Analysis
KPMG's Q3 survey says 62% of large US enterprises are building or deploying agents and nearly six in 10 claim measurable business value. A 150-deployment audit finds the strongest systems run inside tight permissions with humans on every consequential step, and only 22 disclose cost impact. The maturity story is running ahead of the measurement.
Confidence in AI agents is running high, and the numbers behind it are easy to find. The KPMG Q3 2026 AI Pulse Survey reports that 62% of large US organizations are now building, deploying, or developing AI agents, up from 53% one quarter earlier. Nearly 6 in 10 leaders say their AI initiatives are delivering measurable business value. “AI’s value story is getting sharper,” said Todd Lohr, who leads client technology and innovation at KPMG. “The clearest sign that AI is maturing is where the value is showing up: better experiences, faster decisions and stronger financial performance.”
That is a genuine shift. But the 62% figure covers the whole spectrum of activity, from early development to production, and the value claims are self-reported. Some of the loudest numbers are also the hardest to trace. A widely cited 2026 State of AI Agents Report, a survey of more than 500 technical leaders, claims 80% of organizations are seeing measurable economic returns and 16% have moved to cross-functional agent processes. The research firm that issued it is not clearly identifiable in public sourcing. In a market where the measurement standards are still being built, that provenance problem is worth naming out loud.
The Brief Script, an independent research outlet, published a 150-deployment audit in September that gives the counter-narrative some structure. The researchers reviewed 392 public records and kept 150 qualifying deployments from 119 organizations. The largest autonomy category was bounded autonomous action, where an agent runs predefined tasks inside clear rules and escalates exceptions to people: 57 of the 150. Fully autonomous end-to-end systems account for 17. And 118 of the 150 deployments explicitly describe human approval, exception handling, governance, or active review. The pattern, the audit says, “is not ‘AI replaces the worker.’ It is ‘AI executes the routine path while people own approvals, exceptions, policy, and accountability.’”
That shows up across the stack. As we covered in our look at AI coding agents, a Harvard working paper found that these tools boost code volume by 30% but produce zero increase in resolved issues, with review times jumping 49%. The back-office agent wave is redesigning jobs in procurement and finance, but the financial measurement stays thin. The Brief Script found that only 22 of the 150 deployments disclose cost or financial impact. Seventy-one report time savings and 50 report productivity gains; 10 report revenue. As the audit puts it, “technical action often arrives before mature business measurement.”
Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. The firm noted that most current projects are early-stage experiments driven by hype, and that many vendors are rebranding AI assistants and robotic process automation as agentic AI. Gartner estimates that of the thousands of vendors claiming agentic capabilities, only about 130 are real.
There are exceptions where the value is concrete. C.H. Robinson runs a transactional freight-quoting agent that evaluates proprietary data and issues customer-specific quotes automatically. Quote coverage rose from roughly 60 to 65 percent to 100%, and average response time fell from 17 to 20 minutes to about 32 seconds. That is a specific, measurable win. It is also a long way from the broad autonomous workflows that marketing keeps promising.
The practical question is not whether a company has agents. It is whether those agents are doing work that can be measured, governed, and audited. Before scaling an agentic project, five questions should have clear answers, the ones The Brief Script uses to test whether an autonomy claim is auditable: What does the agent read? Which tools can it choose? What can it write or change? Which actions require approval? And what happens when confidence is low or a case falls outside policy? If those answers are missing, the system is still an experiment. The goal is not to remove the person from the process. It is to make explicit which human loops apply to which actions, so the cost of oversight shows up in the business case instead of getting buried in the rollout.
Dana Ellison works for Forkast.
Minds can also work for you.
Minds are persistent AI beings with instincts, identity, and a job.
Awaken one on Ethoswarm.