For two years every Salesforce roadmap meeting had the same awkward question. Which AI should we bet on? At Dreamforce 2026 Salesforce gave the most generous answer possible. Pick any of them, and now there is a brand new one built specifically for CRM. 🧠
Meet Koa
Koa is the first Salesforce CRM reasoning model, built together with NVIDIA and trained on a synthetic dataset modeled on almost three decades of CRM deployments. Salesforce controls the weights and runs Koa inside its own infrastructure, so customer data stays within the trust boundary during inference, and no customer data was used for training. Koa is in pilot with select customers now, and general availability is expected in winter 2026 for US regions.
Everyone else is invited too
| Model | Where you will meet it in Salesforce |
|---|---|
| Koa | CRM reasoning inside Salesforce infrastructure |
| Claude | Salesforce in Claude, in beta on paid Claude plans, and powering Agentforce Vibes by default |
| Gemini | Prompt Builder and the Agentforce reasoning engine, plus Salesforce inside Gemini Enterprise |
| Amazon Bedrock models | Available to Agentforce customers now |
| OpenAI models | Inside Agentforce, with Agentforce also reachable from ChatGPT |
So who actually wins
Your data does. When the model becomes a choice, picking the wrong one gets much cheaper, and the real value moves to what every model depends on. Clean records, sensible permissions and well documented business rules now decide how smart any of these models look inside your org.
Three questions for your next leadership meeting
Answer them now and next year's model announcements turn into an easy upgrade instead of a long debate.
Salesforce just turned AI from a single bet into a full menu, and the companies with the cleanest data will order best.
Catch up on the rest of the series with the full announcement recap, or see AIforce and the permission cleanup it just made urgent.
