Claudeforce: what Salesforce customers can use today, and what is still a pilot

On 26 August Salesforce and Anthropic announced Claudeforce, a single name for three things the press release puts on the same footing. Claude becomes a reasoning model inside Agentforce, Salesforce's agent platform. Claude becomes Slack's default model. Salesforce moves inside Claude through a plugin with 37 skills built for sales people.

For a company running Salesforce those three things have different availability, different contracts and different guarantees about where the data goes. The press release does not help you separate them, so I will, one line at a time.

What you can use now and what you cannot

Claude inside Agentforce is there. The release describes it as the engine of the Atlas Reasoning Engine, the default model of Agentforce Vibes and Agentforce Coworker, and a selectable model in Agent Builder. A general availability date with the list of covered editions appears neither in the release nor on the Claudeforce page. The last dated statement comes from Anthropic, which on 14 October 2025 wrote about Agentforce with Claude being available to select customers.

The cost runs through the Agentforce price list, which as consulted on 10 September makes no distinction by model. Salesforce's pricing page shows $500 per 100,000 Flex Credits, 20 credits for each standard agent action and 30 for a voice action. For customer-facing agents there is the alternative rate of $2 per conversation, which however cannot be bought under the PayGo and Pre-Commit consumption models. Then there are the add-on licences at $125 per user per month, $150 for the industry versions. The free Foundations package includes Agentforce Builder, Prompt Builder, Agent Script, Agentforce Coworker and Agentforce Vibes.

On Slack the release says Claude is the default model and powers Slackbot, without stating plans or usage limits. The Claude app you call in a channel is a different thing. According to Slack's guide it works on all Slack plans but requires a paid Claude subscription (Pro, Max, Team or Enterprise), reads the 20 most recent channel messages as context, and since 3 August has been replaced by Claude Tag.

For developers, the Winter '27 announcement of 31 August lists the Salesforce Development plugin as already available in the Claude Code marketplace. The Winter '27 release as a whole becomes generally available on 12 October.

Then there is the part everyone has been talking about. The Salesforce plugin inside Claude is available to selected pilot customers and the open beta is expected in September, but as of today no official announcement declares it open. The Claudeforce page marks sales as beta and every other area — from Service to Marketing, Commerce, Field Service, Tableau, MuleSoft, Data 360 and Industries — as coming soon. There are no published prices. The release itself warns that pricing and packaging may change and that availability may vary by region.

The optimistic reading

Salesforce has made its CRM readable and writable by a model without asking anyone to build an integration. The plugin talks to Salesforce through MCP, the protocol models use to connect to external tools. Permissions stay the ones the company has already configured. Patrick Stokes, a Salesforce executive, said it to VentureBeat without any hedging. If you do not have permission to see a record, neither does the MCP server, so you can neither read nor write it.

For a sales team that updates deals on Friday afternoon, finding the pipeline review already done on Monday is a benefit that needs no training. On the Agentforce side the advantage is already purchasable. The release states that through Amazon Bedrock Claude is available inside the Salesforce trust boundary and explicitly mentions regulated industries. For a bank or a healthcare organisation already running Salesforce, that removes one vendor from the assessment list. The Claudeforce page also announces, for the autumn, Enterprise Frontier Safeguards, which will let customers keep activity data on their own infrastructure and under their own encryption keys.

Salesforce also brings a number, in the release of 26 August: 8.1 million hours of annualised productivity gains from Slackbot internally, more than doubled in a quarter. It is a vendor's figure about itself and nobody outside Salesforce can check it.

The critical reading

The sentence that convinced regulated industries was already in a Salesforce article of 2 February 2026.

Anthropic is the first large language model provider whose models are fully contained within the Salesforce trust boundary.

In October 2025 Anthropic explained it as all Claude traffic being contained within Salesforce's virtual private cloud through Amazon Bedrock, speaking about Agentforce. The August release ties it to Bedrock again. The plugin, however, works in the opposite direction, because it is Salesforce data that travels into Claude. For this direction the Claudeforce page promises inherited permissions and zero data retention on Sonnet, Opus and Haiku, but does not say on which infrastructure the model runs. And Stokes told VentureBeat that Salesforce and Anthropic bill separately and that for now you cannot buy this on one piece of paper.

Two contracts mean two data processing agreements to hand to whoever handles that in your company. They also mean two invoices to reconcile, for consumption nobody has quoted yet. Amazon Bedrock does have inference profiles tied to a geography, the European Union among them, and for those the documentation states that the list of destination regions never changes. None of the public documents on Claudeforce says whether and how they are used for an Italian organisation.

Then there are the use cases. The customer that appears both in Anthropic's October 2025 announcement and in Salesforce's February 2026 article is RBC Wealth Management. In 2025 Rohit Gupta, Head of Digital Advisor Platforms, said that Claude on Bedrock with Agentforce had saved advisors significant time in meeting preparation. In February Bill Wilson, Director of Applications, spoke of advisors freed from administrative maintenance. Minutes, hours or percentages appear in neither statement.

Where I stand

I stand with the critical reading. The part of Claudeforce a company can sign today with written guarantees about its data is Claude inside Agentforce, which is the part getting the least coverage. The plugin inside Claude may well work perfectly, but today it has no price, no declared place of processing, and it runs through a second contract whose terms nobody has published. It should not go into any Italian company's 2026 budget.

Anyone who already has Agentforce, on the other hand, has a calculation to do with real numbers. At the current list price a standard action costs 10 cents, so the $2 of the per-conversation rate is equivalent to twenty actions. In the agents I have built on other platforms the conversation that wrecks the maths is not the average one but the one where the downstream system returns an error and the agent retries three or four times. On the Salesforce pricing page I could not find how those retries are counted. A demo with clean data does not show them.

By 30 September, that is by the end of the month in which the plugin beta is expected, take one single customer-facing process, run it in Agent Builder with Claude across a hundred real conversations and count the billed actions one by one, errors included. If the average is above twenty non-voice actions per conversation, ask for the per-conversation rate. If it stays below ten, credits cost less than half. With voice actions, which consume thirty credits, the break-even comes sooner, at just over thirteen.