Retiring or replacing an AI agent? How to switch one off without breaking the process
How to retire or replace an AI agent safely: find every job it does, hand the work over, run a cutover, revoke its access, keep its records, confirm it stopped.

The short answer
To retire or replace an AI agent without breaking the process, list everything it does, including side jobs nobody wrote down, and name who does each task the next day. Run the successor in parallel, drain the open work, then revoke every credential the agent holds, keep its logs and confirm it took no further actions.
Key takeaways
- Retiring an AI agent is a change to a business process, not only an IT task, because people and reports downstream depend on work the agent does without being asked.
- The inventory comes first: every trigger, schedule, system the agent writes to and person who uses its output, including duties that were added after launch and never documented.
- In Gartner's survey of 297 cybersecurity leaders in the second quarter of 2026, 54% said their organisation had no defined approach to limiting AI agent access, or relied on predefined human access.
- A retired agent is off only when every credential, token, scheduled job and inbox rule it used is revoked and its account shows no actions for a full business cycle.
- Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering, so the exit terms belong in the contract before the build starts.
In this article
When should you retire or replace an AI agent?
Retire or replace an AI agent when the work it does has changed, when a better option exists for that work, or when it costs more to supervise than it saves. A dull agent that still does its job well can run for years.
The process changed: a new ERP, a new approval rule or a merged team. The vendor changed its pricing or stopped supporting the product. A newer agent, often one bundled into software you already pay for, now covers the same task. Or the agent's error rate rose until the people checking it spend more time than they used to spend doing the work.
Replacing and retiring share most of the steps. In both cases the old agent stops, and its duties go somewhere: to a person, to a successor agent or to nowhere because the task is no longer needed. The difference is who picks up the work the next morning.
Our view: write the retirement trigger into the agent's runbook on the day it launches. "Review if supervision time exceeds the time saved for two months running" is a rule a team can follow. "We will know when it is time" is not.
Why agents keep running after their work has moved on
Agents keep running because switching one off feels riskier than leaving it alone, and nobody owns the decision. The project team that built it has moved on, the budget sits in a general software line and the agent does not complain.
The number of agents makes this a real housekeeping job. Microsoft said on 29 July 2026 that nearly 40 million agents were registered on its Agent 365 platform, a count of registered agents, not active ones.1 Our note on registered agents versus active ones covers what that gap means.
In KPMG's survey of 314 US leaders at firms with $1 billion or more in revenue (24 July–25 August 2026), 62% were building, deploying or developing AI agents.2 Every one of them will reach an end at some point, and an idle agent still holds credentials and may still run on a schedule.
AI agents in 2026
More agents, and loose control over what they can reach
Others have written the lifecycle; this is the operations side
The lifecycle stages for retiring an agent are already well described, and they are worth reading before you start. AWS's Well-Architected Agentic AI Lens has a best practice on decommissioning unused agents and preventing agent sprawl.4 It covers registering each agent with an owner and purpose, tracking usage, moving agents through stages from active to decommissioned, and reviewing the whole portfolio on a regular cycle.
TrueFoundry's decommissioning playbook, published in August 2026, sets out a step-by-step retirement process covering inventory, redirection, revoking access, keeping records and checking the result.5 TrueFoundry sells an agent platform, and the playbook says the method applies on any stack.
What we add here is the view from the team that owns the process the agent works in. That means the undocumented jobs an agent picks up over time, the people who take the work back, the cutover day, and the contract terms that decide whether you can leave a vendor's agent cleanly.
Find out what the agent actually does before you switch it off
Start with an inventory of everything the agent does today, built from its logs rather than from its original specification. Agents gather duties after launch. Someone adds a weekly summary email, another team starts relying on a field it fills in, and neither change reaches the documentation.
Read the logs, not the brief
Pull at least one full business cycle of the agent's audit log: a month for most processes, a quarter for anything tied to closing the books or seasonal demand. List every distinct action it took, every system it wrote to and every message it sent. Our field list for an agent's audit log shows what a log needs to make this possible.
List the triggers
Write down everything that starts the agent: a schedule, a new email in a shared inbox, a record created in the CRM, a webhook from another system, a person typing a request. A trigger you miss is a trigger that keeps firing after the switch-off, often into a queue nobody reads.
Ask who uses the output
For each action, find the person or report that depends on it. Ask the teams that sit next to the process, not only the owner. The finance analyst whose dashboard reads a status the agent sets will not be in the project file, and will notice first when the field stops updating.
