Salesforce published its second annual Agentic Enterprise Index on August 7, 2026, drawing on production usage data from customers who kept AI agents live every month between February 2025 and April 2026, supplemented by a May 2026 survey. The headline figures give the clearest industry benchmark yet for how fast agentic AI is actually being operationalized inside large organizations, as opposed to how fast it's being talked about. The average business went from running five agents in production in February 2025 to thirteen by April 2026, a compound monthly growth rate of about 7%. Deployment friction also dropped sharply: it now takes an average of 1.9 days to get an agent into production, down 53% over the same period, suggesting that tooling and internal processes for shipping agents have matured considerably. Once deployed, agents are also being trusted with more responsibility over time, with the average number of actions performed per account growing at roughly 31% per month. Salesforce also introduced a "Sophistication Index," a five-point scale distinguishing simple agent tasks like record lookups or drafting emails (levels 1-3) from more consequential actions like updating database fields (levels 4-5), and found that manufacturing, financial services, and healthcare have actually built more sophisticated agent deployments than traditionally AI-forward sectors like technology and retail. Worth flagging for a technical audience: Salesforce measures agent output using its own "Agentic Work Unit" metric rather than token consumption, a choice that lets it claim 734 million AWUs performed as of April, but that metric has drawn criticism from independent analysts as not clearly tied to actual business outcomes. For engineering and product leaders evaluating their own agent rollouts, the useful takeaway is less the specific numbers and more the pattern: successful agent programs tend to start with narrow, well-scoped use cases (customer service came up as the best ROI starting point) before expanding scope, and vendor-provided productivity metrics are worth treating skeptically until validated against your own outcome data.