Salesforce Koa is a specialised enterprise language model designed to complete multi-step CRM and tool-use tasks rather than compete only on general chat. Salesforce researchers say they built it by post-training NVIDIA’s open-weight Nemotron-3-Super-120B model with simulated workflows and outcome-based rewards.
- Koa specialises an open-weight base model for CRM and multi-turn tool use.
- Salesforce says training used public and synthetic data, not customer data.
- The practical enterprise test is successful, auditable task completion—not a benchmark win in isolation.
Salesforce Koa: what the model does
Salesforce, the enterprise software company behind its namesake CRM platform, disclosed Koa on September 15, 2026. The Salesforce-authored research paper says the model starts from Nemotron-3-Super-120B and is post-trained with group relative policy optimisation, a reinforcement-learning method that compares candidate responses and rewards better task outcomes.
The distinctive part is a simulation-to-reward pipeline. Workflow specifications are expanded into multi-turn tasks with different user personas, and the reward is tied to whether tool use resolves the task. For CRM scenarios, Salesforce says those specifications are written in Agent Script, its declarative language for defining Agentforce behaviour.
Why task completion changes the model test
Salesforce Koa matters because enterprise agents fail at the seams between steps: choosing a tool, preserving context, obeying permissions and knowing when the job is actually finished. A model trained against task-resolution rewards is explicitly aimed at those seams.
SiliconANGLE independently reported the launch and the model’s CRM focus. The paper reports gains on public tool-use, agentic-reasoning and enterprise CRM benchmarks, with the clearest improvement on multi-turn tool use. It also says Koa beats one strong proprietary baseline while remaining behind the strongest frontier models, an important boundary that prevents the release from being read as universal leadership.
| Disclosed design choice | Enterprise consequence |
|---|---|
| Open-weight Nemotron base | Specialisation builds on an existing model rather than starting from scratch |
| Simulated multi-turn tasks | Training targets sequences and tool decisions |
| Task-resolution reward | Success is tied to completing work, not just producing text |
The claim that no customer data was used applies to Koa’s training, according to the authors. It does not by itself describe the data path of every future deployment. Buyers still need to ask what live CRM data a deployed agent can read, which actions require approval, how failed actions are reversed and what appears in the audit log.
Koa also sits inside a broader Salesforce effort to control enterprise agents. Lapaas Voice’s report on the Salesforce Trusted Enterprise AI Harness explains the orchestration layer, while Claude for Financial Advisors shows how specialised workflows depend on governed connectors as much as model reasoning.
The next proof point is deployment evidence: named workloads, error rates, permission failures and the cost of completing a task. Until those arrive, Koa is best read as a clear training strategy for CRM agents, not a blanket claim that a specialised model can safely automate every customer process.
Frequently asked questions
What is Salesforce Koa?
It is a specialised enterprise language model built from NVIDIA Nemotron-3-Super-120B and post-trained for multi-turn tool use and CRM workflows.
Was Salesforce customer data used to train Koa?
The Salesforce-authored paper says training used public and synthetically generated data and did not use customer data.
Is Koa better than frontier AI models?
The authors report improvements over its base and one proprietary baseline on selected tests, but also say it remains behind the strongest frontier models.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
