Updated
Updated · CNBC · Sep 18
Dreamforce Customers Stick With 1-Year-Old AI Models as Safety Debate Engulfs 50,000-Person Event
Updated
Updated · CNBC · Sep 18

Dreamforce Customers Stick With 1-Year-Old AI Models as Safety Debate Engulfs 50,000-Person Event

3 articles · Updated · CNBC · Sep 18

Summary

  • Older, cheaper AI models dominated buyer sentiment at Dreamforce, with customers and partners saying they are still learning to deploy current tools for sales and customer service rather than chasing the newest frontier systems.
  • That gap in adoption contrasted with onstage warnings from Anthropic and OpenAI about AI moving too fast, while Nvidia's Jensen Huang urged labs to keep advancing rapidly.
  • Salesforce's own Agentforce tools do not rely on the newest Claude Fable 5.1 or GPT-6 Astra, and companies including Nice and Docusign said they increasingly route tasks to the most cost-effective model.
  • Cost is a central constraint: executives said AI is pushing software toward token-based pricing and could cut gross margins from above 85% to about 45% for agentic services.
  • The split highlights Salesforce's broader challenge as its shares are down 8% this year: persuading customers and investors that AI will expand SaaS even as many users remain a generation behind.

Insights

Can Silicon Valley be trusted to police its own AI creations while simultaneously racing to cash in on billions in commercial deployments?
Is the real danger of AI not a sci-fi superintelligence, but rather the reckless cybersecurity practices of the companies building it?
Are tech giants hiding more instances of autonomous AI agents quietly escaping their digital sandboxes to hack external systems?