Salesforce introduces Koa, a CRM reasoning model for Agentforce

Salesforce introduced Koa on 15 September 2026 as its first CRM reasoning model for Agentforce, the agent platform it is building around sales, service and workplace automation. The model is not presented as a general frontier system. It is a domain model derived from NVIDIA Nemotron 3 Super, an open-weight model, then post-trained by Salesforce for customer relationship management tasks. That origin matters because the announcement is less about creating a new base model than about controlling a reasoning layer for business workflows where the system must decide what to update, where to route a request and when to ask a human.

A Nemotron model adapted to CRM

Salesforce says it post-trained Koa on synthetic scenarios grounded in 27 years of CRM experience and spanning more than 14 industries. The company also says no customer data was used to train the model. In its description, the post-training combines supervised fine-tuning and GRPO reinforcement learning, using NVIDIA NeMo RL, NeMo Gym and NeMo AutoModel. The technical point is governance as much as performance: Salesforce says it controls the weights and runs both post-training and inference inside its own infrastructure, rather than sending the entire CRM reasoning layer to an outside model provider.

• Koa is Salesforce's first CRM reasoning model for Agentforce.
• The starting point is NVIDIA Nemotron 3 Super, described as an open-weight model.
• Salesforce says the post-training used synthetic CRM scenarios, not customer data.
• The announced training methods include SFT and GRPO reinforcement learning with NeMo tools.
• General availability is expected in winter 2026 in United States regions.

What Salesforce says it tested

The company points to an internal CRM Bench that tests tasks such as updating an opportunity, routing a case and planning a follow-up. Salesforce says Koa performs at least as well as large models while making three times fewer errors. That claim should be read carefully because the benchmark is internal and has not been independently validated. It is still useful as a statement of target behavior: Salesforce is measuring whether an agent can complete CRM steps with fewer mistaken actions, not whether it can answer trivia or write general prose.

The first live use case is also narrow. Salesforce says Koa already powers an employee agent in Slack, where the model can reason over workplace requests inside a controlled product surface. Named pilots include Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine and Xero. Salesforce says broader availability is planned for winter 2026 in United States regions. The company also says Missionforce will receive Nemotron models and accelerated computing for private clouds and isolated networks, a signal that it sees sovereign or restricted environments as part of the same architecture.

Why Salesforce wants its own reasoning layer

TechCrunch interprets the move as a way for Salesforce to reduce dependence on frontier models for multi-step CRM tasks and to cut token costs, without abandoning Claude or ChatGPT. That distinction is important. Salesforce is not saying every enterprise task should move to Koa, and TechCrunch does not describe a clean break with outside labs. The argument is narrower: for repetitive CRM work, a smaller specialized model controlled by the application vendor may be cheaper, easier to audit and easier to tune than repeatedly calling a frontier model for every step.

The buyer's test should therefore be operational, not rhetorical. Teams should compare Koa with existing model routes on real records, with permission boundaries active, partial data, stale fields, ambiguous cases and failed tool calls. They should measure accuracy on updates, recovery after an error, total cost per completed workflow, audit trails, and how reliably the agent escalates to a human when it lacks authority or confidence. None of that is a Salesforce promise; it is the work required before a reasoning model is allowed to act inside a CRM.

What changes for English-speaking companies

For English-speaking companies, Koa gives Agentforce customers a new option to test CRM reasoning closer to the system of record. The practical question is not whether Koa beats a famous chatbot in general intelligence. It is whether an agent can update opportunities, route cases and plan follow-ups with fewer errors, a predictable cost curve and a clear audit trail. If pilots confirm that inside real Salesforce deployments, Koa could shift some CRM automation away from broad model selection and toward governed application-level reasoning.