Claudeforce Puts Two Vendors on Your Salesforce Bill

A Client Forwarded Me the Announcement and Asked If It Is Safe. That Is the Wrong Question.

Salesforce and Anthropic announced Claudeforce on 26 August, on the Q2 FY27 earnings call. Within a day my inbox had the shape these things always have: forwarded links, a screenshot of a LinkedIn post, and one question. Is our data safe if we do this.

It is a fair question and it has an answer. The answer is mostly yes, and it was mostly yes ten months ago, which I will come back to.

The question nobody forwarded me is the one I would actually worry about. Claudeforce does not bill the way Salesforce has billed for twenty-seven years. It meters. And the Claude part of it does not appear on your Salesforce invoice at all, because Anthropic sends you a separate one.

What Was Actually Announced

Three things, plus a foundation underneath them.

Salesforce in Claude is a Cowork plugin carrying 37 prebuilt sales skills, built jointly by the two companies. Meeting preparation, deal health review, pipeline management. A seller works in Claude, reads live CRM context, and takes action without opening Salesforce.

Claude in Salesforce is the other direction. Claude is now the reasoning model behind the Atlas Reasoning Engine, it powers Agentforce Vibes and Agentforce Coworker by default, and it is available in Agent Builder.

Slack gets Claude as its default model, powering Slackbot.

Underneath all three sits AIforce, which Salesforce describes as its "trusted enterprise harness that brings all your business data and workflows to any agent through MCP servers, APIs, and CLI tools, without complicated and costly integrations." That sentence is doing more work than it looks like it is doing. It is the architectural claim that Salesforce can be a backend for somebody else's interface and still be Salesforce.

The Billing Model Is the Story

Here is the part that will reach your budget.

Salesforce charges this new usage through its headless consumption pricing. In Salesforce's own framing, "depending on your edition of your user license within Salesforce, you effectively get more incremental access to more and more API calls." Read that carefully. Your entitlement is denominated in API calls, and an agent working a pipeline makes considerably more of them than a human clicking through it.

Then there is the second invoice. Customers contract Anthropic separately for the Claude inference itself. A Salesforce executive put it plainly: "You can't buy this on one piece of paper at the moment."

I have not seen an enterprise software launch admit that so directly, and I think it deserves credit rather than mockery. But sit with what it means procurement-side. The thing your sales org wants to pilot has two suppliers, two contracts, two renewal cycles, and quite possibly two different budget owners. The person who signs your Salesforce renewal is not usually the person who signs an AI inference agreement.

A word on the numbers, because this is where I have seen people already get it wrong. Salesforce has published Flex Credits pricing for Agentforce: a minimum of 100,000 credits for 500 US dollars, with an agent action consuming 20 credits, so roughly 10 cents per action. Flex Credits are consumed only when an agent action actually executes inside Salesforce; simple MCP tool reads and writes typically do not consume them. That distinction matters more than the rate does.

But those are Agentforce numbers, not Claudeforce numbers. No price was disclosed for Claudeforce, and Salesforce's own release says pricing and packaging are subject to change. Anyone quoting you a cost per seller today is extrapolating.

Salesforce Already Told You This on the Earnings Call

The consumption shift is not something I am reading into the announcement. Salesforce said it out loud, in the same session, as good news.

Miguel Milano, President and COO: "Fifty percent of the bookings came from customers refilling the tank: they consume Flex Credits, they want more."

Marc Benioff, Co-Chief Executive and Founder: "Customers want that kind of diversity in pricing, we have created a high level of flexibility, and that has really expanded our ability to sign very large transactions."

Refilling the tank. That is a metaphor about a thing that empties. The market liked it — Q2 revenue came in at 11.3 billion US dollars, up 11 percent year on year, and the stock rose 12 percent.

Now the other side, because a post that only quotes the vendor is a press release with a byline. Gartner warned in October 2025 that consumption models like Flex Credits "can be credit-based, with rates that may change unilaterally by the vendor, exposing buyers to unexpected cost increases," and predicts that all-you-can-eat agentic license agreements will convert to "defined quantity contracts at the end of the agreement."

Both of those things can be true. Consumption pricing genuinely is more flexible for a customer whose usage is lumpy, and it genuinely does transfer forecasting risk from the vendor to you. Under per-seat licensing you knew your Salesforce cost the day you signed. Under consumption you find out in arrears. That is not a scandal. It is a different job, and somebody in your organisation now has to do it who did not have to before.

