resources and geographies" it believes better serve its cloud offerings.
Three stories, one set of layoffs
Story one, the machine does the work, is Oracle's own. In March it told investors that AI code generation had let it restructure product teams "into smaller, more agile and productive groups" and "build more software in less time with fewer people." Research and development staff did fall, from about 50,000 to about 43,000 in a year.
Story two, the machine is the excuse, is the one I told in June, when I wrote that "We're using AI to serve you better" should invite a follow-up: "who got cut to pay for it?" MIT Sloan's Paul Osterman told Fortune: "AI is a perfect excuse to justify big layoffs."
Story three is the one Bloomberg has reported since March, when Reuters relayed its report that Oracle was planning thousands of job cuts "as it faces a cash crunch from a massive AI data center expansion effort", some aimed at job categories "the company expects will shrink due to AI." On September 11, Bloomberg reported the cuts had grown "while the company grapples with a cash crunch" tied to building AI data centers for customers like OpenAI. Even CNBC's June story put Oracle among "several tech firms that are slashing jobs to offset massive capex spending on AI buildouts." Two stories at once — from the start.
Isn't Oracle just building what its customers ordered?
Fair question, and Oracle's answer deserves room.
First, customers are paying for much of it. Of the quarter's $23.1 billion in operating cash, $11.4 billion was customer prepayments "with a significant financing component," a kind Oracle received none of a year earlier. In March, Oracle said that for the large AI contracts behind most of that quarter's backlog growth it "does not expect to have to raise any incremental funds," because "most of the equipment needed" is either prepaid by customers or bought and supplied by them.
Second, scale. The whole plan, up to about $2.8 billion over its life, is roughly a tenth of one quarter's capital spending; in fiscal 2026 its $1.8 billion of expense came to about 3% of that year's $55.7 billion. Payroll cannot pay for this buildout, and anyone selling you a dollar-for-dollar swap of workers for data centers is selling bad arithmetic.
Third, restructuring is a habit at Oracle: its filings describe a 2022 plan and a 2024 plan before this one. And Oracle's own account, in a statement to CNBC in June, is about fit, not cash: "As our cloud and AI businesses grow, we will continually balance our resources and restructure our development group to help ensure we have the right people …"
So what survives of story three? Priorities, and cash. Analysts spelled it out in January: TD Cowen, in a note CIO reported, saw customer-supplied chips and workforce reductions as Oracle's likeliest path, since the chips would address the capital bill "while job cuts would improve cash flow." In July, S&P downgraded Oracle to BBB-, the lowest investment-grade rating, and, per Investing.com's summary, expects its fiscal 2027 free operating cash flow deficit to widen to $42 billion, with OpenAI roughly half of Oracle's contracted backlog. (HAIA noted in June that Oracle, Microsoft and Amazon had all signed reported infrastructure commitments "betting that OpenAI will be good for the money.") Even The Motley Fool, which owns and recommends Oracle shares, ran Daniel Sparks's verdict: "My worry is the funding."
To be clear, none of this makes the AI clause false. The honest version of story three is narrower: when cash runs this tight, payroll is the lever a company can pull fastest, and the restructuring note names AI, not the budget.
What Connecticut's new question can and can't hear
So what can a state form hear of all this? Section 26 of Connecticut's Public Act 26-15, signed in May and in force from October 1, 2026, requires any employer serving a federal WARN notice (the 60-day warning larger employers owe before a mass layoff) on the state Labor Department to disclose whether the layoffs are "related to the employer's use of artificial intelligence or another technological change." The section carries no penalty of its own; none of the act's enforcement clauses reaches it. And as of September 24, 2026, the Labor Department's WARN page still listed four things a notice "should include," none of them AI, and linked no form for the new question.
I wrote about that sentence in August, arguing that a filing makes an answer exist: "dated, attributable, public, sitting beside the number of people who lost their work." The trouble is which answer. Oracle's own annual report would support a yes on a Connecticut form, and a yes would say nothing about $28.5 billion.
Oracle already answers one state form. New York's WARN dashboard lists about 110 notices (197 site listings) dated in 2026 through August 25; exactly one names AI (Nespresso's, for 46 workers), and Oracle's only one, dated April 2, covers 62 workers in Valhalla, at Westchester Medical Center's campus, with the reason "Contract Termination." That may be exactly right for that site. Neither that form nor June's annual report is wrong on its face: one sees the immediate reason at one site, the other a strategy, and neither ties the cuts to the capital budget, though the annual report prints it.
New York has asked since March 2025, by checkbox, whether "technological innovation or automation" contributed. In May, lawyers at Hunton Andrews Kurth reported that of more than 160 companies filing in the first year, "not a single notice has attributed layoffs to AI technology or automation." (Nespresso's January notice, posted in April, names AI. Even the tallies disagree.) The federal rule underneath asks where, when, how many, which job titles and whether it's permanent. Read the list and you won't find a line asking why.
France asks the jobs question before the software arrives
France runs the same question in the opposite order. In companies with at least 50 employees, the works council (the comité social et économique, or CSE) must be informed and consulted on measures affecting the size or structure of the workforce and on the introduction of new technologies, and the employer's decisions must be preceded by that consultation. Obstructing the council's regular functioning is an offense punishable by a €7,500 fine.
