SAN FRANCISCO, 29 AUG 2026 — Salesforce and Anthropic have announced Claudeforce, which puts Claude at the centre of Salesforce's products and Salesforce's CRM inside Claude. The quietest part of the deal has the widest reach: Claude is now the default model in Slack, six days after Slack launched a deliberately multi-vendor agent surface with four launch partners.
What was announced
The 26 August announcement covers two distinct integrations, and they are not symmetrical.
Going one way, Salesforce in Claude is a Claude plugin carrying 37 prebuilt sales skills built jointly by the two companies, covering meeting preparation, deal health review and pipeline management. It is with pilot customers now, with an open beta expected in September.
Going the other, Claude becomes the reasoning model for Salesforce's Atlas Reasoning Engine and the default large language model in Slack, powering Slackbot and a new feature called Claude Tag for team decision-making.
The Slack default is the consequential part
Salesforce in Claude will be used by sales teams that opt into it. Slack is used by everyone in the company.
On 20 August, Slack launched Slack Code, putting AI coding agents into project channels, and we covered it as a deliberately multi-vendor surface: the launch partners were Anthropic's Claude Code, Cognition's Devin, GitHub Copilot and Vercel's agent, available on every plan including free workspaces.
Six days later, Claude is the default model for Slack itself. The two moves are not in conflict. One governs which third-party coding agents a team can invite into a channel; the other picks the model for Slack's own features. The sequence still shows a platform claiming neutrality at one layer while choosing a side in the layer just below.
Defaults decide outcomes at this scale. A user who never opens a settings page is using Claude, and most users never do.
Salesforce inside Claude is the stranger direction
The plugin deserves more attention than it is getting, because of what it concedes.
Salesforce spent twenty-five years making its interface the place work happens. Shipping the CRM as a plugin inside another company's chat application concedes that the interface layer may be moving. It accepts that being reachable where users already are beats defending a destination.
That is a defensible strategic call and possibly the correct one. It is also the mistake incumbents historically make: refusing to distribute through a rival's interface until that interface no longer needs them.
The risk sits in the same place as the opportunity. A sales representative who prepares a meeting, reviews deal health and manages pipeline without opening Salesforce is a customer whose relationship with the product is now mediated by Anthropic. We have written before about the threat agentic AI poses to the seat-based SaaS model, and this is a large vendor choosing a side of that question early.
What a plugin with 37 skills actually is
The number invites a shrug, so it is worth asking what a skill actually is here, and why the count is not the point.
A skill in this sense is a defined task the model can invoke against live CRM data with a known shape of input and output: fetch the open opportunities on an account, summarise the last four interactions, flag deals whose close date has slipped twice. It is closer to an API endpoint with a description than to a prompt.
Their value is not in the count but in the fact that they were built jointly. A skill written by Salesforce knows the object model, the permission semantics and which fields are actually populated in real tenancies rather than in documentation. A third party writing against the public API guesses at all three.
Which also explains why more are promised for later this year. Thirty-seven covers sales; service, marketing and commerce are all larger surfaces, and none of them is in this release.
What it means for buyers in this region
For enterprises here the practical questions are procurement questions, and they are answerable before the beta.
Where does the inference run, and does that satisfy the data-residency terms already agreed for Salesforce data? A CRM record processed by a model is a disclosure to that model's operator, and the existing Salesforce agreement was not written with Anthropic in it.
Whether the Slack default can be changed at the tenant level, and whether that is a plan-gated control. An organisation with a policy on which model may see internal discussion needs to know if it can enforce it.
And what happens to the 37 skills if the partnership changes. Jointly built assets are exactly what does not survive a commercial falling-out, and a sales process rebuilt around them is a dependency on two companies staying friendly.
Why Anthropic gets more from this
Both sides gain, and the gains are of different quality.
Salesforce gets a frontier model without building one, at a moment when every enterprise software company is judged on its AI story. The capability is rentable rather than permanent, since the model underneath could be swapped.
Anthropic gets distribution into the daily working surface of a very large share of the world's knowledge workers, and default status is the hardest thing to buy in enterprise software. It also gets usage data of a kind that model quality compounds on, in a domain where its competitors are strong.
The timing is notable. This is the week Anthropic is telling IPO investors its addressable market exceeds US$30 trillion. A default position inside Slack is not thirty trillion dollars, and it is the kind of concrete distribution that makes a large abstract number easier to tell a story about.
View author profile → · Editorial policy
