Koa: Salesforce's AI Reasoning Model for Agentforce, Built on Nvidia Nemotron

On September 15, 2026, Salesforce and Nvidia announced Koa, described as the first CRM-dedicated reasoning model, built to let Agentforce agents reason through complex, multi-step business tasks. Unlike Claudeforce or AIforce, Koa isn't an interface layer: it's the model that decides how an Agentforce agent executes a task.

How Koa was trained

Koa is built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled after 27 years of Salesforce CRM deployments — using no customer data, per the official announcement. Training combines supervised fine-tuning (SFT) and reinforcement learning via GRPO (Group Relative Policy Optimization), using NVIDIA NeMo tooling. Training scenarios span more than 14 industries, including manufacturing, financial services, healthcare and travel.

• Base model: NVIDIA Nemotron 3 Super
• Training method: SFT + GRPO via NVIDIA NeMo
• Data: proprietary synthetic dataset, no customer data used
• Industry coverage of the training set: 14+ industries

Official CRM benchmark results

On Salesforce's "CRM Benchmark" — which tests real tasks such as updating opportunities, routing cases and scheduling follow-ups — Koa "matches or exceeds leading model performance on CRM actions with three times fewer errors," according to the official announcement. No methodological detail on the benchmark's exact composition (number of tasks, models compared by name) has been published yet; this figure should be read as a vendor-reported number, not an independent audit.

Availability

Koa is available in pilot to select customers in Agentforce now, with general availability announced for Winter 2026 in US regions. The model also powers Missionforce Operations (public sector / regulated organizations) and an internal Slack agent used by Salesforce employees. No pricing is disclosed in the official announcement.

What this means for your business

• If you already use Agentforce: Koa will gradually replace (or supplement) the underlying reasoning model with no interface change for your users — Winter 2026 general availability currently covers US regions only, so check on Europe before making any decision
• The "3x fewer errors" figure comes from Salesforce's proprietary benchmark: before any rollout, ask for a test against your own use cases (opportunity types, routing rules) rather than relying on the marketing figure alone
• If you're evaluating an AI CRM agent today: since the current pilot is limited to select customers, it's still early to base a platform choice on Koa alone — worth tracking rather than deciding on for now

Want to make your CRM workflow automation reliable before adding an AI agent layer on top? Our automation offering audits your existing business rules to identify what's actually worth handing to an agent.