Since 2025, judges have applied that order to AI. On February 14, 2025, a Nanterre judge froze an insurer's AI rollout until the works council had its say: the insurer had opened a consultation but put the tools to use in a test phase first, and the judge called that early deployment "un trouble manifestement illicite" (a manifestly unlawful disturbance). The insurer was MetLife France, and its council's lawyer, quoted by Républik IT, said that, lacking information, the council was asking about "des risques de suppression de postes" (the risk of job cuts).
Five months later a Créteil judge went further. A press group had deployed AI tools without involving its council, and the court held that it could not seriously be disputed that artificial intelligence is a new technology, then suspended the tools until the consultation closed. On May 21, 2026, the Paris court of appeal confirmed that order, reasoning, in one lawyer's account, that no office software could previously claim to reproduce intellectual capacities, as AI tools now do.
Two limits keep this honest. Consultation is not consent: the council's opinion, Républik IT noted, is "jamais bloquant" (never binding). And not every AI tool counts: a review of the year's four decisions finds judges weigh a tool's effects on work, not the mere presence of AI, and notes a Paris court annulled a council's appointment of an expert over a voluntary Copilot test. Employer-side lawyers at CMS Francis Lefebvre call one 2025 ruling's analysis "juridiquement défaillante" (legally deficient).
What did consultation buy MetLife's staff? A signed agreement promising, for two years, no economic layoffs or job eliminations "motivé exclusivement" (motivated exclusively) by the AI tools. Exclusively is a narrow word, and this is one company's contract, not French law. But the jobs question got asked in the open, during the consultation, and the answer went in writing.
That is the test. Connecticut asks after the decision, on a form, with no penalty; France asks before it, and obstruction carries a fine. Oracle's own annual report shows one company under both kinds of system: none of its U.S. employees are union-represented, while "in certain foreign subsidiaries, labor unions or workers' councils represent some of our employees." (It names no country, and no record I found shows a French consultation over this plan.)
Picture the first honest yes
It is 2027. A Connecticut subsidiary of a listed company files a WARN notice; its parent's annual report names AI in the restructuring note, so the lawyers answer section 26 with a yes. The state's first tally records an AI layoff. A legislator cites it, and somebody on cable news says the robots have arrived.
Now add a House discussion draft from Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.), which would have employers estimate "the percentage of job losses attributable to AI" whenever AI is a "substantial factor" in a mass layoff. The first filing prints 40% — a judgment call that looks like a measurement and still stops at the technology line.
The parent's quarterly report, filed the same month, shows capital spending tripling and free cash flow below zero, and no form asks for it. Its French subsidiary, rolling out the same tools, sits through a consultation first. Two workforces, two procedures, one budget — and an American record that says AI did it, because AI is all anyone asked about.
Everyone wants the count; hardly anyone asks about the money
The appetite for AI-layoff data is bipartisan. Republican Josh Hawley and Democrat Mark Warner introduced a bill that "would require major companies and federal agencies to report AI related layoffs to the Department of Labor," with Warner arguing that "Good policy starts with good data." It has sat in committee since November 2025.
The skeptics are serious, too. ITIF's Center for Data Innovation warned that the bill's "vague criteria would generate inconsistent and largely meaningless data" and "create a misleading narrative about AI-driven layoffs." Free-market critics at R Street warn that if the rush of state AI bills is not contained, it will "worsen the patchwork of confusing and costly mandates already mounting." And Fox News's CyberGuy report put the flaw in one line: "Each company decides what counts as an AI-related job impact …"
Labor wants more than a count. The AFL-CIO Technology Institute backs state bills with rules for "preventing and tracking potential displacement," and the UC Berkeley Labor Center records a union contract in which "The Company shall meet and discuss with the Union before implementing new forms of Artificial Intelligence (AI) technology." And researchers counsel patience: economy-wide, Yale's Budget Lab finds its analysis "does not yet clearly indicate an AI-related labor market footprint."
Read together, they share one blind spot. Every proposal here, from the right, the left or labor, asks about the technology. Not one asks about the capital budget.
What does this mean for you?
Read the restructuring note next to the capital line. EDGAR's full-text search covers public filings back to 2001. Search your employer (or its parent) for "restructuring," then find "capital expenditures" in the same filing.
Check the notices yourself. Connecticut posts WARN notices in a public library indexed by year, and New York's dashboard lets you filter and download its notices. Compare the stated reason with what the company told investors.
In Connecticut, watch where the AI answer goes. From October 1, 2026, the law sends it "to the department," not necessarily to the public. A UConn study the same act orders, due January 1, 2027, must address public reporting; if none follows, ask why.
Don't take a round number on faith. When a headline says AI "eliminated" some number of jobs, check whether it's a net headcount change and whether any filing says so.
If you have a union, bargain for notice before the technology arrives, the way that contract does.
Tell Congress which number you want. The House draft was released to gather public feedback before it is introduced. If you'd rather see where the money went than a percentage guess, tell your representative.
The lesson, as I see it
Each of these disclosures asks the headline's question: was it AI? That question deserves an answer. But a company can answer it truthfully and still leave out why the cut came when it did, and Oracle is simply the company whose paperwork makes that easiest to see.
In July I wrote that the cover story and the real thing "turned out to be roommates." Oracle's filings, read beside Bloomberg's reporting, suggest a third tenant — and no layoff form asks about it.
My vote? Keep the AI question and add the one it's missing: when a WARN notice comes from a public company or its subsidiary, put the parent's latest capital-spending figure beside the answer. It would cost the filer next to nothing (the number is already public), and it would let a state's tally carry both stories at once. France asks the jobs question before the software arrives. American law could at least ask it next to the checkbook.
When a company says AI did it, turn the page.