Salesforce Explains How It Trained Its Own Reasoning Model
Salesforce's AI engineering leaders published an essay on September 16, 2026 explaining how and why they post-trained Koa, the CRM reasoning model Salesforce announced a day earlier with NVIDIA. This is the how-it-was-built piece, not the product announcement.
Transcript
Salesforce explained today how it trained its own reasoning model. It says the industry default gave limited gains.
So Salesforce used reinforcement learning. It says Koa repeated simulated customer scenarios thousands of times while a judge scored each result.
Salesforce says a general model without the right tool can confirm an action that never happened. Koa was trained to stop instead.
Salesforce says an open model let it change the weights. It could also see the training data NVIDIA published.
Salesforce says your data never trains the model. It calls early results encouraging, with no numbers and no outside audit.
Inboxsmith builds AI phone and email agents for small business. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- Supports every claim in the script: that supervised fine tuning gave limited gains, the reinforcement learning simulation with synthetic customers and a scoring judge, the deliberate missing tool behaviour, the open model weights and published training data rationale, the claim that customer data and traces never train the model, and the unquantified early benchmarking claim.(Salesforce (salesforce.com/news), first party corporate newsroom essay titled "Why We Post-Trained Our Own Reasoning Model", bylined Jayesh Govindarajan and Silvio Savarese, published September 16, 2026)