Salesforce

Marc Benioff said in June 2025 that AI was doing 30% to 50% of the work at Salesforce. Eleven months later, Srinivas Tallapragada, President and Chief Engineering Officer, published the engineering half of that claim in detail: past 90% AI adoption, his organization standardized company-wide on «Claude Code» and removed all token limits, on the doctrine that a token cap is friction to eliminate. April 2026 against April 2025: work items per developer +50.8%, merged PRs +79%, and an Effective Output score (an ML measure of the code's real value) +151.3%. One migration of 33 API endpoints, estimated at 231 person-days, closed in 13 days across 5 PRs, the largest delivering 21 endpoints at 100% coverage. Incidents fell 5% over the same period.

The same company invented the seat-based SaaS model that Elliot Greenwald identifies as the structural conflict of legacy customer-experience vendors: the better their AI, the fewer seats their clients need. Salesforce cut 4,000 customer support roles, raised prices 15% between 2020 and 2024, and is read by Jaya Gupta and Ashu Garg as a system of record that stores current state rather than the state at the moment of decision, which is why they argue it cannot build a context graph. Yamini Rangan positions «HubSpot» against it as AI-native rather than AI-added.

Tallapragada names what stayed hard: variable quality in CLAUDE.md files, agents that act rather than suggest, and a blast radius that he says demands a fundamentally different security model.