The Grade of the Verb: Reading AI in Workforce Filings

Two Sentences in One Filing

Oracle's annual report for the year to 31 May 2026 says two different things about artificial intelligence and its workforce, and the distance between them is the subject of this paper.

The first sentence sits in the risk factors, under a heading about the disruption caused by periodic restructurings: "the adoption and deployment of AI technologies across our operations have resulted, and may continue to result, in reductions to our workforce"¹. It is past tense and causal. AI adoption has resulted in fewer people.

The second sits in the restructuring note, the place where a company describes the plan it is actually paying for. There, Oracle says management approved plans "to implement certain strategic measures and further improve operational efficiencies, including through the adoption and integration of AI technologies across certain functions and other operational activities"¹. The plan's estimated cost is up to $2.1 billion, $1.8 billion of it recorded in fiscal 2026, primarily employee severance¹. Here AI is not the cause. It is one of the means, nested inside efficiency, alongside unnamed "other operational activities".

Read together, the two sentences say something more useful than either alone: the company will state that AI has reduced headcount, and will not state how much of this plan AI accounts for. That gap is not evasion. It is the honest shape of what most organisations can currently evidence. The mistake is to plan on the stronger sentence as if the weaker one did not exist.

The argument here is narrow. AI attribution for workforce cuts has moved from press narrative into filed text, and that matters. But filed text is graded evidence. It shows how firms choose to frame reductions. It does not show how much work AI has displaced. Leaders who must write, read or act on these rationales should treat filed AI language as a signal of capital and narrative choices, and test it against measured substitution before building workforce plans on it.

Grading the Language

Read closely, filed attribution falls into at least three grades, each licensing a different conclusion.

The monday.com case shows why the strategic grade must be kept separate. Its Form 6-K of 22 July 2026 announces a reduction of approximately 20% of the workforce, with $45–55 million of net charges, to align the organisation with its AI Work Platform². The filing does not say that AI replaces any role. The co-founder's memo to employees, as reported by TechCrunch, said the move "was not made to reduce costs or replace people with AI"⁹. This is a substantial reduction tied explicitly to AI, in which the company explicitly denies substitution. Any reader who counts it as evidence of AI displacing work has added a claim the company declined to make.

The grades matter because they travel. A causal sentence in one company's risk factors becomes, in a board paper elsewhere, "peers are cutting because of AI". A strategic realignment becomes a substitution ratio. Each step drops a qualifier the original filer kept for a reason.

The Scale Check

Numbers sharpen the reading. Oracle's capital expenditure was $55.7 billion in fiscal 2026, against $21.2 billion the year before¹. Restructuring and other expenses were $1.8 billion, against $374 million¹. The restructuring charge is therefore about 3% of the year's capital spend, and Oracle still reported approximately 141,000 full-time employees at the year end¹.

Two inferences follow, both to be held loosely. Severance on this scale cannot by itself fund a capital programme of this size. And the restructuring, large in absolute terms, is small beside the strategic bet. It is entirely compatible with a story in which AI is the reason for the capital programme and the workforce reduction is one way of showing discipline while that programme runs.

That is the strongest version of the sceptical case, and it deserves to be stated plainly rather than waved at.

The Strongest Case Against Taking the Words at Face Value

The sceptic's argument runs as follows. Boards are under pressure to fund AI infrastructure; reducing headcount releases cash and signals discipline; naming AI recasts a cut as strategy rather than weakness. Oxford Economics, as reported by Sherwood News, suggested that AI's role in recent layoffs may be "overstated" and that attributing cuts to AI "conveys a more positive message to investors" than citing weak demand or past over-hiring⁸. On this reading, filed AI language is capital-reallocation cover and carries no information about substitution at all.

There is supporting evidence. A Harvard Business Review survey of more than 1,000 executives, again as reported by Sherwood News, found that 60% of organisations had already reduced headcount in anticipation of AI's future impact, while 2% had made large layoffs tied to actual AI implementation⁸. A Gartner survey of 350 executives at companies with at least $1 billion in revenue found that 80% of those that had piloted AI or autonomous technology reported workforce reductions, and that reduction rates were nearly equal whether the technology was producing high returns or not⁷. Microsoft, announcing role eliminations in July, said the roles were "not being replaced by AI", as quoted by TechCrunch⁹.

The best reply to the sceptic is that filed language is not free. The SEC has already treated exaggerated AI claims as securities violations. In March 2024 it settled charges against two investment advisers, Delphia and Global Predictions, for false and misleading statements about their use of AI, with $400,000 in combined penalties⁵. In January 2025 it brought a settled action against Presto Automation, which had claimed its AI product eliminated the need for human order-taking when, in fact, the vast majority of drive-thru orders required human intervention⁶. A company that writes "AI adoption has resulted in reductions" into a 10-K is accepting some exposure. That makes the sentence more than marketing.

The reply is real, but it does not rescue the strong reading. It sharpens what the sentence can carry. A causal statement in a risk factor is defensible if some reduction somewhere involved AI. It does not have to be large, and the filer has not quantified it. Legal exposure disciplines the existence of the claim, not its magnitude. So the honest conclusion is not that filed attribution is meaningless, nor that it proves displacement. It is that filed attribution is credible evidence that AI was a factor and almost no evidence of how much of a factor. The surveys point in the same direction: reductions are running ahead of measured implementation, and they do not track returns.

Legal exposure disciplines the existence of an AI attribution, not its size. A filing can be truthful and still tell you almost nothing about how much work AI has taken over.

