One Year After the AI Agent Boom: What Enterprise Adoption Really Looks Like
In early 2025, Operator and Manus ignited talk of the year of the agent. By November 2025, McKinsey's survey told a cooler story: 62% of companies are experimenting with agents, yet scaled use tops 10% in no single function. A look at what sits between the hype and the numbers.
Key takeaway
McKinsey's November 2025 survey: 62% of companies are at least experimenting with AI agents and 23% have scaled somewhere, but in no single function does scaled use exceed 10%. The gap is not model capability but permissions, data, exception handling and acceptance criteria. Start narrow.
Look back at the 2025 calendar and three dates trace the arc of the agent boom. January 23: OpenAI released the Operator research preview, an AI taking over a cloud browser to complete web tasks for its user. March 6: Manus launched, pushing the idea of a general-purpose agent into the mainstream — invite codes were resold for hundreds to tens of thousands of yuan. July 17: Operator was folded into ChatGPT's agent mode and its standalone site retired. The agent went from standalone experiment to a standard feature of a mainstream product.
After a year of noise, what actually happened inside companies? In November 2025 McKinsey published its annual State of AI survey, fielded June 25 to July 29, 2025, with 1,993 respondents across 105 countries. The numbers repay a careful read.
62% experimenting, under 10% scaled in any single function
The survey found 62% of companies at least experimenting with AI agents — 23% had scaled agentic systems somewhere, while another 39% were just beginning to experiment. Zoom into the scaling detail, though, and the optimism thins: most of those who scaled cover only one or two functions, and within any single function, no more than 10% of companies report scaled agent use.
Between a 62% experimentation rate and under-10% scaling in any single function sits not model capability, but permissions, data, exception handling and acceptance criteria.
Put differently, when a company said we use agents in late 2025, the accurate version was usually we are trying agents in one corner. The gap between expectation and reality is not about whether to do it — it is the unglamorous stretch between a pilot and daily operations.
Why so many trials, so few successes
From what we observe in enterprise projects, the blockers cluster in three places. One: permissions and data are not ready. For an agent to act on someone's behalf it needs access to business systems and clean data interfaces, and most enterprise systems were designed for human users — no account model or operating boundaries for a machine employee.
Two: the cost of exception handling is underestimated. In the demo the agent glides through; in production it meets redesigned pages, forms with missing fields, ambiguous requests — and every exception type needs a hand-back-to-human rule, engineering that often outweighs the happy path itself. Three: acceptance criteria stay vague. Smarter and more efficient cannot be signed off; without checkable pass conditions of the kind described in from AI pilot to production, a pilot can only remain a pilot.
The scenarios that work are narrow
The two areas where agents are used most, per the survey, are IT (the service desk) and knowledge management (deep research); by industry, tech, media and telecom, plus healthcare, lead the pack. Note what these scenarios share: crisp task boundaries, enumerable actions, controllable error costs — a failed password reset can be retried, a missed document re-queried. Which confirms a plain judgment: agents land first not where intelligence shows best, but where success is easiest to define.
High performers don't have better models; they have better processes
McKinsey classifies roughly 6% of surveyed companies as high performers, and this group is more than three times as likely as peers to have scaled agents across most functions. From the public material, their common trait is not pricier technology but treating agents as an occasion to redesign the process: take the workflow apart, define each step's inputs, outputs and human-machine boundary, then hand the suitable steps to an agent — rather than stacking an agent on top of the existing process and hoping it adapts.
For smaller companies: win one scenario, skip the do-everything agent
Smaller companies lack a big firm's budget for failed experiments, which makes these numbers a map rather than a cold shower. The pragmatic route: skip the general-purpose agent, pick one narrow scenario — auto-answering routine inquiries, assembling fixed-format documents — and define permissions, exception rules and acceptance criteria before switching it on. To judge whether your process wants a workflow or an agent, see AI workflow or AI agent; if the concept itself still feels fuzzy, start with what an AI agent is.
Three signals worth watching in 2026
For the coming year, skip the demo reels and watch three kinds of signals instead. One: vendors publishing operational metrics rather than capability demos — task completion rates, human-takeover rates, cost per task. Two: whether scaled cases spread beyond IT and knowledge management into functions closer to revenue, such as sales and finance. Three: what next year's figure is for no more than 10% scaled in any single function. Numbers are more honest than stories; the ledger is worth checking again this time next year.