Salesforce Unveils AIforce and Core CRM Inference Model at Dreamforce
The AMW Read
AIforce and Core materially extend Salesforce's enterprise-operations AI stack by combining CRM context, governed actions, and model deployment options.
Named counterparties: Slack
Salesforce Unveils AIforce and Core CRM Inference Model at Dreamforce
Salesforce introduced AIforce, a real-time interface layer for its CRM platform, alongside Core, a CRM-focused inference model built with NVIDIA on Nemotron 3 Super. Announced at Dreamforce 2026 in San Francisco, AIforce connects Salesforce data, workflows, business logic, permissions, and governance across interfaces including Claude, Slack, and Agentforce Coworker. The company says Core is designed to interpret complex business context and select tools for CRM tasks; Salesforce also cited internal benchmarks showing lower error rates than leading models.
The release shifts Salesforce's agent strategy beyond a standalone assistant toward an enterprise control plane for context, permissions, and action execution. Rather than asking employees to move between dashboards, the company is positioning natural-language requests as the entry point for dynamically assembled work surfaces. Its zero-data-retention approach for external model providers addresses a central enterprise constraint: companies want access to model capability without rebuilding identity, authorization, and data-control systems around each new AI interface. This extends Salesforce's recent push into multi-agent governance and usage-based Agentforce pricing by tying adoption more directly to existing CRM workflows.
For builders, the implication is that integration with systems of record may matter more than another generic conversational interface: tool reliability, semantic context, and permission-aware execution become product requirements. For investors, Salesforce's partnership with NVIDIA and support for private-cloud and air-gapped deployments underscore how regulated-industry demand can reward vendors that package models with deployment control, not just model performance.
