Blog · Research · 31 August 2026 · 10 min read

Salesforce's Filing Counts 7 Billion AI Tasks. It Does Not Count Jobs.

AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)

On 26 August 2026, Salesforce filed a Form 8-K carrying its second-quarter fiscal 2027 results. The third bullet of the company highlights reads: "7.0 billion Agentic Work Units ('AWUs') delivered to date across Agentforce and Slack, with 3.2 billion in Q2, growing 97% quarter-over-quarter".

Seven billion is the sort of number that gets repurposed within a day — as evidence that AI is doing an enormous amount of work people used to do, or that the agent economy is finally producing something measurable.

Both readings run past the definition. Salesforce supplies one, in the same document, and it is narrow.

One date note first. The filing is from 26 August; this was written on 31 August. An EDGAR full-text search run while writing returned zero filings mentioning "artificial intelligence" dated 29, 30 or 31 August, against 458 on Friday 28 August — two of those days were a weekend, and the third is the day of writing, when same-day filings may not yet be indexed. Nothing fresher exists on this beat. Saying so is part of the reporting.

What is an Agentic Work Unit, in Salesforce's own words?

Right at the end of the release, under "Other Metrics", Salesforce defines the thing it just counted seven billion of:

"The Company defines an Agentic Work Unit ('AWU') as a measure of discrete tasks executed by AI agents in production across the Salesforce platform, including Agentforce and Slack. AWUs represent the conversion of generative AI capabilities into measurable business outputs, such as resolving customer cases, updating records or triggering automated workflows."

Read the examples, because they set the scale of the unit. Resolving a customer case. Updating a record. Triggering an automated workflow. One AWU is one of those, once — an event counter on a software platform, counting the smallest possible events.

Note also what the definition does not contain. There is no quality gate — nothing about whether the case was resolved correctly, or whether a person had to touch it afterwards. There is no counterfactual — nothing about whether a human previously did the task, or whether the task existed before an agent could perform it cheaply. And there is no conversion factor: no hours, no full-time equivalents, no cost per unit. The filing does not offer one anywhere.

So the arithmetic people are quietly doing in their heads when they see seven billion — tasks divided by some notional human throughput, equals jobs — has no denominator in this document.

Is AWU a standard metric anyone else reports?

No, and this is checkable. Salesforce introduced the measure six months ago, in its fourth-quarter fiscal 2026 release of 25 February 2026, whose bullet reads "Introduced Agentic Work Units ('AWUs') to measures tasks accomplished by an AI Agent, with 2.4 billion AWUs delivered to date across Agentforce and Slack, growing 57% quarter-over-quarter".

A full-text search of SEC EDGAR for the phrase "Agentic Work Unit", run on 31 August 2026, returns three filings. All three are Salesforce's own quarterly earnings exhibits: February, May and August 2026. No other registrant uses the term.

That is not a criticism — companies invent operating metrics constantly, and publishing a written definition is better practice than the alternative. It does mean there is no peer number to compare it against and no auditor's opinion on it. Non-GAAP measures in the same release get a page of reconciliation; operating metrics like this one get a paragraph.

Do the AWU numbers reconcile across quarters?

Line up the three releases and one number does not close.

The cumulative figures behave: 3.8 billion at the end of the first quarter plus 3.2 billion in the second gives exactly the 7.0 billion reported. The growth rate does not. Subtracting the two earlier cumulative totals implies about 1.4 billion AWUs in the first quarter, and 97% growth on 1.4 billion is roughly 2.8 billion, not 3.2 billion.

Rounding cannot close a gap that size. Salesforce publishes these to one decimal place, so the first quarter's implied volume can be anywhere between about 1.3 and 1.5 billion; 97% growth on the top of that range is under 3.0 billion, while the smallest number that rounds to "3.2 billion" is 3.15 billion. That subtraction is ours, not Salesforce's, and Salesforce never publishes a first-quarter in-quarter figure to check it against.

