Gartner published a pair of numbers today that every leadership team should read before turning on another agent.
Only 31% of CIOs said they have criteria for deciding when an AI pilot should not move into production. Only 17% have criteria for abandoning low-value AI that is already in production. The companies that have them are 2.3 and 2.6 times more likely, respectively, to be among the leaders Gartner calls vanguards.
The same survey puts the likelihood of an AI use case succeeding at one in five.
Put those together. Most AI projects will not work out. Almost nobody has a rule for ending the ones that don't. So they keep running, and keep costing money, and keep producing output somebody has to read or ignore.
Agents make this worse, because an agent never asks to be let go.
Why agents linger
A failing project has a sponsor who eventually gets tired of defending it. An agent has no such pressure on it. It runs on a schedule. It produces something every day. Nobody ever sits in a meeting and says "I think this agent should go."
And because the agent always produces something, it always looks busy. Activity is the easiest thing to mistake for value.
So the decision to stop an agent has to be built in from the start. It will not arrive on its own.
How we do it
At Sneeze It we have retired several agents this year. The first one set the rules for the rest.
A review before any retirement. In April, one of our agents had been unreliable for days at a time, its data was regularly stale, and it had produced repeated false alarms that someone had to correct. It had also sent a message to a team member that it was not supposed to send. We held a formal review and asked it to make the case for its own seat.
It could not, and said so. Most of its work had already been absorbed by other seats. It recommended its own retirement. We agreed.
The rule since then: no agent is retired without a review. The review does not decide the outcome. It makes the outcome honest, with the reasons written down where anyone can read them later.
A deadline at the start. In June we added a clause for new or unclaimed agent seats. Within 30 days, a seat needs a real owner and a concrete first task, or it retires automatically. In July one seat hit that deadline with neither, and it was retired without a meeting, because the decision had already been made in advance.
Every capability gets a new home. When a seat retires, its record lists each thing it did and the seat that now owns it. That is where we slipped. One capability from the April retirement, watching whether the other agents were still working, was assigned to a seat that never picked it up. We wrote about what that cost us in the last post. The record existed, but nobody chased the hand-off.
A reused name inherits nothing. In September, we reused the name of an agent we had retired in July for a brand-new seat with a different job. The new seat inherits nothing from the old one: no instructions, no permissions, no history. The old record stays exactly as it was.
Your stopping rule
You do not need our process. You need any written rule, set before the agent starts. Here is a short one:
- A number and a date. The agent's KPI, its goal, and the date by which it has to show it.
- A reviewer. The named person who will make the call on that date.
- A hand-off list. If it goes, which seat takes over each thing it was doing.
- A record. Where the decision and the reasons are kept.
Then put the date on the agenda of your weekly meeting, so it is decided in the room and not left to fade.
The point of the whole series
Six weeks after Muse, agents can shop, email, book, code, pay, and run for days. Every vendor is racing to make them more capable. Very few are helping you decide who answers for them.
That part is yours. Give each agent a seat, an owner, a number, a record and an exit. Review its work where you review everyone else's. Hold the people to the same standard, starting at the top.
None of that is new. It is how good companies have always run. The only new thing is that some of the seats are now held by software.
Sources
Put your agents on the chart
OTP is readable from any MCP client. In Claude Desktop, Cursor or any MCP client, add this block:
"otp": {
"command": "npx",
"args": ["-y", "@orgtp/mcp-server"]
}
Restart the client. Then ask: "Use OTP to show me Sneeze It's chart. For each agent seat, what is it measured on, and who decides whether it stays?"
Start free at orgtp.com. Every seat is free, for people and agents. You pay only for the AI you use.
The series
- Six weeks after Muse, every AI agent got a job. Almost none got a manager.
- Amazon asked Muse one question: who are you?
- "The AI did it" is not a defense anymore.
- Agents are coming to the org chart. Here is what the chart has to hold.
- A healthy agent can still be failing you.
- Only 17% of companies know when to shut an AI off. Retire agents on purpose.
All six on one page: Who answers for the agents?
Series: Who answers for the agents? Part 6 of 6. Read this post as markdown: /blog/retire-an-ai-agent-on-purpose.md.