While everyone was arguing about AI safety panels at Dreamforce this week, Salesforce quietly shipped something that matters a lot to anyone who works with customers for a living. On Tuesday, September 16, the company announced Koa — its first purpose-built CRM reasoning model, co-engineered with NVIDIA and designed to run inside Salesforce's own trust boundary. If your office runs on CRMs (and in India, most sales and support teams do), this one's for you.
01.Why build your own reasoning model at all?
The team behind Koa — Jayesh Govindarajan, EVP of Software Engineering, and Silvio Savarese, Chief Scientist at Salesforce AI Research — made a blunt observation in their announcement: a general model reasons about a refund policy the same way it reasons about a physics problem. Same gears, same process, no context. That works fine for homework. It's wasteful for an enterprise pipeline where an agent has to decide whether a ticket escalates, a lead gets called, or an order gets held.
So instead of renting general intelligence for every micro-decision, Salesforce took NVIDIA's open Nemotron 3 Super model and post-trained it on a proprietary synthetic dataset modelled on nearly three decades of CRM deployments — how real businesses actually run their processes, workflows and operational policies. We covered the bigger industry drift in our post on Google's Gemini 3.8 Live launch, and Koa is the same story wearing a suit.
02.What Koa actually does for AI agents
Koa isn't a chatbot you talk to. It's the reasoning engine underneath Agentforce, Salesforce's agent platform. When an agent needs to update an opportunity, route a case, or schedule a follow-up, Koa is what figures out the multi-step logic and picks the right tool for the job. In Salesforce's own CRM Benchmark — a suite of real-world tasks like those — the company says Koa already matches or exceeds leading general-purpose models on CRM actions with 3x fewer errors.
Here's the thing nobody's talking about: "3x fewer errors" is the whole ball game in enterprise AI. A consumer chatbot that hallucinates one time in fifty is a curiosity; a support agent that closes the wrong ticket or promises the wrong refund one time in fifty is a liability. If those numbers hold up outside Salesforce's own benchmark, this changes procurement conversations.
03.The trust boundary detail that matters for India
Salesforce controls the model weights, and both post-training and inference happen entirely inside its own infrastructure. That sounds technical, but it's the difference between "your customer data helped train a model you don't control" and "the model lives where your data lives." For Indian banks, insurers and government departments — organisations that have strict rules about where citizen and customer data is processed — this is the part worth underlining.
Salesforce and NVIDIA are also bringing Nemotron-based models and accelerated computing into Missionforce, extending that same control to government and regulated organisations — private clouds, air-gapped networks and other secure environments. Given how much of India's AI push is now explicitly sovereign (something we saw with the NTA running NEET translation on its own GPUs), that "deploy anywhere" posture is well aimed at this market.
04.Rollout: live internally, customer pilots next
Koa is already running inside Salesforce's own employee workflows, and the company has moved into customer pilots with a small group running agents in service, sales and commerce. A broader release will presumably follow the usual Dreamforce-to-GA cycle. No India-specific pricing has been announced yet, so treat any rupee figures you see online as speculation.
05.The lesson for small teams
You're probably not going to post-train Nemotron this weekend. But the playbook is copyable at any scale: don't use one giant model for everything. Use a cheap, specialised model for the high-volume, repetitive decisions and save the expensive general one for the messy edges. It's the same cost-optimization logic we walk through in the agents and cost-control chapters of our GPT-6 Astra — Zero to Hero Guide — Salesforce just did it with 27 years of data and a chip partner.
Key Takeaways
- Koa is Salesforce's first CRM reasoning model, built on post-trained NVIDIA Nemotron 3 Super
- Trained on a synthetic dataset distilled from ~27 years of CRM deployments
- Reportedly matches or beats leading models on CRM tasks with 3x fewer errors (Salesforce's own benchmark)
- Runs inside Salesforce's trust boundary; Nemotron models are also headed to Missionforce for governments and regulated industries
- Already live internally at Salesforce; customer pilots underway in service, sales and commerce
Frequently Asked Questions
What is Salesforce Koa?
Koa is Salesforce's first CRM reasoning model, announced September 16 at Dreamforce. It was built by post-training NVIDIA's Nemotron 3 Super and powers the company's Agentforce AI agents.
What data was Salesforce Koa trained on?
Koa was trained on a proprietary dataset distilled from nearly three decades of CRM deployments — roughly 27 years of Salesforce's own deployment experience.
How accurate is Salesforce Koa?
Salesforce says Koa matches or beats leading general models on CRM tasks with 3x fewer errors, according to its own benchmark. It runs inside Salesforce's trust boundary, and Nemotron-based models are also coming to Missionforce for governments and regulated industries, including private clouds and air-gapped networks.
What This Means for You
If you work in sales, support or operations at an Indian company, the era of "one AI for everything" at work is ending — you'll soon have quiet, specialised models making the routine calls behind the scenes. That's a genuine opportunity for professionals who understand how agents and workflows fit together, because someone has to design those processes and catch the errors the model doesn't. If you're early in that journey, our GPT-6 Astra — Zero to Hero Guide takes you from basic prompting to building your own working agents, step by step — the same shift Salesforce just spent millions riding.