Where the Agent Pilot Stops

Earlier this month I wrote about how far companies have got with AI agents. The short version, from McKinsey's 2026 State of AI survey: nearly nine in ten respondents say their organization regularly uses AI in at least one business function, and about two in ten say it has scaled agents across the organization. Among organizations with more than $1 billion in revenue, the share of respondents scaling agents in at least one function rose from 27% to 40%; among smaller organizations it stayed at 22%. A year earlier, no single business function had more than 10% of respondents scaling agents in it. Those are different measures and do not convert into one another, but none of them is a majority.

So if your organization has an agent, the odds are that it is a pilot rather than a process. This piece is about what happens to the pilot. I want to describe four places where it stops, in the order a pilot reaches them, and give you a way to tell which one you are at. The evidence is thinner than I would like, and I will say where it is thin.

Nobody can accept the result

The first stop is the demo that cannot be signed off. The agent produces an output, the output looks right, and when someone asks how we know it is right, the only answer is to read it.

I ran into this directly. In one test I graded 120 runs of an agent whose checklist asked it to record how it had verified its own work. With the checklist advisory, none of the 60 runs with public records wrote that note. On a summary task with a local model (qwen3:14b), all 17 runs that finished declared the work done. Enforcing the checklist made the note common and did not fix more bugs. The agent produced the paperwork saying it had checked.

That is a small sample on a local model, not a production system. But it sits next to what separates McKinsey's high performers from everyone else. High performers, the respondents who attribute at least 5% of EBIT to AI and say their organization has seen "significant" value from it, are 6% of the sample, unchanged from last year. Nearly three-quarters of them report fundamentally redesigning workflows because of their AI use, up from 55% a year ago. One-quarter of the other respondents report doing so. High performers are also twice as likely to say their organization has defined processes to measure the impact of its AI initiatives.

The survey cannot tell you that redesigning the workflow causes the value. These are self-reported answers, and the same respondents describe both the redesign and the result. It also asks nothing about acceptance steps. My guess at the link is this: redesigning a workflow around an agent forces someone to write down what done means for that workflow, in a form a person or a test can check. Dropping an agent into the old workflow does not force that, so the demo stays a demo.

The bill arrives per task

The second stop comes after acceptance. Once a task counts as finished, someone can divide the invoice by the number of finished tasks, and in my experience the result is larger than the pilot budget implied.

In the McKinsey survey, one in five respondents say their organization is limiting AI use because of operating costs. For each of chatbots, agents and coding agents, about one in ten say use of that tool has been constrained by cost. The detail I find most telling is which companies hit the constraint. High performers report being constrained by costs in their use of coding agents about three times as often as others do. They are also twice as likely to be scaling coding agents, 2.7 times as likely to be scaling other agentic AI, and more than three times as likely to be scaling agents in most of the business functions surveyed. The cost figure covers coding agents only, and it is a correlation. My reading of it is that cost shows up as a problem for the companies that got far enough to have one.

Gartner said something similar in June 2025. Its analyst's view was that most agentic projects were proofs of concept "mostly driven by hype", and that this can blind an organization to the real cost and complexity of running agents at scale, stalling projects before production. Two McKinsey authors, in commentary boxes alongside the survey, describe the fix as treating operating costs as a design constraint rather than an afterthought, and note that even as the price per token has fallen, the number of tokens consumed and generated has grown faster.

The pilot that stops here has not been let down by the model. In the cases I have seen, it was costed per month or per seat and then met a per-task reality. If you do not know what one accepted task cost last month, you have not reached this step yet. You will.

The model is deciding things a rule should

The third stop is quieter than the first two. The agent is accepted, the cost is tolerable, and it has been placed where it makes decisions that were never its to make.

My other test had a local model draft support replies for six synthetic cases, one run each. In three, it chose a return policy before it had the information that decides which policy applies. Two of those corrected themselves later in the same output or turned conditional. Two of the three hinged on a field like a receipt date, which a few lines of ordinary code can check, and refuse on when the date is missing.

Six cases say nothing about the market, but they do say something about placement. The model was handed a decision that follows from a date, and in two of the three cases it made the decision without the date. Gartner's release points in a similar direction: "Many use cases positioned as agentic today don't require agentic implementations." Its suggested mapping is agents where a decision is needed, automation for routine workflows and assistants for simple retrieval.

This stop is the one I have no survey data for, only my own experience. In that experience, a pilot that stops here looks healthy. The output is fine most of the time, the cost is fine, and the failures are rare enough to be explained one at a time. It stops because the owner of the process loses trust in it for reasons nobody writes down, and the work is routed around it.

Nobody decided what happens next

The fourth stop has the most public numbers behind it and the weakest. In June 2025 Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." That is a forecast, not a count. The poll in the same release, 3,412 attendees of a Gartner webinar in January 2025, found 19% reporting "significant" investment in agentic AI, 42% conservative, 8% none, and 31% waiting or unsure. Webinar attendees are not a sample of companies, so read it as the mood of a room, not a measurement of a market.

What I take from the three reasons, and this is my reading rather than Gartner's, is that each of them is decidable before the pilot starts and usually is not decided. Escalating costs is the second stop with no cost per task. Unclear business value is the first stop with no acceptance criterion. Inadequate risk controls is the third stop with no list of which decisions belong to the model. A pilot that has none of these written down does not fail. It lingers, and then gets cancelled in a budget round for reasons that sound like all three at once.

The missing piece is an end state. A pilot should carry a condition under which it becomes someone's production process, with a named owner, a budget line and a way to switch it off, and a condition under which it is stopped. McKinsey's survey finds that 80% of respondents say AI has improved their own productivity, while 37% attribute any EBIT impact to it, the same share as last year. Those are two different questions about AI in general, not about agents or pilots. My reading of the gap, and it is only a reading, is that it describes a lot of tools that help the people using them and have not changed how the organization is run. A pilot without an end state is one of those tools.

Which step are you at

Four questions, one per stop. Answer them for one pilot, not for the programme.

The earlier piece had seven questions to ask before an agent goes into a process. These four are for after it is in, and they locate the problem rather than fix it. The fix for each is in the section above, and none of them needs a platform.

If your pilot stopped at one of these steps, or at one I have not listed, I would like to hear which, and what you did next.