AI agent decommissioning is the decision to stop an agent that is already in production, and most companies have no rule for it. They have a rule for starting agents. They often have a rule for stopping pilots. Once an agent is running, it keeps running until it breaks something or someone notices the bill.
A retirement rule fixes that. It names the conditions that end an agent, the person who makes the call, and the step that actually stops the work. Write it before the agent goes live, and it costs one paragraph. Write it after an incident, and it costs a meeting with legal.
Companies stop pilots more readily than production agents
Gartner’s October survey of more than 2,000 CIOs found that 31% have rules for stopping AI pilots, but only 17% have criteria for abandoning low-value AI already in production.1 The gap makes sense. A pilot has a review date built in. A production agent has users, a budget line, and a team that would rather not admit it stopped earning its place.
The pilot article covers the first problem. This one covers the second: the agent that cleared every gate, went live, and is now quietly worse than it was, or costs more than the work is worth, or belongs to someone who left in March.
When companies do look, they find agents worth stopping. In EY’s survey of senior AI leaders at large US public companies, a quarter of those that ran assurance reviews fully stopped at least a quarter of their AI systems as a result, and 29% paused at least that many.2 A review that ends a quarter of what it examines is finding real problems. A company with no retirement rule leaves those agents running.
Four triggers retire an agent
Each trigger is a fact someone can check, so nobody has to win an argument to start the review. Any one of them starts a retirement review. The owner can renew the agent, but has to do it on purpose and on the record.
The first trigger is the owner leaving or changing roles. An agent without a named business owner stops on the owner’s last day unless someone else signs for it. This is the most common way agents go stale, because the work keeps arriving and nobody is reading it. The agent register makes this checkable: filter for rows whose owner is no longer in the role.
The second is a falling accepted-output rate. Every production agent should have a number for how much of its output people use without rework: tickets routed correctly, drafts sent with light edits, summaries nobody corrects. Set the floor when the agent launches. If it falls below the floor for a full month, the agent goes to review. Models change, source data drifts, and the process around the agent moves on; the rate is where all three show up.
The third is cost per accepted job passing its ceiling. The cost-per-task calculation gives each agent a cost per piece of work someone accepted, including the human review. Give it a ceiling, usually what the work cost before the agent. An agent above the ceiling for a quarter gets redesigned, moved to a cheaper model, or stopped.
The fourth is the renewal date. Every row in the register carries a shutdown or renewal date, 90 days by default. When the date arrives with no renewal, the agent stops. This is the trigger that catches everything the other three miss, because it does not depend on anyone measuring anything.
The business owner decides, and the platform owner executes
Retirement fails when the decision and the off switch sit with the same person, or with nobody. Split them.
The business owner decides whether the agent continues, because they answer for its output. If they do not decide by the trigger date, the default is stop, and the steering committee hears about it at the next quarterly read. A default of stop is what makes the rule work. A default of continue is the situation you have now.
The platform owner, usually IT, executes the stop and confirms it. They do not get to veto a retirement on technical grounds, and the business owner does not get to declare an agent stopped because they stopped reading its output.
Write down what actually stops the work
Most retirement plans fail here. Turning off an agent takes several actions, and they differ by product.
Speed matters. In Deloitte’s third-quarter CFO survey, 74% of CFOs at companies with at least $1 billion in revenue said their organizations would typically need several business days to several weeks to pause AI usage that exceeded approved limits.3 Weeks is a long time for an agent that sends customer email or writes to the CRM.
The obvious move, cutting off the person who set it up, often does not stop the work. Anthropic’s documentation for its new Claude Enterprise sign-out control says plainly that signing a member out does not end their API keys or their Claude Code cloud sessions and routines.4 Scheduled work keeps running after its author is locked out. Other platforms differ in the details; the pattern is the same. Agents run on keys, connectors, service accounts, and schedules, and each of those has its own off switch.
So each row in the register needs a stop procedure, written when the agent launches:
- Disable the schedule or routine itself, in the product where it runs.
- Revoke the API keys and service-account credentials it uses.
- Remove the connectors that let it read or write company systems.
- Confirm in the usage logs or the spend dashboard that activity stopped.
The last step is the one that gets skipped. An agent is retired when its usage goes to zero. Test the procedure on one agent each quarter, the same way you test a backup restore.
What to do this quarter
Add two columns to the register if they are not there: the accepted-output floor and the cost ceiling. Make “stop” the default at every renewal date. Ask the platform owner for the written stop procedure for your three agents with the most authority, then run one of them on a test copy and check the logs.
The risk guide treats a kill switch as part of the agent inventory; the retirement rule decides when someone uses it.
The sentence for the board
The board sentence is this: every AI agent we run has a named owner, a quality floor, a cost ceiling, and a renewal date, and when any of those fails, the agent stops by default and we confirm in the logs that it stopped. A company that can say that has a reasonable basis for approving more agents.
Footnotes
-
Gartner, “Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI”, October 19, 2026. Survey of 2,002 CIOs and technology executives. ↩
-
EY, “EY survey finds that autonomous AI implementation outpaces oversight”, September 15, 2026. Survey of 202 US senior AI decision-makers at public companies with at least $1 billion in revenue; the stop and pause figures apply to organizations that ran assurance reviews. ↩
-
Deloitte, CFO Signals, 3Q 2026. Survey of 200 North American CFOs at companies with at least $1 billion in revenue. Reported by TechTarget, October 23, 2026. ↩
-
Anthropic, “Sign Out User”, Claude API reference (beta, Claude Enterprise). ↩