Why Everyone Is Quoting the Wrong Line

The line doing the rounds this week is that Claude is now the first model provider fully integrated inside the Salesforce Trust Boundary.

That is true. It has also been true since 14 October 2025, when Salesforce and Anthropic announced exactly that, along with all of Claude's traffic being contained within the Salesforce virtual private cloud, delivery through Amazon Bedrock, a regulated-industries collaboration covering financial services, healthcare, cybersecurity and life sciences, and CrowdStrike and RBC Wealth Management as early adopters.

Salesforce's own Claudeforce release treats it accordingly, in the present tense, as settled background: "Through Amazon Bedrock, Claude is available within the Salesforce Trust Boundary." Not a new capability. A precondition for the new capability.

The word carrying the falsehood is "now." Strip it out and the sentence is accurate and ten months old.

This is the second time in eight days I have had to write this paragraph. Last week the claim was that Winter '27 was shipping a 41-skill Salesforce plugin for Claude Code; the count was right and the release framing was invented, because that plugin has shipped from GitHub since November 2025. I would rather not make a habit of this. But when a real announcement is large enough, the recycled claims travel further than the actual ones, and the actual ones here are more interesting anyway.

What Regulated Orgs Should Actually Take From This

What is genuinely settled: the Trust Boundary is real, Claude traffic runs through Amazon Bedrock inside Salesforce's virtual private cloud, and this predates Claudeforce by ten months. Permissions inherit through the MCP servers, so actions are governed by the platform's existing permissions and sharing rules. Claude does not receive unrestricted access to your CRM. If your institution already approved Salesforce's environment, the perimeter review is genuinely shorter.

What is not settled is everything downstream of the perimeter. David Girvin, founder and chief executive of Assury, argues that model-in-the-loop review is fundamentally unreliable in regulated environments, on the grounds that even the strongest models miss a meaningful share of violations. That is a claim about output quality under audit, and no network diagram answers it.

I want to be careful about the boundary of what I can tell you here. Nothing in the published material addresses business associate agreements, or any specific compliance regime, and I am not going to speculate about one. What I will say is structural: a perimeter guarantee is not a compliance programme. Knowing your data stayed inside a VPC tells you nothing about whether an agent gave a customer advice you are obliged to supervise.

Those are two different reviews, and in most organisations they are run by two different people. Do not let the first one being easier convince anyone that the second one happened.

Nothing Here Is Generally Available

| Component | Status |
| --- | --- |
| Salesforce in Claude | Select pilot customers only |
| Salesforce in Claude, open beta | Expected September 2026 |
| Service, marketing and commerce skills | Q3 |
| Additional prebuilt skills | Beginning late 2026 |
| Claude in Agentforce and Slack | Shipping |
| Claudeforce pricing | Not disclosed, subject to change |

Read that table before you promise anyone a date. The headline capability is in pilot, the open beta has not opened, most of the skill catalogue does not exist yet, and the pricing is explicitly provisional. Anyone planning a Q3 rollout is planning around a beta whose commercial terms can move.

What To Do Right Now

1. Find out who would own the Anthropic contract. It is not your Salesforce renewal and it is probably not your Salesforce buyer. If the answer is a shrug, you have found the thing that will delay your pilot by a quarter, and you have found it early, which is the only good time to find it.

2. Ask for a consumption forecast before you pilot, and ask what happens if rates change. Model your pipeline volume against API-call entitlements by license edition. Then ask, in writing, what notice you get if credit rates move. Gartner's warning is about unilateral changes, and the answer to that question belongs in a contract, not a conversation.

3. Separate the perimeter question from the compliance question. The perimeter answer is good and it is old. The supervision, audit and evidence questions are open and belong to a different reviewer. Book two meetings, not one.

4. Wait for open beta before you commit to a date, then test with real permission sets. Permissions inherit through MCP, which is exactly why an admin login will tell you nothing useful. Run the pilot as a regular seller with a regular profile, and see what the agent can and cannot reach.

Trying to work out what Claudeforce would actually cost your org, as opposed to what it would do? Let's talk — I will help you get the forecast and the contract questions straight before the pilot, not after the first invoice.