An AI system that answers well still needs to understand the customer, respect permissions and take the right action to create value inside a company. Salesforce's Dreamforce 2026 announcements focus on that gap between model intelligence and business operations.
The event, held September 15–17 in San Francisco, lists Anthropic's Dario Amodei and NVIDIA's Jensen Huang in the main keynote program, alongside a separate conversation between Marc Benioff and OpenAI's Sam Altman. Their presence shows the scope of the agenda. The product announcements explain its direction: making CRM a source of context and execution for work with AI.
AIforce brings CRM to where work happens
Announced on September 15, AIforce is an interface layer that brings Salesforce data, workflows and business logic into environments such as Claude and Slack. Its proposed capabilities include retrieving information, updating records and triggering workflows through those interfaces, using existing permissions and rules.
For sales leaders, the potential value is a shorter path from identifying a problem to acting on it. A pipeline analysis improves operations when it reaches the owner, supports a decision and leads to a recorded action. That continuity matters more than adding another chat window to the team's working day.
For CIOs and architects, the implication is significant: interfaces can change, while authorization and execution must remain tied to business systems. This is where a compelling demonstration starts to become a usable architecture.
Claudeforce connects reasoning to sales context
The expanded Anthropic partnership, announced in August as Claudeforce, makes that integration more concrete. Salesforce in Claude includes 37 sales skills, covering work such as meeting preparation and opportunity analysis. Claude is also available as a reasoning model in Agentforce and is the default for Agentforce Vibes and Coworker.
The technical challenge is to connect model reasoning to actions that make sense in the business. Preparing for a sales meeting requires account history, opportunity status and next steps. Updating that opportunity also requires permission to change the record. These are different responsibilities that must work together.
This changes how a pilot should be assessed. Beyond answer quality, teams need to establish whether they can complete work with less effort while keeping changes traceable.
Koa adds specialization to the strategy
On September 15, Salesforce introduced Koa, a CRM reasoning model based on NVIDIA Nemotron 3 Super. It is post-trained on synthetic business workflow scenarios, without customer data. Salesforce controls the weights and runs inference on its own infrastructure. Koa is in selected Agentforce customer pilots.
This approach brings specialization and model control into the same discussion. Architecture teams should test whether it improves tasks such as routing a case or updating an opportunity, while accounting for latency, cost and maintenance effort. A specialized model needs to justify its adoption in the company's actual workflow.
Start the executive decision with a process
The Benioff–Altman conversation is part of a program connecting enterprise platforms and model developers. The integration and product announcements provide a concrete starting point for assessing what that connection delivers.
The strategic reading is that Salesforce is competing for the place where intelligence meets business context and becomes action. That matters to executives who need a return on investment and engineers who need to deliver reliable operations.
A company's next step is to choose a bounded process and measure the outcome: completion time, errors, human intervention and cost per completed task. Decisions to expand adoption can then rest on operational evidence, with clear owners and limits on agent actions.