Salesforce on September 15, 2026 announced Koa, a new CRM-focused reasoning model developed in cooperation with NVIDIA and intended to power its Agentforce platform. The koa model is described by Salesforce as a post-trained version of NVIDIA’s Nemotron family that was tuned on synthetic enterprise scenarios to help AI agents carry out complex, multi-step customer‑relationship tasks. The launch matters because it shows a major enterprise software vendor moving from general-purpose large language models toward vertically specialised models designed to operate inside corporate trust boundaries.
Salesforce published technical and product details on its website and said the koa model was created by post‑training NVIDIA Nemotron 3 Super on a synthetic dataset modelled on nearly three decades of CRM workflows. The company said it trained the model with a mix of supervised fine‑tuning and reinforcement learning techniques, and that no real customer records were used in the model’s training corpus. Salesforce said Koa is in pilot with select customers today and that broader availability is planned in the U.S. in winter 2026.
According to Salesforce, Koa was designed to help Agentforce agents reason through multi‑step processes such as generating and qualifying leads, routing and resolving service cases, and scheduling follow ups. Salesforce published benchmark results for its internal CRM benchmark showing Koa reduces certain CRM action errors versus leading generalist models; the company also listed pilot customers including 1‑800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. Salesforce and NVIDIA said the collaboration will extend Nemotron‑based models and accelerated computing into Missionforce, a variant aimed at government and regulated environments that require strict control over models and data.
Salesforce says Koa was created by starting with NVIDIA’s open‑weight Nemotron architecture and then post‑training it on synthetic scenarios that replicate enterprise workflows across multiple industries. In plain terms, post‑training means taking a high‑capacity base model and tuning it with examples and reinforcement methods so it develops stronger, task‑specific reasoning rather than only conversational fluency. The synthetic scenarios mirror the sequences of steps, decisions and tool calls an AI agent needs to perform common CRM work without exposing real customer data during training.
Why this development matters
Specialised models like Koa shift the enterprise AI conversation from raw general‑purpose intelligence to domain expertise and governance. For companies that must meet regulatory, privacy and security requirements, a model that runs inside a vendor‑controlled environment and was not trained on live customer records can be easier to audit and govern. The move also reflects a commercial logic: models tuned for predictable enterprise tasks can be cheaper to run and more reliable on specific actions than frontier generalist models that excel at broad creativity.
Salesforce launched Agentforce and related products over the last two years to embed autonomous agents across CRM workflows; Agentforce couples model reasoning with deterministic business logic and governance controls. Enterprises and cloud providers have been building what they call “compound AI” systems that combine several models, rule engines and connectors to keep agents aligned with business rules. Partnering with NVIDIA gives Salesforce access to Nemotron open models and accelerated compute tooling, while NVIDIA benefits from a reference enterprise customer and engineering collaboration that validates Nemotron in real‑world agent workloads.
Limits, unanswered questions and what happens next
Important uncertainties remain. Salesforce’s announcement gives a high‑level description of the training approach and internal benchmarks, but independent third‑party evaluations of Koa are not yet available. The company says no customer data was used to train Koa, but customers and auditors will likely press for evidence of how synthetic scenarios were generated and whether any downstream inference involves customer inputs crossing trust boundaries. Pricing, enterprise licensing terms, regional availability dates beyond ‘winter 2026’ and detailed performance comparisons against other enterprise models were not disclosed in the press material.
Operationally, customers will want to know how Agentforce will route sensitive requests, how Koa will be updated and governed over time, and what tools Salesforce will provide to monitor agent decisions. Salesforce said Missionforce will support deployments in private clouds and air‑gapped systems for regulated customers; practical timelines for those government‑grade deployments were left at a high level.
Sources
- Salesforce: Announcing Koa: Salesforce’s First CRM Reasoning Model, Built on NVIDIA Nemotron (press release)
- SiliconANGLE: Salesforce debuts Koa, a specialized model built to reason about CRM data