What the Counts Count

The same distinction applies to the aggregate data now quoted in boardrooms. Challenger, Gray & Christmas reported that AI had been cited in 87,714 announced US job cuts by the end of May 2026, 22% of all cuts that year, and that AI accounted for 40% of cuts announced in May alone³. By the end of September, AI had been cited in 120,136 announcements, about 21% of the year's 573,195, and remained the leading reason year to date. But in September it was the fifth most cited reason, at about 9% of the month's total⁴.

Two further facts belong beside those figures. Total announced cuts through September were down 39% on the same period of 2025⁴. And these are announcements, classified by the reason employers give. They are a count of citations, not a measure of work displaced. A rising AI share inside a falling total says that AI has become the preferred explanation for a smaller volume of cuts. It does not say that AI is removing more jobs than it did. The monthly share moves sharply, from 40% of May cuts to 9% of September cuts, and should not be read as a trend in either direction.

Three Things That Look the Same on Paper

For the workforce leader, the practical problem is that three quite different organisational events produce near-identical filed language.

1. Substitution. Specific tasks are now performed by a system, the people who performed them are no longer needed for that work, and output holds or improves. This is the only one of the three in which AI has displaced work.
2. Anticipation. Headcount is reduced or hiring slowed on the expectation that AI will absorb the work later. The tasks still exist. Someone, somewhere, is still doing them.
3. Capital reallocation. Cost is taken out to fund investment, including AI investment, or to correct earlier over-hiring. AI is the reason for the spending, not the mechanism of the cut.

A composite illustration makes the difference visible. Consider a customer operations function of 1,200 people in a large software company, presented here as an illustration rather than an observed case. The board paper records a 15% reduction "enabled by AI-assisted case handling". Behind that single phrase, three different things may be true. In one version, a deployed assistant resolves routine password and billing queries end to end, the volume handled by people has fallen by a measured amount, and the 180 roles that handled it have gone. In a second, the assistant is in pilot, the volume has not moved, and the 180 roles were removed on the forecast; the remaining team absorbs the work through overtime and longer queues. In a third, the function was cut to meet a cost target set when the capital budget was approved, and "AI-assisted" was the most defensible label to hand.

The filed sentence is identical in all three. The workforce consequences are not. In the first, the organisation has a real capability shift to manage: fewer people doing different work, a new dependency on a system, and a need to keep enough human expertise to handle exceptions. In the second, it has a capability gap disguised as a saving, which tends to surface later as service decline, quiet rehiring or contractor spend. In the third, it has made a capital decision and should account for it as one, including what it will do if the investment case slips.

This is why the grading matters for workforce and capability planning in particular. A workforce plan built on the second or third event, but labelled as the first, will under-invest in the redeployment, reskilling and exception-handling capacity that genuine substitution requires, and will over-estimate the capacity that remains.

The Tests Before the Sentence

The useful question is therefore not whether to mention AI in a workforce rationale, but what must be true before the stronger grade of language is used. Three tests follow from the mechanism. Each applies to one's own rationale before it is written and to a peer's filing after it is published.

1. Can the tasks be named? Substitution leaves a trace at task level: which activities are now performed by a system, at what volume and with what error rate. If the rationale cannot name them, the honest grade is anticipation or reallocation.
2. Does unit cost fall while output holds? Genuine substitution should show up as falling cost per unit of output in the affected function, not just falling headcount. A reduction that shifts work onto remaining staff, contractors or longer cycle times has moved cost rather than removed it.
3. Is there later rehiring? If the work was never absorbed, it returns as backfill. Tracking rehiring, backfill and contractor spend in the same functions over the following year is the cheapest available test of whether the original attribution was substitution.

None of these tests is exotic. What is unusual is applying them before the sentence is disclosed. The Presto action shows the cost of getting the order wrong: the claim that AI had eliminated human work was tested against what actually happened in the drive-thru, and it failed⁶. A workforce rationale that says "AI has resulted in reductions" now sits in the same category of claim, and disclosure controls should treat it accordingly. That means the finance, legal and HR owners of a restructuring agreeing which grade of language the evidence supports, rather than letting the strongest available sentence migrate from investor narrative into the filing.

Where This Reading Stops

Three limits apply. No filing examined here quantifies the share of any reduction caused by AI, so the central claim concerns the absence of that evidence, not its direction. The reading of Oracle rests on its risk factors, restructuring note and financial statements; the management discussion in the same report may frame the plan further. And the HBR and Oxford Economics findings are cited as reported in secondary press, from an evidence base that is mainly US-listed technology. Where task automation is documented, customer support being the usual example, the substitution grade may be earned and the argument applies with less force.

Nor does the paper claim that AI-attributed cuts are failing. The Gartner finding that reductions did not track returns is correlational⁷. It is a reason to ask the tests above, not evidence of an outcome.

The Question That Changes

The arrival of AI language in binding filings changes what leaders can say and be held to. It does not change what they know. For the moment, the most informative thing about a filed AI attribution is usually the grade of the verb: resulted, including through, align with. The causal sentence and the operative note in a single annual report can differ in strength, and the difference is evidence.

So the question to bring to the next workforce rationale, one's own or a competitor's, is not "is AI behind this?" It is "which of the three events is this, and could we name the tasks?" Where the answer is substitution, the organisation has a capability transition to manage and should resource it as such. Where it is anticipation or capital reallocation, the honest sentence is a different one, and the workforce plan should be built on that sentence instead.