Salesforce Koa Turns Workflow Knowledge Into a Model
Salesforce built Koa by training an open model on synthetic CRM workflows. The moat may be the workflow specification, not the base model.
Salesforce has introduced Koa, a CRM reasoning model for Agentforce built by post-training NVIDIA Nemotron 3 Super. Its training data uses synthetic CRM scenarios rather than customer data, covering workflows such as qualifying opportunities, routing cases and scheduling follow-ups.
Salesforce kept an open-model foundation, then trained it on workflow specifications, tool sequences and task outcomes. It says Koa runs inside its trust boundary and is in select customer pilots, with US general availability expected in winter 2026.
For PMs, the product decision is whether proprietary task structure justifies a specialist model. Before commissioning one, ask whether the team can describe its states, tools, permissions, completion criteria and representative failures. If not, better prompts, data and application controls may create more value than model customisation.
Treat the benchmark story as vendor-reported. On its own CRM benchmark, Salesforce says Koa produces one-third as many errors as leading models. The accompanying paper reports that Koa beats one proprietary baseline but remains below the strongest frontier models. Independent production evidence and pricing are not yet available.