We are not alleging an error. An unaudited usage metric can legitimately change basis between quarters — Salesforce says openly, two bullets away, that its Agentforce ARR figure changed scope this quarter to include Slackbot and Headless 360. No equivalent note appears for AWUs. A metric with no audit, no peer and no reconciliation table is one whose quarter-to-quarter arithmetic you cannot check from outside.

Why did earnings per share nearly double?

The other number from this release travelling as an AI story is the earnings beat. Non-GAAP diluted earnings per share came in at $5.90, up 103% year over year, against a consensus of $3.27 reported by MarketWatch. The stock rose 13% in extended trading.

Almost none of that gap is operations. Footnote 3 to Salesforce's own income statement says so plainly: "gains on strategic investments impacted GAAP diluted net income per share by $2.43 and $0.00" for this quarter and the year-ago quarter, "and non-GAAP diluted net income per share by $2.53 and $0.00". Take $2.53 off $5.90 and you get $3.37 — which is precisely the like-for-like figure MarketWatch reports: "Stripping out the gain on investments, Salesforce's adjusted earnings would be $3.37 a share."

Against a $3.27 consensus, that is a beat of about 3%, not the 80% the headline pair implies. MarketWatch identifies the gain as a paper gain on Salesforce's stake in Anthropic. The filing itself never says Anthropic — the word does not appear in the release at all. It reports a line called "gains (losses) on strategic investments, net" of $2,613 million for the quarter, against $6 million a year earlier.

Salesforce is careful about this in its own guidance. The forward-looking table's footnote states that projected earnings per share "assumes no change to the value of our strategic investment portfolio as it is not possible to forecast future gains and losses". Guidance for the next quarter, on that no-gain basis, is $3.42 to $3.44 — the neighbourhood of the $3.37, nowhere near the $5.90.

Whatever seven billion AWUs are worth, this quarter's earnings jump is not the evidence. It came from a revaluation of an equity stake.

What did the operating lines actually do?

One level up from earnings per share, the income statement is flatter than the headline suggests. Total revenue rose 11%, from $10,236 million to $11,345 million. GAAP income from operations went from $2,332 million to $2,331 million — down one million dollars on a billion dollars more revenue. GAAP operating margin fell from 22.8% to 20.5%; the non-GAAP margin slipped from 34.3% to 34.1%.

Read that carefully rather than triumphantly. Flat operating income is not proof that agents do nothing; the same quarter absorbed an acquisition, and the release makes no claim about AI's effect on the operating line in either direction. What it does establish is the narrow thing: nobody reading this release can point at the operating result and call it the AI dividend.

Where would a workforce cut show up in this filing?

In two places.

The first is the 8-K's item list. A newly board-approved plan with material exit or disposal costs is disclosed under Item 2.05. The filing index for this 8-K shows Item 2.02, Results of Operations and Financial Condition, and Item 9.01, Financial Statements and Exhibits. There is no Item 2.05.

That fact carries exactly one meaning: no new material restructuring plan was announced with these results. It does not mean nobody left. Attrition, a hiring freeze, a quiet team closure, a non-severance reorganisation — none of those trigger an Item 2.05, and none would appear here. The words "headcount", "layoff" and "reduction in force" appear zero times in the release, which is a fact about the vocabulary of an earnings document and not a fact about Salesforce's workforce.

The second place is the restructuring expense line, and it moved. Restructuring ran to $94 million in the quarter against $4 million a year earlier, and $174 million against $40 million for the six months. The release defines what sits in that line: "costs associated with a formal restructuring plan and may include employee notice period costs and severance payments, lease or contract termination costs, asset impairments, accelerated depreciation and amortization and other related expenses." Full-year guidance carries restructuring and acquisition-related costs at 0.8% of the revenue midpoint.

So Salesforce is spending materially more on restructuring than it was a year ago, and the release attaches no headcount, no cause and no AI linkage to any of it. Severance and lease terminations are bundled together in one number. Anyone claiming this filing shows AI cutting jobs at Salesforce, and anyone claiming it shows the opposite, is reading past the same silence.

What did Salesforce's own finance chief say the number means?

This is the part that most cuts against the displacement reading, and it comes from the company. MarketWatch reports that Robin Washington, Salesforce's President and Chief Financial and Operating Officer, "noted that higher usage of agentic work units is complementing core seat growth."

