Announced at Dreamforce 2026, Salesforce Koa is Salesforce’s first purpose-built CRM reasoning model for Agentforce. Built in collaboration with NVIDIA on the Nemotron foundation, Koa is designed for the high-volume, multi-step workflows that make up most enterprise CRM work—updating opportunities, routing cases, scheduling follow-ups, and similar actions—rather than competing as a general-purpose frontier model.
The strategic value is not just capability. It is unit economics, data trust, and intelligent model routing.
What Is Salesforce Koa?
Koa is a specialized reasoning model that helps Agentforce agents plan and execute complex, multi-step CRM tasks. Salesforce post-trained NVIDIA’s Nemotron open models using supervised fine-tuning and reinforcement learning on NVIDIA NeMo tooling. Training used a proprietary synthetic dataset modeled on nearly 27 years of Salesforce CRM knowledge across more than 14 industries. No customer data was used in training.
Key points:
- The base Nemotron model has open weights, but Salesforce controls Koa’s weights and runs the model entirely inside its own infrastructure.
- Customer data never crosses Salesforce’s trust boundary during training or inference.
- Koa is available as a managed LLM in the Data Cloud generative models catalogue, selectable org-wide in Agentforce Setup, and usable at the agent or sub-agent level in Agentforce Builder.
Performance Claims and the CRM Benchmark
Salesforce reports that on its internal CRM Benchmark (real-world style tasks such as opportunity updates, case routing, and follow-up scheduling), Koa matches or exceeds leading general models on CRM actions while producing roughly three times fewer errors. Additional reported lifts include approximately 11% higher precision in calling the correct action, 2.1× greater reliability in recalling customer context, and 15% better context retention in longer conversations.
These figures come from Salesforce’s own benchmark. Independent head-to-head results against models such as Claude or GPT on identical agentic CRM tasks have not been published. Treat the numbers as a strong signal for piloting rather than a final proof. Also check out our guide on Claudeforce Explained: Salesforce and Anthropic Bring Claude Into the CRM.
Why the AI Gateway Matters More Than Any Single Model
Koa does not replace frontier models. It complements them. Inside Agentforce, an AI gateway automatically routes work: routine, well-scoped, high-volume CRM actions go to the cheaper, CRM-tuned Koa; tasks that require deeper judgment go to frontier models.
This architecture is the real lever. At Salesforce scale, routing the majority of agent actions to a specialized, lower-cost model while reserving expensive frontier tokens for harder problems can materially improve the cost of running digital workers. NVIDIA has described the combination as delivering sovereign AI control, fast time-to-first-token, and efficient reasoning.
Data Trust and Enterprise Controls
Because Koa is hosted inside Salesforce’s trust boundary and was trained only on synthetic scenarios, regulated industries and risk-sensitive teams gain a clearer data-protection story than models that require data to leave the platform. Salesforce also notes enterprise controls including instruction adherence, temperature-0 serving for consistency, and an additional serving harness for trust and safety.
Availability and Early Customers
Koa is in pilot with select customers (examples cited include UChicago Medicine, 1-800Accountant, Baxter Credit Union, Engine, Formula 1, and Xero). General availability is expected in U.S. regions in Winter 2026, with broader rollout to follow. Salesforce itself is using Koa internally across help, employee, event, and web agents.
What Mid-Market and Enterprise Teams Should Do Next
You will rarely choose Koa manually—the gateway decides. Your focus should be outcomes:
- Identify high-volume, repetitive CRM workflows that currently consume frontier-model tokens.
- When Koa reaches GA in your region, pilot those workflows and measure first-pass success rate and escalation frequency.
- Track true cost and quality on your data and processes rather than relying solely on vendor benchmarks.
- Keep frontier models available for the judgment-heavy cases the gateway routes upward.
A purpose-built CRM model for routine work plus intelligent routing is a pragmatic architecture. Most agent activity in a CRM is repetitive and well-scoped; a specialized, cheaper, privately hosted model is a logical fit for that majority, provided escalation rates stay low.
Frequently Asked Questions
What is Salesforce Koa?
Salesforce’s first CRM-specific reasoning model for Agentforce, built on NVIDIA Nemotron and trained on synthetic data reflecting 27 years of CRM workflows across 14+ industries.
Does Salesforce train Koa on customer data?
No. Training used only synthetic scenarios. No customer records were involved.
Is Koa better than Claude or GPT?
On Salesforce’s CRM Benchmark it claims equal or better performance on CRM actions with fewer errors. It is not positioned as a general-purpose replacement for frontier models.
When is Koa generally available?
Pilot now for select customers; expected GA in U.S. regions in Winter 2026.
Does Koa replace other models in Agentforce?
No. An AI gateway routes routine work to Koa and more complex judgment work to frontier models.
Koa’s real contribution is not another general intelligence claim. It is a specialized, cost-efficient, trust-boundary-respecting model for the everyday multi-step work that dominates CRM agents—paired with routing that makes the economics of large-scale Agentforce deployments more sustainable. For teams already investing in Agentforce, the practical next step is a controlled pilot that measures first-pass accuracy and escalation rates on real workloads. Working with an experienced Salesforce Agentforce partner can help design that pilot, interpret the results, and ensure the gateway routing and model mix deliver measurable ROI on your specific processes.
