Agentic AI — software that can plan and carry out multi-step tasks on its own rather than just answering a single prompt — moved from pilot projects to something closer to standard infrastructure in large organizations this year, according to Salesforce's 2026 Agentic Enterprise Index. The report, published by a company that sells agentic AI products, found that the number of activated AI agents per organization nearly tripled, rising from roughly 5 to 13 over about 14 months. By April 2026, organizations using Salesforce's agent platform had collectively performed 734 million "agentic work units" — a measure of discrete tasks completed autonomously by an agent.
The report points to one specific metric as evidence that trust is catching up with usage: agent-led customer service chats grew roughly 170-fold over five quarters, while the rate at which those conversations escalated to a human agent held flat at around 32%. In plain terms, as agents handled dramatically more conversations, they weren't escalating a larger share of them to humans — a sign, Salesforce argues, that agent performance is holding up even as volume scales.
A separate, independent survey from Futurum Group tells a more cautious story about what's stopping agentic AI from going further. It found that 55% of organizations name reliability and hallucination management as their top challenge in adopting AI agents, with privacy and security concerns close behind at 53%. Put together with Salesforce's own numbers, the picture is less a story of trust fully resolved and more one of usage significantly outpacing confidence — companies are deploying agents faster than they're fully solving the problems that make agents risky to deploy.
For job seekers and people already employed in roles that touch customer service, operations, or IT support, this gap between adoption and trust is actually where a lot of the near-term hiring demand sits. Organizations rolling out agents at this pace need people who can monitor agent output, design escalation rules, audit for errors, and fix the parts of a workflow where an agent predictably gets things wrong — not just people who can configure an agent and walk away. Roles with titles like "AI operations," "agent orchestration," and "AI quality" have been appearing more frequently on job boards this year, often as an evolution of existing QA, support, or operations roles rather than entirely new departments.
The practical signal for anyone building a career around this: agentic AI is real and growing fast, but the skills in shortest supply right now are less about building agents from scratch and more about the oversight, correction, and governance work that keeps them from making costly mistakes at scale.