Salesforce and NVIDIA Announce Koa CRM Reasoning Model

Summary

Salesforce and NVIDIA announced Koa. This CRM-focused reasoning model was fine-tuned on synthetic enterprise data using the Nemotron 3 Super base. The model runs entirely on Salesforce's infrastructure; no customer data is used during training or inference.

Why it matters

This move signals a significant shift toward vertical, domain-specific reasoning models that address enterprise data sovereignty concerns by keeping weights and inference within a single vendor's stack. In CRM-heavy sectors, a model specifically engineered for these tasks can offer an advantage over general-purpose large language models (LLMs) in narrow business tasks such as opportunity updates, case routing, and follow-up planning. This also deepens the NVIDIA–Salesforce alliance, as Agentforce becomes increasingly dependent on a single hardware/AI partner.

Highlights

• Koa was fine-tuned via NeMo RL using synthetic scenarios across more than 14 industries; real customer data does not touch the training pipeline.
• According to Salesforce's own CRM Bench, Koa matches or surpasses leading models in CRM actions, with an error rate 3 times lower.
• Customer pilots (Formula 1, UChicago Medicine, Xero) begin in October 2026; general availability in the US is targeted for winter 2026.

Critical take

The '3 times fewer errors' figure in the headline was measured on Salesforce's own designed CRM Bench. There is no third-party validation; this is akin to grading one's own homework. Moreover, the model was trained on synthetic scenarios. How it will handle the complex, contradictory, and multi-workflow reality of a real enterprise CRM at scale remains entirely unproven.