Salesforce Put Its CRM Inside Claude

Salesforce reported Q2 FY27 this afternoon and the stock jumped double digits in extended trading. Most of the coverage is quoting the wrong number.

Adjusted EPS was $5.90. That matched the LSEG consensus. It was not an 80 percent beat, and if you saw that figure somewhere today, it came from a stale consensus estimate. The number that actually jumped was GAAP: $4.29 a share, up 119 percent year over year, on net income of $3.53 billion against $1.89 billion a year ago. The lift came from a $2.6 billion gain on strategic investments, tied in part to Salesforce's stake in Anthropic.

That distinction matters more than it sounds. Non-GAAP is the operating number. GAAP is the one that moved. So the loudest number in the AI CRM's quarter was a mark on an equity position, not a product line.

The short version

Salesforce's Q2 FY27 adjusted EPS of $5.90 matched consensus; the 119 percent GAAP jump came from a $2.6 billion strategic investment gain tied to its Anthropic stake, now worth around $5 billion and roughly two thirds of its entire investment portfolio. The real news landed the same afternoon: Claudeforce, which ships Salesforce as a plugin inside Claude with 37 prebuilt sales skills, the first time Salesforce has put its "force" suffix on someone else's product. Benioff's own quote calls Salesforce "a trusted enterprise harness," which is the substrate position stated out loud. Keep the scale honest, though: Agentforce ARR of $1.5 billion against a $46 billion guide is about 3 percent of the business and the seats did not collapse, so this is still an attach motion. The metric to watch is Agentic Work Units, 3.2 billion in the quarter and 7 billion to date, because the largest seat-based software company on earth just put units of completed work on an earnings slide.

The actual headline

While everyone was reading the EPS line, Salesforce and Anthropic announced Claudeforce.

Salesforce is shipping its CRM as a plugin inside Claude. Thirty-seven prebuilt sales skills, so a seller can reason over live revenue context, update pipeline, and take governed action without ever opening Salesforce. Pilot customers now, preview next month.

It is the first time Salesforce has ever put its "force" suffix on another company's product.

Sit with that. The most valuable interface in enterprise software just agreed to be a plugin in someone else's chat window.

Benioff's quote in the release names the position without flinching: "with AIforce, our trusted enterprise harness, we're unlocking the data, workflows, business logic, actions, and governance inside Salesforce and making it available to every agent, model, and interface."

Harness. Every agent, model, and interface. That is a company describing itself as plumbing, on purpose, in a quarter where it also raised guidance.

I made a version of this argument last month about Alation moving from catalog to governance layer: the serious companies stop trying to own the agent and start trying to be the thing every agent plugs into. I have a longer piece on the pattern going up this Saturday. I wrote it before today's announcement, which is either good timing or a sign the move was obvious enough that I should have expected Salesforce to make it first.

The catch is that the move has a bill attached. Plumbing has no daily-active-user ratio, because nobody opens plumbing. Salesforce is doing this from strength rather than desperation, which is the version worth studying, and it still means handing the daily habit to Claude and keeping the governed record underneath.

The number I actually care about

Agentic Work Units. Seven billion delivered to date across Agentforce and Slack, with 3.2 billion of those in the quarter, up 97 percent quarter over quarter.

Read that line again. The largest seat-based software company on the planet is now putting a work-completion metric on its earnings slide, next to revenue and cRPO.

You can argue the metric is soft. It is. Salesforce defines it, Salesforce counts it, and nobody outside can audit it. Fine. It still matters, because metrics on earnings slides become the metrics buyers ask about in the room. Once your biggest competitor reports units of work completed, per-seat stops being the default and becomes a position you have to defend.

I wrote about Sierra's version of this two weeks ago, getting paid on resolution. Sierra is a startup with the freedom to price that way from day one. Salesforce has thirty years of seat contracts and a renewal base to protect, and it is still walking toward the same unit. That is the tell.

The disclosure nobody is quoting

One line in the release deserves more attention than it is getting:

Agentforce ARR exceeded $1.5 billion, up over 240% Y/Y. Effective Q2 FY27, Agentforce ARR includes our AI offerings, Slackbot and Headless 360.

The definition changed in the same quarter the growth rate was reported. Some of that 240 percent is scope, not sales, and there is no way from outside to tell how much.

Salesforce disclosed it plainly, which is to their credit. Almost nobody covering the print has mentioned it. If you are going to cite 240 percent in a board deck this week, cite the footnote with it, because someone in the room will find it eventually and you want to be the person who brought it up.

Keep the scale honest

$1.5 billion of Agentforce ARR against a $46.1 to $46.4 billion full-year guide is about 3 percent of the business. Revenue grew 11 percent with roughly four points of that coming from Informatica, so organic growth is still mid single digits. The guidance raise was $200 million, and the composition is instructive: $100 million organic, $200 million from the pending Contentful and Fin acquisitions, less $100 million of FX. Two hundred forty percent growth on 3 percent of revenue does not reaccelerate a company this size, not this year and probably not next.

Which cuts both ways, and this is the part most takes will miss.

If agentic AI were going to eat seat-based software first anywhere, CRM was the obvious candidate. Service seats especially. It has not happened. Salesforce sold agents hard into its own base and the seats stayed. So far this is an attach motion, not a replacement cycle.

Meanwhile the company is running a $25 billion accelerated share repurchase, with 103 million shares already delivered under it, and returned another $364 million in dividends this quarter. When the AI line cannot move the growth rate yet, the balance sheet moves EPS instead. Both things happened in this print. Only one of them is a product story.

What I am taking into Heidi's roadmap

Three things, and they are not abstract for me. I am building an agentic productivity platform, and this quarter changed what I think we have to be able to show a buyer.

Define the work unit and count it. Tokens are activity, not value. A work unit is a task a person would otherwise have done, with a start, an end, and a result you can point at. If you cannot report it, you cannot price against it, and you will get benchmarked by a competitor who can.

Consumption pricing tells the truth faster than seats do. Credits purchased and never burned are churn on a delay. That is uncomfortable for a forecast and enormously useful for a product team, because it surfaces the gap between what someone bought and what actually worked inside their workflow. You find out in week six instead of month eleven.

Rent the model, own the record. Today's print is evidence on both sides of this. The model layer captures enormous value on its own, and Salesforce booked $2.6 billion of that as investment income rather than product revenue. You do not beat that by building a slightly better wrapper around the same reasoning engine. You beat it by owning the accumulated observation record for a customer: the entities, the decisions, the receipts of what your agents did and what happened next. That is the asset that does not reprice when the next frontier model ships. It is also, not coincidentally, exactly what Salesforce is now selling access to.

Investor Day is September 16, alongside Dreamforce. That is where the work-unit framing either becomes a pricing model or stays a slide. I would bet on the former, and I would rather be early to the unit than right about the timing.

One thing to try this week: take a single workflow your product touches, define one completed work unit for it, and start counting. Not tokens, not sessions, not monthly actives. Work. You will learn more from that number in a month than from another quarter of usage dashboards.

Sources: Salesforce Q2 FY27 press release, Salesforce and Anthropic announce Claudeforce, Quartz on the print, CNBC on the Anthropic partnership, TechCrunch on Anthropic's Series H.

Related answer: What did Salesforce's Q2 FY27 earnings say about agentic AI?