Salesforce Koa: What the CRM Reasoning Model Means for Your AI Stack

Salesforce and NVIDIA announced Koa on September 15, 2026 at Dreamforce: Salesforce's first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. According to Salesforce, Koa is purpose-built to help AI agents reason through complex, multistep workflows and use the right tools to complete them — and it is already in customer pilots, with U.S. general availability expected in winter 2026.

This is not another copilot feature. It is Salesforce shipping its own foundation-class model, with weights it controls, running inside its own infrastructure. For businesses already running multi-model AI strategies — Claude via Claudeforce, OpenAI models, or others — Koa turns model selection into a standing governance decision, not a one-time capability purchase.

What changed: what exactly did Salesforce announce?

Koa is Salesforce's first CRM reasoning model — a model trained to work through multistep enterprise tasks (like updating an opportunity, routing a case, or scheduling a follow-up) by reasoning step by step and invoking the right tools along the way, rather than simply generating text.

According to the Salesforce announcement:

• Foundation: Koa was built by post-training NVIDIA Nemotron 3 Super with a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of Salesforce CRM deployments.
• No customer data in training: Salesforce states the training corpus was built entirely from synthetic scenarios simulating real enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare, and travel.
• Training method: Salesforce applied Supervised Fine-Tuning (SFT) and reinforcement learning with Group Relative Policy Optimization (GRPO), using NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel.
• Control and trust boundary: Salesforce controls the model weights and performs post-training and inference entirely within its own trust boundary. Salesforce states that no customer data crosses that boundary during inference.
• Claimed performance: in Salesforce's CRM benchmark — a suite of real-world tasks like updating an opportunity or routing a case — Salesforce says Koa already matches or exceeds leading model performance on CRM actions with three times fewer errors. These are vendor-stated benchmark results; independent validation will come as pilots mature.

Koa is already running inside Salesforce — including an agent in Slack that helps employees find information and complete everyday tasks — and is now in customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.

A related piece of the announcement: Salesforce and NVIDIA are bringing Nemotron-based models and accelerated computing into Missionforce for government and regulated organizations, including deployment into private clouds and air-gapped networks. Post-trained NVIDIA models for Missionforce Operations reach select customers in October 2026.

Koa vs. Claudeforce: do not confuse these two

Readers coming from Dreamforce coverage will conflate Koa with Claudeforce, the Salesforce–Anthropic partnership announced three weeks earlier. They are architecturally different:

The practical takeaway: Salesforce now offers both a first-party model (Koa) and a deeply integrated third-party model (Claude via Claudeforce). As we covered in our analysis of the Claudeforce trust boundary, where a model runs shapes your compliance review. Who owns the weights shapes your roadmap risk. Koa raises both questions at once.

Who's affected

• Agentforce customers and evaluators. Koa will become a candidate reasoning engine for the agents you build. Pilot participants are already testing it on sales, service, and workflow tasks.
• Organizations with multi-model AI strategies. If your teams use Claude, OpenAI models, or others alongside Salesforce, you now have a Salesforce-hosted option to slot into that mix — which means model routing decisions, not just model adoption decisions.
• Regulated and government organizations. The Missionforce side of the announcement extends model, data, and deployment control into private clouds and air-gapped networks — relevant if public-cloud AI has been a blocker.
• CRM admins and operations teams. A reasoning model that takes actions (not just answers questions) raises the stakes on permission sets, sharing rules, and field-level security. Those were mostly configured for humans reading one record at a time.

Why this matters: model choice is now a governance decision

For the past two years, most businesses treated AI model selection as a capability question: which model writes the best email, summarizes the best call, scores the best lead? Koa signals the next phase. When your CRM vendor ships its own reasoning model — trained on synthetic versions of workflows like yours, hosted inside a trust boundary you have already assessed — the questions change:

• Provenance: Who trained the model, on what, and who controls the weights? (For Koa: Salesforce, synthetic data, Salesforce.)
• Data boundary: Does customer data cross any new perimeter during inference? (Salesforce states it does not for Koa.)
• Workflow fit: A model post-trained on CRM task sequences may outperform a general model on CRM actions — Salesforce's benchmark claims exactly that — while a general model may still win elsewhere.
• Portability and lock-in: A Salesforce-controlled model deepens the value of staying on-platform and the cost of leaving it.
• Accountability: When an agent takes the wrong action in your CRM, which vendor's model made the call — and under which agreement?

