If you only read launch announcements, you’d think AI agents were already running half of corporate America. If you read the surveys, you get a different picture: a lot of experiments, a lot of demos, and surprisingly few agents doing real work at scale.
I spend a lot of time thinking about this gap, partly because as a business owner I’m on the buying side of it. Here’s what the data says, why I think the gap exists, and what seems to separate the companies that get past the pilot stage.
What the numbers say
The most useful long-running data comes from McKinsey’s annual global survey on AI.
In its 2025 edition, based on a survey of nearly 2,000 people in mid-2025, 62% of respondents said their organizations were at least experimenting with AI agents. Only 23% said they were scaling an agentic system anywhere in the business, and in any single business function, no more than 10% said they had scaled agents.
The 2026 edition, from a survey run in May and June of this year, shows movement, but very uneven movement. Among organizations with more than $1 billion in annual revenue, the share scaling agents in at least one function rose from 27% to 40% over the year. Among smaller organizations, it stayed flat at 22%. Even now, the function where agents are most often scaled, IT, comes in at just 12% of respondents.
Meanwhile, the payoff is still elusive. In the 2026 survey, 37% of respondents attributed any EBIT impact at all to their use of AI, which McKinsey’s data shows is essentially unchanged from a year earlier.
Analysts expect plenty of failures along the way. In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
Why agents get stuck
None of this means agents don’t work. In the right setting, they clearly do. The problem is getting from “it worked in the demo” to “it runs every day and we trust it.” I see five recurring reasons.
1. The demo hides the plumbing
A demo agent works on clean sample data with a few hand-picked tools. A production agent has to connect to the real CRM, the real ticketing system, the real permissions model, and the twenty-year-old system nobody wants to touch. That integration work is usually most of the project, it rarely shows up in the pitch, and it’s a big part of why escalating cost tops Gartner’s list of reasons agent projects get canceled.
2. Long tasks expose small error rates
An agent that’s right 95% of the time on each step sounds impressive. Chain twenty steps together and the odds of a clean run drop sharply. Short demos hide this. Real workflows reveal it. I explain why in more detail in our guide on what an AI agent is.
3. Nobody agreed who’s accountable
When an agent sends the wrong email to a customer, whose problem is that? The team that built it? The team that owns the customer? The vendor? Companies that never answered this question tend to keep agents in “pilot” forever, because no one is willing to own them in production. That’s a governance problem, not a technical one.
4. The value case was never measured
Many pilots start because a technology is exciting, not because someone measured the cost of the current process. Without a baseline, it’s impossible to prove the agent is better, and the project dies at budget time. Gartner’s advice is blunt: pursue agentic AI only where it delivers clear value or ROI.
5. Some “agents” aren’t agents
Gartner also warned about “agent washing,” the relabeling of existing chatbots and automation tools as agents. It estimated that only about 130 of the thousands of vendors claiming agentic AI were real. If a product was never really an agent, it’s not surprising that it doesn’t scale like one.
Why big companies are pulling ahead
The jump among $1 billion-plus companies, from 27% to 40% in a year, while smaller firms stood still, is the most interesting number in the 2026 data. My read is that the bottlenecks above favor scale. Large companies can fund integration teams, build internal platforms, and afford a few expensive failures. Smaller businesses usually can’t.
But I don’t think small companies are destined to fall behind. They have one big advantage: fewer systems and fewer approval layers. A 20-person company can pick one painful, repetitive process and fix it in a month. The trick is to choose carefully.
What the companies that scale seem to do differently
Across the reporting and case studies I’ve read, the pattern is consistent:
- They start narrow. One process, one team, one measurable outcome. Not “an agent for operations.”
- They pick checkable work. Tasks where a person can verify the result faster than doing it themselves: triaging tickets, drafting first responses, reconciling records, writing code that has to pass tests.
- They often start with a workflow, not an agent. A fixed sequence of AI steps is cheaper and more predictable. They add autonomy only where the path can’t be scripted.
- They measure first. They know how long the task takes today, what it costs, and how often it goes wrong, so they can prove improvement.
- They decide ownership upfront. Someone owns the agent’s output, the same way someone owns a process or a product.
- They design for oversight. Approval steps for anything with side effects, logs of what the agent did, and limits on what it can access.
That last point isn’t only about internal risk. Agents operate in the outside world too. The Wikimedia incident this month showed what can happen when agents act on other people’s systems without enough oversight. And if your agents can read company email or files, the data risks in our guide on using AI at work without leaking sensitive data apply directly.
My take
The gap between experimenting and scaling isn’t a sign that agents are overhyped. It’s a sign that they’re being treated like a feature to switch on, when they’re really an operational change that needs an owner, a budget, and a way to measure success.
If you’re deciding where to start, I’d pick the most boring, repetitive, easy-to-check process you have, measure it, and give one person responsibility for making an agent handle it well. That’s less exciting than a demo. It’s also how the 40% got there. And as agents take over routine tasks, it’s worth thinking now about what that means for the people who used to learn the job by doing them, a question I explore in why entry-level jobs are the front line of AI.
Sources
- McKinsey, The state of AI in 2025: Agents, innovation, and transformation, November 2025
- McKinsey, The state of AI in 2026: On the road to ROI, 2026
- IDM, Agent use follows industry value, not hype, 2026
- Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, June 25, 2025
