Salesforce and Nvidia built their own reasoning AI so enterprises can stop paying OpenAI and Anthropic

Key points

• Salesforce unveiled Koa, its first reasoning AI model, at its Dreamforce conference in 2025.
• Koa is built on Nvidia's Nemotron, an open-weight model (meaning the underlying code is publicly available, not locked behind a paywall).
• The model was trained on synthetic, made-up data rather than real customer records, so no private information was used.
• Koa costs less to run than sending the same tasks to ChatGPT or Claude because it uses fewer tokens (the small chunks of text an AI processes; fewer tokens means a lower bill).
• Salesforce is also launching ClaudeForce, a partnership that lets companies use Anthropic's Claude AI while keeping their data inside Salesforce's own secure infrastructure.

Salesforce built its own reasoning AI model, and the frontier labs should pay attention.

The company announced Koa this week at Dreamforce, its annual mega-conference. Koa is Salesforce's first model capable of reasoning, meaning it can work through long, multi-step problems rather than just answering a single simple question. Before Koa, whenever a Salesforce customer's AI agent hit a tricky task, the system handed the job off to a frontier model like Anthropic's Claude or OpenAI's ChatGPT. That hand-off cost money every single time. Now Salesforce wants to keep that work in-house.

What does Koa actually do?

Koa lives inside Agentforce, Salesforce's platform where businesses build AI agents (software programs that carry out repetitive tasks like answering customer queries or scheduling appointments without a human stepping in). Until now, those agents relied on Salesforce's own small, task-specific models for simple jobs and outsourced harder reasoning to OpenAI or Anthropic.

Koa handles the harder stuff itself, at lower cost. The model was built by taking Nvidia's Nemotron as a starting point. Nemotron is open-weight, which means any company can download and customise it without paying a licence fee. Salesforce and Nvidia then post-trained it, a process where you take a general-purpose AI and teach it to specialise, drilling it on sales conversations and customer-service scenarios. We first covered Nemotron on 16 July 2026, and this is the most significant enterprise deployment of it we've reported since.

None of that training used real customer data. Instead, the two companies built synthetic data: artificial conversations that mimic real interactions without containing actual private details. Salesforce EVP of AI Jayesh Govindarajan told TechCrunch they simulated everything from frustrated callers to sales reps trying to close a deal.

Why does the training data matter to ordinary users?

It matters because leaked customer data is one of the biggest fears businesses have about AI. Koa was designed so that risk simply does not exist at the training stage. Customer records stay inside Salesforce's own infrastructure and were never fed into the model. That is a concrete privacy safeguard, not a marketing promise.

Koa also fits inside Salesforce's AI gateway, the internal routing system that decides which AI model should handle which request. A business can set rules about data security and cost, and the gateway automatically sends work to the right model without anyone managing it manually.

What about OpenAI and Anthropic?

Salesforce isn't cutting them out. The company announced ClaudeForce at the same event: a partnership with Anthropic that lets businesses use Claude as their main AI interface while all underlying data stays locked inside Salesforce's own systems. It's less a break-up and more a renegotiation of terms. Salesforce's pursuit of adjacent capabilities is aggressive right now; our 10 September story on Listen Labs walking away from a $1.5 billion funding round shows how far the company is reaching to own its AI stack.

But Koa signals something real. Big enterprise customers are getting comfortable building, or co-building, their own models rather than writing a blank cheque to a frontier lab each month. Koa is a proof of concept that a purpose-built, data-safe, cheaper model can beat a general-purpose giant at a specific job. Other large software platforms will be watching closely.