Salesforce Redefined Its Headline AI Metric in the Quarter It Reported It

What did Salesforce report, and what did most coverage leave out?

Salesforce closed its second quarter of fiscal 2027 on 31 July and reported on 26 August. Working from the company's own results release, total revenues were 11,345 million dollars against 10,236 million a year earlier, which is growth of 10.8%. Current remaining performance obligation was 33.5 billion dollars, up 14% year on year and in constant currency. GAAP diluted earnings per share were 4.29 dollars, up 119%, and non-GAAP diluted earnings per share were 5.90 dollars, up 103%. Operating cash flow was 1.3 billion dollars, up 71%.

On the artificial intelligence line, Agentforce and Data 360 annual recurring revenue reached nearly 3.9 billion dollars, up over 210% year on year, and Agentforce annual recurring revenue on its own exceeded 1.5 billion dollars, up over 240%.

Then comes the sentence that matters more than any of those, and that most coverage of the quarter did not carry. In the release's own words, effective the second quarter of fiscal 2027, Agentforce annual recurring revenue includes the company's AI offerings, Slackbot and Headless 360.

Why that one sentence breaks the comparison everyone wants to make

The obvious thing to do with a 1.5 billion dollar figure is set it against the 1.2 billion reported a quarter earlier and call the difference growth. That comparison cannot be made, because the two numbers do not describe the same set of products. The later figure counts Slackbot and Headless 360. The earlier one did not.

The direction of the distortion is knowable even if the size is not. Widening a definition can only add revenue to the later number, so the true like-for-like growth in Agentforce is lower than the arithmetic suggests, by an amount the company has not quantified. Anyone who published a quarter-on-quarter growth rate for this line, including an earlier version of this report, published a number that does not mean what it appears to mean.

The usage figure is real, and it measures something wider

The release reports 7.0 billion Agentic Work Units delivered to date across Agentforce and Slack, with 3.2 billion of those in the second quarter, growing 97% quarter over quarter. That growth figure is disclosed, specific, and the most concrete evidence in the filing that automated work is being consumed at scale.

It is also counted across Agentforce and Slack together, which is a wider surface than the Agentforce revenue line, and a work unit is defined by the vendor. So the honest reading is that consumption is growing very fast on a broad surface, and that the filing does not support a clean ratio between consumption and the revenue attached to it. That ratio is the number a buyer would most like to have, and it is precisely the one nobody has published.

What the same period did to the measurement layer

While the resolution layer was being bid up, the feedback layer was being written down, and those numbers are not in dispute.

On 22 April, Thoma Bravo agreed to hand Medallia to its lenders in a debt-for-equity swap that wiped out 5.1 billion dollars of equity against a 6.4 billion dollar take-private completed in 2021, with 2.8 billion dollars of debt outstanding and a syndicate including Blackstone, KKR, Apollo Global and Antares Capital converting into majority ownership. It is the largest equity write-off in Thoma Bravo's history.

On 18 May, Qualtrics finalised its 6.75 billion dollar purchase of Press Ganey Forsta. Reporting on the financing describes banks carrying more than 500 million dollars in paper losses on a 5.3 billion dollar debt deal, and JPMorgan halting that deal over concerns about disruption from artificial intelligence. Qualtrics paid more for Press Ganey Forsta alone than Thoma Bravo paid for the whole of Medallia four years earlier. Both were named Leaders in the same Gartner Magic Quadrant for voice of the customer platforms published in March.

Three headline numbers, three missing definitions

Put the quarter's most quoted figures next to each other and a pattern appears that is more useful than any single one of them.

Agentforce annual recurring revenue was redefined in the quarter it was reported. Agentic Work Units span two products and rest on a unit the vendor defines. And the 3.6 billion dollar purchase of Fin, formerly Intercom, was announced on 15 June against a published claim that its agent resolves an average of 76% of support volume end to end, a figure with no published denominator, so nobody outside the company can say what counts as resolved.

None of that is evidence of bad faith. It is evidence of a category still deciding what to count, at exactly the moment buyers are being asked to make multi-year commitments against the counting. That is also a plausible part of why the measurement half is being repriced downward: a market that cannot yet agree on its own units is a hard market to prove returns in.

What a buyer should do differently this quarter

Ask for the definition before you compare the number. For a resolution rate, get the denominator in writing, including the treatment of handoffs, abandonments and reopened tickets. For a consumption metric, get the unit defined and get it fixed for the contract term, because a vendor that can redefine a unit can reprice you without changing a price.

Price measurement and resolution as two purchases even when one vendor sells both, since the two halves are being valued in opposite directions and a bundle hides which half carries the cost. And when a platform in the measurement layer changes hands, ask the new owner about roadmap funding rather than continuity of service, because a creditor recovering a position and a sponsor pursuing growth do not invest alike.

What we could not establish

How much of the Agentforce revenue line comes from the newly included products, because the company did not quantify the change. Whether any of that revenue is net new against migrated from existing service licences, because the release does not break it out. What a resolved conversation is at Fin. And what Medallia or Qualtrics actually earn, since both are privately held and neither files.

This report was corrected before publication. Its first draft asserted a quarter-on-quarter growth rate for Agentforce revenue and a comparison against usage growth, both drawn from secondary coverage. Reading the filing itself showed the series had been redefined in the same quarter, which made the comparison unusable. The error is recorded here rather than removed quietly, because a publication that asks readers to inspect its reasoning has to show the corrections too.