Seats are how Salesforce charges: per user, per month. A finance chief describing agent usage as complementary to seat growth is describing an upsell running alongside headcount-linked revenue, not a replacement for it. That is a commercial claim made by the company selling the product, and should be weighted accordingly — but it points the opposite way from the story the seven-billion figure is being used to tell. The word "seat" does not appear in the earnings release at all, so that framing lives in the reporting, not in the filed document.

How is this different from adoption rising while headcount holds?

Earlier this month we covered the survey version of this question: the ONS surveyed 38,637 UK businesses and a Fed executive panel asked nearly 750 executives, and both found adoption rising steeply while aggregate headcount mostly held. That post is about many firms, measured by outside statisticians, answering a standard question about what already happened.

This one is the opposite object. One company, one self-defined counter, no external statistician, no comparison group, and a metric whose own definition rules out the inference people want to draw from it. The two pieces are not two takes on the same evidence. They are a population measurement and a company disclosure, and the second cannot be aggregated into the first, no matter how large its number gets.

Is Salesforce unusual here?

Not especially. Autodesk filed its own second-quarter results the next day, and the only executive quote in the release is entirely about AI — chief executive Andrew Anagnost opens with "AI turns connected data and context into actionable project intelligence that can ease endemic capacity constraints". Its restructuring, other exit costs and facility reductions line came in at a $1 million credit for the quarter against a $6 million charge a year earlier, and $29 million against $111 million for the six months. Its 8-K carries Items 2.02, 7.01 and 9.01 — again, no Item 2.05.

Two large software companies, one week, heavy AI language in both, restructuring lines moving in opposite directions, and neither release connecting the two. That is the state of the evidence at the company-filing level: AI in the narrative, and a separate accounting line it never touches.

Compare that with the exception. HP did file an Item 2.05 naming artificial intelligence, with a headcount range attached — and even there, the only mechanism the filings named was a voluntary early-retirement offer. When a filing genuinely connects AI to a workforce reduction, it looks like that. It does not look like a task counter in a highlights list.

What would actually settle this, and when?

Not a company metric, and not a headline. The measurements pointed at the labour market as a whole arrive within days of this being written. The Bureau of Labor Statistics publishes JOLTS with July 2026 data on 1 September 2026 at 10:00 a.m., and the August 2026 Employment Situation on 4 September 2026 at 8:30 a.m.

Neither will settle the AI question either, and it is worth being blunt about why: no BLS instrument has a field for cause. They measure openings, hires, quits, layoffs and payrolls; they do not ask why. What they can do that no company disclosure can is cover the whole economy on one consistent definition.

While waiting, the useful discipline is the one this filing rewards. Find the definition before you use the number. Check whether the figure sits above or below the operating line, and which item of the 8-K it was filed under. And check whether a number has a primary source at all — the "Oracle may cut 30,000 more jobs" figure currently circulating does not, with no filing, no company statement and no WARN notice behind it that we could find.

Salesforce's disclosure is better than most. It published a definition alongside its metric, and the definition is the thing that answers the question. Seven billion discrete tasks executed by software. Not seven billion tasks taken from anyone, not a headcount, not a saving. The company counted what it counted, and said so.

Sources: Salesforce, Inc. (NYSE: CRM), CIK 0001108524. Form 8-K filed 26 August 2026, accession 0001108524-26-000187, Exhibit 99.1 and filing index; Q1 fiscal 2027 earnings release, Exhibit 99.1 to Form 8-K filed 27 May 2026; Q4 fiscal 2026 earnings release, Exhibit 99.1 to Form 8-K filed 25 February 2026. Autodesk, Inc. (NASDAQ: ADSK), CIK 0000769397, Form 8-K filed 27 August 2026, Exhibit 99.1 and filing index. Consensus estimates, the $3.37 like-for-like figure, the Anthropic identification and the CFOO's remark: MarketWatch/Dow Jones, Christine Ji, via Morningstar. Release dates: U.S. Bureau of Labor Statistics release calendars.