Salesforce

Marc Benioff stated in June 2025 that AI was doing 30% to 50% of the work at Salesforce. On 27 May 2026, Srinivas Tallapragada, President and Chief Engineering Officer, documented the engineering share: +50.8% work items per developer, +79% merged PRs, and a +151.3% Effective Output score, an ML measure of code value rather than volume. A migration of 33 API endpoints estimated at 231 person-days shipped in 13 days, across 5 PRs. The two levers he names are org-wide standardization on Claude Code and the removal of every token limit, which he treats as friction rather than a budget guardrail ("we removed all token limits"). Incidents dropped 5% as PR volume rose. Tallapragada also names what stays hard: variable CLAUDE.md quality, and agentic security, which he argues needs a fundamentally different security model. Didier Girard cites these figures as evidence for his AI4IT thesis: code verifies itself through compilation, tests and CI, while business processes do not.

The same company appears on the other side of the ledger. David Heinemeier Hansson places Salesforce and SAP "en sérieuse difficulté": the value has left the code and sits only in process intuitions he judges uncertain. Jaya Gupta and Ashu Garg note that Salesforce records current state, not state at the moment of decision, so the decision traces agents need accumulate elsewhere. Salesforce created the seat-based SaaS model; it also cut 4,000 customer support roles and raised prices 15%.