At Dreamforce, Salesforce is showcasing AIforce, an interface layer that makes CRM data and workflows available in Slack, Claude, and other AI environments. Or is it simply Agentforce in new packaging?
Salesforce announced AIforce during Dreamforce. The company describes it as a layer that makes data, workflows, business logic, semantics, access rights, security, and governance from its own platform available in any AI interface. Those who wish to update a record, look up information, or start a workflow no longer need to be within Salesforce itself, but can do so from Slack or Claude. According to Salesforce, AI agents can search hundreds of records simultaneously and act autonomously based on that data.
AIforce is being launched with three so-called surfaces: Claudeforce, Slackforce, and Agentforce Coworker. The first two are in beta, while Agentforce Coworker is immediately available to customers. Beneath it lies the Headless Toolkit, an open architecture that exposes MCPs, APIs, plugins, skills, and development tools. Salesforce is not communicating pricing for the time being.
“AI is revolutionizing the way we work with technology,” says CEO Marc Benioff. “We are combining the intelligence of AI models with all the context that customers have built up in Salesforce.”
New layer or new name? #
The question arises as to what distinguishes AIforce from Agentforce, which Salesforce launched two years ago and has since received new version numbers at a rapid pace. Even then, it concerned agents that reason and act independently based on company data. In the 2024 architecture, channels such as Slack, WhatsApp, and SMS were already explicitly intended as places where these agents appear.
Last year at Dreamforce, Slack rebranded itself as an ‘agentic OS’, with direct access to Agentforce Sales, IT Service, and HR Service from within a conversation. This was followed in April of this year by Headless 360, which made the platform accessible via API, MCP, and CLI. The Headless Toolkit under AIforce builds directly upon that.
Claudeforce is also not a premiere this week. Salesforce already mentioned the integration with Anthropic during the quarterly results at the end of August. To kick off Dreamforce, Salesforce introduced another series of ready-to-use agents under the name Agentforce.
With the launch of AIforce, Agentforce is not disappearing. According to Salesforce, it remains the agent layer, with AIforce as the interface layer on top. In other words, AIforce answers the question of where the user works, not what the agent can do. This makes the announcement more of a repackaging of the gateway than of the technology.
Governance becomes the real work #
For IT departments, the center of gravity is shifting. Once actions originate from Slack or Claude, the core question is no longer whether an agent can do something, but which rights and which logging travel with it. Salesforce states that access rights, security, and governance move into the interface and points to Zero Data Retention. Salesforce is not communicating a price for AIforce for now. Gartner previously identified the licensing terms surrounding AI as ‘total chaos’ and explicitly cited Agentforce as an example of inconsistency within a single portfolio. Agentforce has also not yet fully managed to convince Salesforce’s customers and partners.
Koa: proprietary reasoning model on Nvidia Nemotron #
Alongside AIforce, Salesforce is also introducing Koa in collaboration with Nvidia, the first proprietary CRM reasoning model for Agentforce. The model is built on Nvidia Nemotron and trained on a proprietary dataset of synthetic scenarios, based on 27 years of experience with CRM systems.
Koa targets business actions such as updating sales opportunities, routing service cases, and planning follow-up actions. “The most valuable thing Salesforce has built is not our platform, but the accumulated knowledge of how companies actually work,” says Benioff. “With Koa, that knowledge becomes part of the model itself.”
In Salesforce’s own CRM Benchmark, Koa reportedly matches or exceeds leading models on CRM actions, with three times fewer errors. For now, this remains a claim by Salesforce on a self-designed benchmark, without external validation. Koa currently runs within Salesforce itself and is in several pilots; there is no date for general availability.