Salesforce Koa: First CRM Reasoning Model on NVIDIA Nemotron
Salesforce unveiled Koa on September 15, 2026 at Dreamforce, its annual customer conference in San Francisco. Koa is Salesforce’s first purpose-built CRM reasoning model for the Agentforce platform, developed through deep technical collaboration with NVIDIA. The announcement lands at a moment when enterprise buyers are increasingly questioning whether a specialized domain model can outperform a general frontier model on the specific tasks that matter most to their business.
For CRM workflows, the early data suggests the answer is yes.
What Koa Does
Koa is not a general-purpose AI assistant. It is optimized for the reasoning tasks that Agentforce agents perform across the customer lifecycle: updating sales opportunities, routing support cases, qualifying inbound leads, scheduling follow-ups, and executing multi-system workflows. On Salesforce’s internal CRM Bench benchmark, a suite built from real-world business tasks, Koa matches or exceeds the performance of leading general-purpose models with three times fewer errors.
That reliability gap matters because it speaks directly to the failure mode that has blocked enterprise AI agent deployments. When an agent routes a support case incorrectly or calls the wrong action in a multi-step workflow, the cost is not just a failed transaction. It is lost customer trust, manual remediation time, and reduced confidence in agentic AI adoption across the organization.
Koa targets that reliability problem from inside the model itself, rather than patching around it with external guardrails.
How Koa Was Built
The foundation is NVIDIA Nemotron 3 Super, a 120-billion-parameter open-weight foundation model. Salesforce selected Nemotron specifically because it is a sovereign American pre-trained model with clear data provenance. Salesforce EVP of AI Jayesh Govindarajan made the criteria explicit: “We have no idea what Qwen trains on,” he told TechCrunch, explaining why open weights with known origins mattered more than raw benchmark scores.
From that base, Salesforce applied two post-training stages using NVIDIA’s NeMo tooling: supervised fine-tuning (SFT) followed by reinforcement learning with Group Relative Policy Optimization (GRPO). The training corpus contains no actual customer data at any stage. Salesforce constructed it entirely from synthetic scenarios derived from nearly 27 years of internal CRM deployment knowledge: the processes, workflows, and operational policies documented across real enterprise deployments, translated into simulation-ready training data.
The technical architecture is described in detail in a companion arxiv paper submitted on September 14, 2026.
The Simulation-to-Reward Pipeline
The core innovation is a specification-driven RL pipeline. Enterprise training tasks were written in Agent Script, Salesforce’s declarative language for configuring Agentforce agents. A specification captures an agent’s routing logic, sub-agents, typed actions, tool scope, and workflow instructions. A simulation pipeline then expands each specification into scenario- and persona-conditioned multi-turn tasks, with grounded task-resolution rewards driving GRPO training.
The result is that the same specifications used to configure a production Agentforce agent also structure its training tasks. Agent authoring and model post-training share the same declarative layer, which means improvements to Agentforce specifications feed directly back into model training.
CRM Benchmark Performance
Salesforce benchmarked Koa against its CRM Bench, a suite of real-world tasks including opportunity updates, case routing, and follow-up scheduling. The results across key dimensions:
| Dimension | Koa vs. General Frontier Models |
|---|---|
| CRM action accuracy | 11% more precise: fewer wrong tool calls |
| Response reliability | 2.1x greater: higher relevance, faster answers |
| Context retention | 15% better across long conversations |
| Error rate on CRM tasks | 3x fewer errors overall |
The arxiv paper reports that across public tool-use, agentic-reasoning, and enterprise CRM benchmarks, Koa improves over the open-weight Nemotron base with the clearest gains on multi-turn tool use. It surpasses GPT-4.1 as a strong proprietary baseline while remaining below the strongest frontier models on general benchmarks. The research team describes this as a demonstration that specification-driven RL is a practical path to specializing open-weight foundation models for enterprise agentic tasks without trading away general-purpose capability.
Trust Boundary and Data Privacy
Salesforce controls the Koa model weights and performs all post-training and inference entirely within its own infrastructure. No customer data crosses the trust boundary during inference. Koa runs at temperature 0 by default, which ensures consistent and repeatable responses rather than open-ended generation. A dedicated serving harness adds trust and safety controls above the model layer.
This architecture directly addresses the most common objection from enterprise buyers evaluating agentic AI deployment: data cannot leave the controlled environment. With Koa, it does not need to.
Missionforce: AI for Regulated Industries
Alongside Koa, Salesforce and NVIDIA are extending the Nemotron collaboration to Missionforce, Salesforce’s secure cloud platform for high-security operations. Post-trained NVIDIA models will power Missionforce Operations, a product that automates government and regulated-industry workflows including procurement, supplier management, and logistics.
These models can run on private clouds and fully air-gapped networks, which is a hard requirement for organizations whose systems cannot reach public infrastructure. Government agencies, defense contractors, and regulated financial services firms are the primary addressable segment. Post-trained Nemotron models for Missionforce Operations will be available to select customers in October 2026. Current Missionforce Operations is generally available in US regions today.
What Changes for Enterprise Buyers
Before Koa, when an Agentforce agent needed to reason through a multi-step CRM task, that reasoning was routed to a frontier model such as Claude or ChatGPT through Agentforce’s AI gateway. That routing carried real costs: per-token spend at frontier pricing, added latency, and data leaving the Salesforce environment for external inference.
Koa is designed to capture that reasoning workload within Salesforce’s infrastructure for CRM-specific tasks. Marc Benioff, Salesforce Chair and CEO, framed the strategic intent directly: “The most valuable thing Salesforce has built isn’t our platform. It’s the accumulated knowledge of how enterprise business actually works. With Koa, that knowledge is put inside the model itself.”
For enterprise AI teams, the practical change is a new model selection option inside Agentforce. Koa becomes the fourth model provider available in the generative AI models catalogue, selectable org-wide in Agentforce setup or at the individual agent and sub-agent level. The deployment pattern mirrors how teams already configure model routing in Agentforce today.
This development connects to broader platform moves Salesforce has made in recent weeks. The Trusted Enterprise AI Harness and the named Agentforce agents announced earlier this week establish the governance layer and role-specific agent products. Koa provides the domain-specialized reasoning model underneath them. Together they represent a more complete AI stack that enterprise buyers can run within a single vendor’s trust boundary.
The move also connects to the ClaudeForce partnership with Anthropic announced in August. ClaudeForce lets enterprises use Claude as their AI interface while their data stays in Salesforce. Koa inverts that: reasoning happens inside Salesforce, keeping cost and control inside the platform. Both patterns point in the same direction. Enterprise software companies are bringing the reasoning layer inside their own stacks rather than permanently delegating it to external frontier model providers.
TechCrunch noted the competitive implications: a model that is cheaper (because it burns fewer tokens on CRM tasks), more accurate on the tasks that matter, and housed entirely within the customer’s trust boundary is “everything the AI labs should fear.” Not because it replaces general frontier models for all work, but because it displaces them for the specific high-volume, high-stakes tasks that drive the largest agentic AI workloads in enterprise CRM.
Availability
Koa is currently in expanded pilot through Agentforce with customers including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. General availability in US regions is expected in Winter 2026. An open beta is planned shortly after GA launch.
Access: select Koa as a managed LLM in the Data Cloud generative models catalogue, enable it org-wide in Agentforce setup, or apply it at the agent or sub-agent level in Agentforce builder.
For enterprise AI and GTM teams evaluating Agentforce, the Koa pilot program is the relevant next step. Salesforce has stated it expects to expand availability after the initial US GA window.