For decades, Salesforce's value has centred on the application employees open to view customer records, update opportunities and manage service cases. At Dreamforce 2026, however, the company is making the argument that the interface matters less than what sits beneath it.
AIforce brings Salesforce's data, workflows, business logic, permissions and governance into other AI interfaces, including Claude and Slack, letting employees query and act on Salesforce without necessarily opening Salesforce itself. Salesforce isn't abandoning the CRM so much as trying to make the infrastructure underneath the interface more important than the interface itself, extending the headless CRM strategy it has been developing.
This architecture now spans a layered stack: Data 360 for unified customer and business context, Customer 360 for business logic and permissions, Agentforce to carry out the work, and AIforce as the layer that lets people and agents reach all of it from wherever they are working. The transition is from Salesforce as an application employees use to Salesforce as a layer that other applications and agents can use, with MCP among the technologies connecting the pieces.
Salesforce doesn't need to own the model
Salesforce isn't insisting that its own model must sit at the centre of this architecture. Alongside opening the platform to external models such as Anthropic's Claude, it has built its own CRM-specific reasoning model, Koa, on NVIDIA Nemotron.
Salesforce says Koa was trained using synthetic scenarios based on nearly three decades of CRM knowledge and that it matches or exceeds leading models on its own CRM benchmark, while producing three times fewer errors on tested CRM actions. Those are Salesforce's own benchmark results. A research paper from Koa's developers found that the model improves substantially on its Nemotron base and performs strongly on enterprise agentic tasks, while still falling short of the strongest frontier models.
The bigger strategic point is that Salesforce does not necessarily need to control the model. It appears more focused on controlling the enterprise context, permissions, workflows and actions surrounding it, a layer that could become as important as the model itself as agents grow more capable.
Agents have to do the work
Enterprise AI is increasingly moving from pilots into live deployments, and Salesforce's clearest example is Siemens, which is using Agentforce agents to process inbound leads across 132 countries, with more than 2,500 unqualified leads arriving each month. The agents qualify and route those leads for 18,000 sellers, and Salesforce also describes agents connecting to Siemens' Teamcenter system and an agent-to-agent workflow for supplier onboarding.
That expansion is also central to Salesforce's growth ambitions. The company's investor-day materials describe an “Agentic Enterprise Opportunity” and reaffirm a FY2030 revenue target above $63bn.
Its adoption scenarios range from customers still on core licences through to those running Agentforce “wall to wall” across their internal and external operations, the top tier of which Salesforce says could see a 3–4x-plus ARR uplift. Those are Salesforce's own projections rather than independent forecasts, and they describe the most advanced tier of adoption rather than a typical outcome. Even so, the point is not simply to make Salesforce available in more places. Salesforce is trying to increase the amount of work, data and spending that runs through its platform.
What happens when the layer underneath fails?
During Dreamforce, Salesforce experienced a widespread service disruption, with customers across multiple regions reporting severe delays, intermittent errors and login failures. Salesforce attributed the problem to an internal login service buckling under load and said it had validated and rolled out a fix within a few hours, with services returning to normal the same day. There is no indication the outage was connected specifically to Agentforce or AIforce.
But the more business processes, agents and customer interactions depend on the platform underneath them, the more consequential the reliability of that platform becomes.
The interface is becoming the least interesting part
Salesforce's bet isn't that agents will simply replace dashboards. It's that the CRM can become the governed system of record and action behind whatever interface an employee or agent happens to be using.
This connects with a broader pattern already visible in customer experience, where coordinating agents across systems is becoming a more valuable capability than any single interface.
If Salesforce succeeds, employees will spend less time thinking about Salesforce as an application and more time relying on it as the system that gives agents permission to act, supplies the context they need and records what they do. That would make the CRM far less visible to users, but potentially much harder for enterprises to operate without.