What Salesforce announced
The plan is called AIforce, and it is built to carry Salesforce data, business rules and workflows out of Salesforce and into whatever AI a company already uses, without setting up a new permissions system and without the AI provider keeping the data once a query is done.
The headline piece is Claudeforce, which drops Salesforce into Anthropic's Claude as a ready-made connection with 37 prebuilt sales skills for everyday work like call prep, pipeline reviews and record updates. There's a Slack version in the mix too, and a new reasoning model called Koa that Salesforce built with Nvidia and trained on 27 years of its own CRM data.
It was when the conversation moved from features to safety that the mood on stage shifted.
Where the partners split
Nvidia chief Jensen Huang framed safety as something you build your way through rather than a reason to slow down.
"Safety is paramount. In a lot of ways, it's job one. However, safety is an engineering problem." Huang said.
Anthropic's Dario Amodei has spent years arguing closer to the opposite, that the risks of powerful AI are serious enough to pull back and move carefully, and those two views sat uneasily side by side in front of a room full of the very customers Salesforce now wants to hand its data to.
One partner was selling confidence, the other has made his name selling caution, and the gap between them was hard to miss.
What has not been confirmed
Salesforce has not said what any of this will cost, when it will be generally available, or which regions get it first. It also has not spelled out how its safety promises hold up once a customer's data is sitting inside a model the company itself does not control.
Nvidia and Anthropic have not put out any shared line on where they actually agree, and for all the talk of trust on stage, the harder question of who is on the hook when an outside AI mishandles CRM data went unanswered.