None of these are reasons to avoid Koa. They are reasons to evaluate it the way you would evaluate any system that takes actions in your CRM: deliberately, with owners and criteria named in advance.

What to do now: a practical checklist

1. Inventory your current AI models. List every model in use across sales, service, and marketing — including Claude via Claudeforce, any OpenAI usage, and embedded features like Einstein. You cannot route work between models you have not cataloged.
2. Map candidate workflows for Koa. Prioritize multistep CRM processes: lead qualification sequences, case triage and routing, opportunity hygiene, follow-up scheduling. These match the task types Salesforce says Koa was trained on.
3. Define evaluation criteria before GA. Agree now on accuracy, error tolerance, escalation paths, and audit requirements, so the winter 2026 GA window is a decision point rather than a scramble.
4. Review permissions for agent-executed actions. A reasoning model that updates records and routes cases at scale will exercise sharing rules and field-level security far harder than human users do. Fix over-permissive access first.
5. Update your AI governance register. Add Koa (and Claudeforce, if not already there) with owner, data boundary, vendor, and review date. Model choice is now a standing governance item, reviewed quarterly.
6. Watch for pricing. Salesforce has not announced Koa pricing. Budget planning should treat it as an open variable until GA terms are published.

If your team is evaluating how Koa fits alongside Salesforce, HubSpot, and your broader AI stack, Vantage Point can help you assess the right next step and build a practical implementation plan.

Timeline

FAQ

What is Salesforce Koa?

Koa is Salesforce's first CRM reasoning model for Agentforce, announced September 15, 2026 and built by post-training NVIDIA Nemotron 3 Super. According to Salesforce, it is designed to help AI agents reason through complex, multistep CRM workflows — like updating an opportunity or routing a case — and use the right tools to complete them.

Is Koa the same thing as Claudeforce?

No. Koa is Salesforce's own model, built on NVIDIA Nemotron with weights controlled by Salesforce. Claudeforce is the Salesforce–Anthropic partnership that makes Anthropic's Claude the default model in Agentforce Vibes, Agentforce Coworker, and Slack AI, running inside the Salesforce Trust Boundary via Amazon Bedrock. Salesforce now offers both, and many organizations will evaluate them side by side.

Was customer data used to train Koa?

Salesforce states that no customer data was used to train Koa. Its training corpus was built entirely from synthetic scenarios simulating enterprise workflows across more than 14 industries. Salesforce also states that inference runs entirely within its own trust boundary, so customer data does not cross that boundary when the model runs.

When will Koa be generally available, and what will it cost?

Koa is available to select pilot customers now, and Salesforce expects general availability in U.S. regions in winter 2026. Salesforce has not announced pricing; treat cost as an open variable in planning until GA terms are published.

How well does Koa perform?

Salesforce says that on its own CRM benchmark — a suite of real-world tasks like updating opportunities and routing cases — Koa matches or exceeds leading model performance on CRM actions with three times fewer errors. These are vendor-stated results; independent validation will emerge as pilots mature.

Should we wait for Koa or keep building on Claude and other models?

Do both, deliberately. Keep building workflows that deliver value today on your current models, but design them so the reasoning engine can be swapped or routed. Then evaluate Koa against your documented criteria when it reaches GA. Vantage Point helps organizations structure exactly this kind of multi-model evaluation.

Does Koa matter if we are not a Salesforce customer?

Indirectly. Koa signals that major platform vendors will increasingly ship their own specialized models alongside third-party options. Whatever your CRM — Salesforce or HubSpot — the governance questions about model provenance, data boundaries, and accountability now apply to your stack too.