Salesforce and Nvidia unveiled Koa at Dreamforce on September 15, 2026, Salesforce's first in-house reasoning model and the clearest sign yet that enterprise buyers want something different from what the frontier labs are selling. Koa is built on Nvidia's open-weight Nemotron model, post-trained jointly by the two companies to handle sales, marketing, and customer-support tasks inside Salesforce's Agentforce platform. It ships as an alternative to routing reasoning-heavy prompts out to Claude or ChatGPT, and Salesforce is pitching it explicitly on cost, data control, and provenance.
Until Koa, any agent built on Agentforce that needed to reason across a multi-step task punted the work through Salesforce's AI gateway to a frontier model. That gateway still exists — customers can still route to Claude or ChatGPT — but Koa is now the default in-house option for the tasks Salesforce sees most often.
Jayesh Govindarajan, EVP of Salesforce AI, framed the model as the missing piece of a portfolio Salesforce has been assembling for years. Small task-specific models already handled narrow jobs like scheduling or classification; reasoning was the gap.
Key facts
- 01Salesforce and Nvidia unveiled Koa at Dreamforce on September 15, 2026 — Salesforce's first in-house reasoning model.
- 02Koa is built on Nvidia's open-weight Nemotron model and post-trained for sales, marketing, and customer-support tasks.
- 03The model was trained on synthetic data simulating customer service and sales scenarios, not on real Salesforce customer data.
- 04Koa slots into Agentforce as an alternative to Claude and ChatGPT, routed via Salesforce's AI gateway.
- 05Salesforce also announced ClaudeForce, a partnership with Anthropic that keeps customer data inside Salesforce infrastructure.
The core pitch to enterprise buyers is a stack of specifics that the closed frontier labs cannot match on their current terms. Koa is open-weight rather than closed. It was trained without ingesting any actual Salesforce customer data, which removes a class of leakage risk that has haunted enterprise deployments of ChatGPT and Claude. It uses fewer tokens than a frontier model on the same task, which cuts the per-inference bill. And it inherits Salesforce's existing data-residency and security controls by default.
The provenance argument is the one Salesforce is leaning on hardest, and it doubles as a shot at the Chinese open-weight ecosystem. Nemotron, in Govindarajan's framing, is the first sovereign American open base model good enough and clean enough to build on.
“Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance. We have no idea what Qwen trains on.”— Jayesh Govindarajan, EVP of Salesforce AI
The training method underscores the enterprise focus. Rather than fine-tune on real customer transcripts, Salesforce and Nvidia generated synthetic data by simulating a full customer service environment — including irate customers phoning support lines and sales reps working to close deals. The model learns the shape of the work without ever touching a real customer record.
Nvidia's contribution runs deeper than the base weights. Kari Ann Briski, Nvidia's VP of Generative AI Software for Enterprise, said the Nemotron architecture was designed for token-efficient inference, which is the metric that actually drives enterprise AI budgets once a deployment scales past pilot.
That framing — sovereign, fast, efficient — is a direct answer to the two objections that have slowed enterprise AI rollouts most: unpredictable inference costs and unclear data handling. Koa exists to make both problems go away for the specific work Salesforce customers do most.
Salesforce is not walking away from the frontier labs. The company simultaneously announced ClaudeForce, a partnership with Anthropic that lets enterprises use Claude as a conversational interface while their records stay in Salesforce's system and under Salesforce's security perimeter. The AI gateway routes accordingly — Koa for the tasks it does well and cheaply, Claude or ChatGPT for the ones that still need frontier capability.
The obvious open question is how Koa actually performs against Claude or ChatGPT on the specific enterprise tasks it targets. Salesforce has not published head-to-head benchmarks, and "cheaper and good enough" is a claim customers will test in production before they believe it. A model trained on synthetic customer service simulations will also face edge cases that no simulation anticipates, and the AI gateway model means Salesforce can quietly fall back to a frontier provider whenever Koa comes up short.
Koa is the shape of enterprise AI Salesforce has been signaling for two years, and it is exactly the product OpenAI and Anthropic should have been worried about. The frontier labs want enterprises pouring proprietary data, prompts, and feedback into closed models and paying per-token rates that scale with usage. Salesforce is offering the opposite deal: an open-weight model that never sees your data, costs less per call, and lives inside the compliance perimeter you already trust. If Koa performs, every reasoning task that used to route to Claude or ChatGPT inside an Agentforce deployment is a token the frontier labs no longer bill for — and the enterprise AI market stops looking like a winner-take-all race between three closed labs and starts looking like a distribution game the incumbents are well-positioned to win.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
