- 01Salesforce unveiled AIforce at Dreamforce on September 15, 2026, as an interface layer for CRM data, workflows, permissions, security and governance.
- 02The launch includes Claudeforce, Slackforce and Agentforce Coworker; Claudeforce is in beta with 37 prebuilt sales skills.
- 03Headless 360 supplies the underlying MCP servers, APIs, plug-ins, reusable skills and developer tools.
- 04Koa is a separate Salesforce CRM reasoning model built on NVIDIA Nemotron 3 Super and available to select pilot customers in Agentforce.
- 05Salesforce says AIforce requests inherit existing permissions and business rules, while organizations still face practical questions around approvals, auditability, data quality and custom configurations.
Salesforce AIforce brings CRM data and workflows to Claude, Slack and AI agents. Here is what launched, what is in beta and where Koa fits.
Salesforce unveiled AIforce at Dreamforce on September 15, 2026, positioning the CRM platform as a governed data and workflow layer that can operate through Claude, Slack, Salesforce Lightning and other AI interfaces. The launch brings three products—Claudeforce, Slackforce and Agentforce Coworker—under one architecture, while the separate Koa model supplies Salesforce-specific reasoning for Agentforce.
The idea is straightforward: you can ask an AI interface to work with Salesforce data and processes without spending every minute inside the traditional CRM screens. Salesforce says the requests still inherit existing permissions and business rules, and that external model providers do not retain business data under its Zero Data Retention architecture.
Salesforce moves the CRM beyond its own screen
AIforce is Salesforce’s interface and orchestration layer. It exposes CRM data, workflows, business logic, permissions, security and governance to AI systems, rather than treating the Salesforce application as the only place where employees can interact with that information.
That makes AIforce different from Agentforce. Agentforce is Salesforce’s agent workforce layer; AIforce is the access layer that lets agents and third-party AI tools work with Salesforce capabilities through other interfaces. In practical terms, Agentforce supplies the workers while AIforce helps determine where and how those workers interact with the CRM.
Salesforce’s strategy also preserves its role as the system holding the underlying customer records and business rules. The visible interface might be Claude or Slack, but the data and governed workflows remain connected to Salesforce.
Three products, three ways into Salesforce
The initial AIforce lineup covers different entry points into the same CRM foundation:
| Component | Primary interface | Supported role | Current scope |
| Claudeforce | Claude | Sales prospecting, pipeline hygiene, meeting preparation, deal health and pipeline review | Beta with 37 prebuilt sales skills |
| Slackforce | Slack | Live Salesforce context, shared interactive surfaces, record updates and workflow triggers | Packaging and availability vary by customer agreement |
| Agentforce Coworker | Salesforce Lightning | AI teammate that reasons across accounts, activity and history and can invoke Agentforce agents | Salesforce says it reached 100,000 activations in its first 35 days |
Claudeforce is the clearest early example. Salesforce says it is available to customers in beta and includes 37 prebuilt sales skills. A separate Salesforce Development plug-in for Claude includes more than 40 skills for Salesforce development tasks.
Slackforce takes a different route: instead of opening Salesforce to inspect a record or build a report, a user can work from Slack and trigger Salesforce actions there. Agentforce Coworker stays inside Salesforce Lightning but is designed to reason across account information, activity and history before invoking existing Agentforce agents.
Salesforce also cites Fulton Bank as an early customer. The bank says it moved from zero to more than 20 Agentforce Coworker use cases in production, serving approximately 3,000 users.
MCP is the plumbing underneath
AIforce sits above Headless 360, Salesforce’s technical foundation for exposing platform capabilities outside the standard interface. Headless 360 uses Model Context Protocol servers, APIs, plug-ins, reusable skills and developer tools.
The distinction matters. MCP servers and APIs provide the connections and callable functions; AIforce packages those capabilities into interfaces intended for employees and AI agents. Salesforce executive Patrick Stokes described Headless 360 as the underlying exposure of Salesforce APIs and MCP servers, with AIforce positioned one layer above it.
That architecture is also why the launch is more than a new chatbot window. The interface can reach workflows and business logic, not just retrieve isolated CRM fields. Whether that feels like a meaningful shift or a more polished MCP connection will depend heavily on how an organization has customized Salesforce and how much write access it grants to agents.
Koa is a separate CRM reasoning bet
Koa is not the AIforce interface itself. It is a Salesforce-hosted CRM reasoning model for Agentforce, built on NVIDIA Nemotron 3 Super and trained with synthetic CRM scenarios, supervised fine-tuning and GRPO reinforcement learning.
Salesforce says the training scenarios cover CRM reasoning and tool use across more than 14 industries, without using customer data for training. The company also reports that Koa produced three times fewer errors than leading models on its own CRM benchmark.
For U.S. organizations, Koa is available to select pilot customers in Agentforce, with general availability expected in U.S. regions during Winter 2026. That is a separate rollout from Claudeforce’s customer beta.
Koa’s role is to handle multistep CRM tasks and tool use, while AIforce determines how Salesforce capabilities are exposed through interfaces such as Claude, Slack and Lightning. Put simply: AIforce is the doorway; Koa is one of the reasoning engines that may work behind it.
The real enterprise test is governance
Salesforce says AIforce requests continue to run through existing permissions and business rules. It also says its Zero Data Retention architecture means external model providers use business data to answer a request without retaining it. That policy describes Salesforce’s architecture; it does not by itself determine an organization’s broader compliance outcome.
The harder question is what happens when agents operate those controls at software speed. Organizations evaluating AIforce will need to examine permission drift, conflicting records, approval controls, audit trails, data quality, write access and compatibility with heavily customized Salesforce environments.
Those issues are especially important because a permission model built for a person moving through screens may behave differently when an agent can make repeated requests or trigger workflows rapidly. Existing governance remains part of the system—but it may become more consequential once the interface can act across Claude, Slack or other agent platforms.
The launch therefore changes Salesforce’s product boundary without removing the platform’s central responsibilities. Employees may no longer need to work primarily through Salesforce’s own screens, but the CRM still supplies the records, workflows, permissions and business logic on which those new interfaces depend.
Read this article in another available language.
ItalianoPortuguêsNederlandsDeutschPolskiFrançaisEspañol
Huawei’s reported AI-chip push takes aim at Nvidia with massive clusters
Huawei’s reported AI-chip push combines Ascend accelerators, CPUs, connectivity and Atlas clusters as it challenges Nvid…
Senate inquiry puts OpenAI’s Hugging Face incident under scrutiny
The U.S. Senate is examining OpenAI’s Hugging Face incident, including a reported May warning signal, agent coordination…
Anthropic alleges Chinese AI labs extracted Claude capabilities at industrial scale
Anthropic alleges Chinese AI labs extracted Claude capabilities at scale. Here is what the campaigns involved, what U.S.…
