The average company now runs roughly 13 AI agents, up from around 5 just over a year ago — a real, documented near-tripling in enterprise AI agent adoption. But the same period also produced a real, uncomfortable finding on the human side of that shift: a majority of developers using AI tools daily describe the experience as feeling more like dependence than a productivity advantage. Both numbers are real. Understanding what each one actually measures — and doesn’t — is the difference between reading this moment accurately and reading the headline.
The real adoption numbers
Salesforce’s Agentic Enterprise Index, now in its second edition, combines usage-log data from its Agentforce platform across thousands of business customers with a survey of 4,689 respondents fielded in May 2026, tracking the period from February 2025 through April 2026. The headline figure: the average number of AI agents deployed per business rose from roughly 5 to roughly 13 over that window — a 7% compound monthly growth rate sustained for 15 straight months. Alongside that, the time needed to build and deploy a new agent fell 53%, from about 4 days down to 1.9 days, and the number of distinct skills each agent could perform roughly tripled, from 2 to 6. Salesforce logged 734 million “Agentic Work Units” in April 2026 alone, growing roughly 15% month-over-month, and reported that retail businesses using agents on digital commerce saw 4x higher online sales growth than non-adopters.
The real caveat that changes how to read those numbers
Here’s the detail worth taking seriously before treating that growth curve as a picture of the broader economy: this index measures Salesforce’s own “committed adopters” on its Agentforce platform — not a random cross-section of businesses generally, and not companies still evaluating or piloting agent technology. It’s a real, useful dataset for understanding how fast agent usage scales once a company has actually committed to the platform, but it isn’t independent, third-party research on enterprise AI adoption broadly. The honest read is: among businesses that have already bought in, usage is scaling fast and getting more capable. That’s a genuinely different claim than “AI agents are sweeping the average workplace,” and the distinction matters for anyone using this data to make a real business decision.
The real skepticism on the human side
Set against that adoption curve, a 2026 developer survey found 80% of developers describe their daily AI coding tool usage as feeling more like dependence than a genuine productivity advantage — alongside real, specific related findings: 43% keep coding with AI after hours despite intending to stop, 51% report a real burnout risk tied to the always-on nature of the tools, and developer trust in AI-generated code accuracy actually fell, from 40% down to 29%, even as usage kept climbing. That’s a real, documented gap between adoption and confidence — more people using the tools, with less trust in what the tools produce.
A rigorous, independent randomized controlled trial adds real weight to that skepticism. METR’s 2025 study gave 16 experienced open-source developers 246 real GitHub issues, randomly assigning each one to be completed with AI tools allowed or disallowed. The actual result: developers were 19% slower when using AI tools — but afterward, they believed AI had made them 20% faster. That’s roughly a 39-percentage-point gap between the real, measured outcome and what the same developers genuinely believed had happened, one of the clearest documented examples of the AI productivity paradox measured under real experimental conditions rather than self-reported survey impressions alone.
How to actually hold both findings at once
These aren’t contradictory numbers — they’re two real, separate measurements of two different things. Enterprise agent deployment is genuinely scaling fast among companies that have committed to the technology, with real efficiency gains in how quickly new agents get built and how much they can do. Separately, and just as genuinely, individual developers using AI coding tools day to day report a real, measurable gap between how productive the tools feel and how productive they actually are, alongside real dependence and trust concerns. A company can be legitimately accelerating its agent deployment at the platform level while its own developers are individually experiencing exactly the kind of productivity paradox METR documented — both things are true at the same time, and treating either number as the whole story would be a real mistake.
The honest takeaway
The real story in 2026’s enterprise AI data isn’t “adoption is exploding” or “AI tools are overhyped” — it’s that both a genuine platform-level acceleration and a genuine individual-level productivity paradox are happening simultaneously, documented by real, separate data sources. Any organization scaling AI agent deployment should be tracking both curves, not just the one that makes for a better headline.
For the full, regularly updated picture of what else is happening in AI right now, see our daily AI news briefing.
