Salesforce’s AI bet combines a specialist model with control of the workflow

Salesforce’s Dreamforce announcements, flagged by The Information, combine a specialised model with a way to make its existing platform available inside other interfaces.

Koa is built by post-training Nvidia’s Nemotron 3 Super. Salesforce says the training scenarios are synthetic, modelled on CRM workflows across more than 14 industries; it explicitly says customer data was not used to train the model. The company controls the weights and hosts inference within its own infrastructure.

Its pitch is narrower than general intelligence: select the right tools and carry out business processes accurately. Salesforce reports fewer errors on its own CRM benchmark and has begun customer pilots. That is a vendor evaluation and an early deployment stage, not independent evidence of results across every customer’s environment.

AIforce addresses a different question: where does the work happen? The proposed interface layer exposes Salesforce’s records, workflows and business logic through places such as Claude and Slack. Users can ask for a customised live view or trigger an action without navigating the traditional application.

Salesforce says those requests retain existing permissions and business rules, with actions routed back through its platform. It also advertises zero data retention by the model provider for the business data used in a request.

The strategic implication is that an application vendor can lose its position as the screen people open while retaining its role as the system that authorises and records their work. A specialised model is one part of that defence. The accumulated workflows, permissions and operational history are another.

Whether this improves the experience will depend on execution: agents have to understand those rules, complete the intended task and leave a record that the business can inspect.