Salesforce and Claude: SaaS value is leaving the screen

On 26 August 2026, on the evening of its quarterly results, Salesforce announced an expanded partnership with Anthropic called Claudeforce (United States · Salesforce). Claude becomes the default model in Slack, and the customer relationship management (CRM) software moves inside Claude as a plugin with 37 sales skills. Its chief executive, Marc Benioff, summed up the bet in one line: “Here, the UI is the AI”.

Software as a service (SaaS) is not disappearing. Its unit of value is changing: value is leaving the screen billed per seat and moving to three less visible layers — data, agent orchestration and governance. Yet most contracts are still written for the screen.

What is already settled

• Claude by default (United States · Salesforce): since 26 August 2026, Claude has been the default model for Slack, Slackbot, Agentforce Vibes and Agentforce Coworker, and one of Agentforce’s reasoning models. Inside Salesforce it runs through Amazon Bedrock, within the vendor’s trust boundary. The announcement mentions no exclusivity.
• Salesforce in Claude (United States · Anthropic): in beta since 15 September 2026 on all paid Claude plans. The plugin reads data according to each seller’s Salesforce permissions and, by default, asks for approval before every write. The CRM remains the system of record.
• Headless 360 (United States · Salesforce): since 15 April 2026, the whole platform has been exposed as APIs, MCP tools (the open protocol that connects an agent to tools) and command-line tools, with no browser required. Agents inherit the permissions already in place.
• The numbers (United States · Salesforce): for the quarter ended 31 July 2026, revenue of USD 11.3 billion (up 11%). Agentforce annual recurring revenue, which now includes Slackbot and Headless 360, exceeded USD 1.5 billion, up more than 240% year on year.
• The same move elsewhere (Germany · SAP): on 12 May 2026, SAP announced plans to make Claude a reasoning capability of its Joule assistant, while remaining open to other models.

What is still in progress

• The bill for agents (international · Salesforce Ben): Salesforce plans to count every successful call by a registered agent, over MCP or an API, as an “interaction” paid for in prepaid credits. As of 24 September 2026, the rate had not been published; 30 days’ notice has been promised. Each agent will also get its own identity instead of acting under the user’s.
• Two contracts instead of one (United States · VentureBeat): on 26 August 2026, a Salesforce executive acknowledged that Claude usage is, for now, contracted separately with Anthropic.
• Other business functions (United States · Salesforce): skills beyond sales are due to start arriving from late 2026.

The real issue: the seat is no longer the unit of value

On 30 January 2026, Anthropic launched business plugins for its Claude Cowork agent. Over five sessions, the S&P 500 software and services index lost nearly 13%, and 26% from its October 2025 peak (international · Reuters). The market was asking a simple question: if an agent does the work, who still pays for a seat?

Salesforce’s answer is to become the gateway agents must go through rather than the screen humans look at, and to sell three layers. Data: in the second quarter, the Data 360 platform ingested 104 trillion records, up 355% year on year. Orchestration: tools, APIs and skills that any agent can call. Governance: inherited permissions, approval before writing, a separate identity for each agent.

Yet the seat is still the heart of the bill. Even with Slackbot included, Agentforce accounts for about 3% of the USD 46 billion in revenue expected for the fiscal year. The two models will coexist, and that is where the risk lies: paying both for seats that agents make less useful and for consumption that nobody can yet forecast.

Three points to watch

1. A default is not a choice. “Default” does not mean “exclusive”: Salesforce says it is opening its data and workflows to any agent, through MCP, APIs and command-line tools, and presents its Agentforce Vibes development tool as multi-model. But the default model is the one your teams will use, and a change of version or price will be decided by the vendor unless the contract says otherwise.

2. Two meters, two contracts. Interactions will be billed by Salesforce, model usage by Anthropic: two sets of terms, two data processing agreements, two lists of sub-processors. If an agent writes wrong data into the CRM, the split of liability has to be written down before the incident.

3. The data changes sides. With Salesforce in Claude, CRM data is read inside Anthropic’s environment, under its terms. The beta is open to all paid plans, but no training on data by default is only stated for the Team and Enterprise plans. For a European organisation, processing location, retention and transfers outside the EU must be checked contract by contract.

What a board or a CIO should review

1. List the seats actually used by people, and those whose use is shifting to agents, before every renewal.
2. Demand the consumption rate card before signing: price per agent interaction, monthly cap, alerts and notice of price increases written into the contract.
3. Read the vendor’s and the model provider’s contracts together: service levels, data processing, sub-processors, hosting, a ban on training models on your data.
4. Negotiate the choice of model: the right to change the default model, notice of any change of version or provider.
5. Govern agents like privileged accounts: one identity per agent, least privilege, human approval of sensitive writes, retained logs.
6. Extend reversibility to the agent layer: export of data, but also of skills, instructions and configuration, in open formats.
7. Pilot before scaling up: a limited scope, a consumption budget, indicators measured before and after.

We used to buy screens; we now buy access for agents, and the contract has to follow.

The question to ask on Monday morning

To your CIO and your procurement team: “For our three main software vendors, do we know what we will pay if, tomorrow, agents do the work of our users?”