The inventory to build before an agent is switched off
- Actions
Every distinct action from a full business cycle of the audit log, not the original brief.
- Triggers
Schedules, inboxes, record events, webhooks and requests that start the agent.
- Systems
Every system it reads from and writes to, with the account and permission it uses in each.
- Dependants
The people, reports and other automations that rely on what the agent produces.
- Successor
For each action: a person, a new agent, another system, or a decision that the task ends.
Who does the work the day after?
Every row in the inventory needs a named successor before the agent stops: a person, a new agent, another system, or a written decision that the task ends. A blank in that column is a task that will quietly stop being done.
Handing work back to people is the case teams plan worst. Give them the agent's instructions, its test cases and a sample of its recent outputs, so they start from how the work is done today, not how it was done two years ago. If those instructions read like a vague process document, our guide to rewriting an SOP so an agent can follow it works in the other direction too: a clear SOP serves a person as well.
For a successor agent, check that it really covers each action rather than the headline task. A new support agent may answer questions well and still not tag tickets the way the old one did, which breaks the weekly report that reads those tags.
How do you run the cutover without losing work?
Run the old and new set-up side by side, stop new work going to the old agent, let it finish or hand back what it has started, and only then switch it off. A hard stop in the middle of open cases is how invoices go unpaid and customer requests go unanswered.
Run in parallel first
For a replacement, let the successor work the same cases in draft mode for a period while the old agent still acts. Compare the two outputs case by case. Where they differ, decide which one is right and fix the successor before it takes over. One full business cycle of agreement is a sensible bar for routine work.
Freeze and drain
On cutover day, stop routing new work to the old agent first. Then deal with what it has open: cases it has started, drafts waiting for approval and messages it has promised to follow up. Finish them, move them to the successor with their history, or hand them to a person with a note. Keep a list, and do not switch off until it is empty.
Keep a way back for a short time
Disable the old agent rather than deleting it for the first weeks. If the successor fails badly, you can restart the old one in draft mode while you fix the problem. Set a date for removing it fully, or the fallback becomes another agent that nobody owns.
Retiring an AI agent, from decision to final removal
Owner signs off the retirement, the trigger and a cutover date; dependants are told.
The successor works the same cases in draft mode; differences are reviewed and fixed.
New work stops reaching the old agent; open cases are finished, moved or handed to people.
Credentials, tokens, schedules and inbox rules are revoked; the agent is disabled, not deleted.
Its account shows no actions for a full business cycle; dependent reports still update.
The fallback is removed, records move to retention and the inventory row is closed.
Revoke every credential the agent holds
An agent is switched off only when everything it could use to act has been revoked. Closing its chat window or pausing a workflow is not enough if an API key, a token or a scheduled job still works.
List the agent's own service accounts in each system, the API keys issued to it, the OAuth tokens it was granted on a person's behalf, and any shared mailbox or channel it reads. Add webhooks pointing at it, scheduled jobs that call it and inbox rules that forward mail to it. Revoke each one and record the date.
Agents that borrowed a person's access are the hardest case. In Gartner's survey of 297 cybersecurity leaders in the second quarter of 2026 (geography not stated), 54% said their organisation had no defined approach to limiting AI agent access, or relied on predefined human access.3 If the agent ran under a person's login, you cannot revoke its access without affecting that person, and you cannot easily tell its actions from theirs in the change history.
Our view: that is the strongest argument for giving every agent its own account from the start. Our least-privilege checklist for ERP and CRM agents covers how to set those accounts up, and it makes retirement a list of revocations rather than an investigation.

What records must you keep after the agent is gone?
Keep the records that explain what the agent did and why: its audit log, the instructions and model versions it ran, its test set and any approvals people gave it. Those records outlive the agent because questions about its past decisions can arrive months later.
A customer may dispute a refund the agent refused, or an auditor may ask how an invoice was approved. Move the records to the same retention rules as the business records the agent worked on, such as invoices or customer files, rather than deleting them with the agent.
Personal data pulls in the other direction. Under the GDPR's storage limitation principle, personal data should be kept in a form that identifies people for no longer than the purpose needs.7 Keep what your retention policy requires, remove copies the agent made for its own working memory, and ask your data protection lead where the two duties meet.
Leaving a vendor's agent: what the contract must give you back
When the agent belongs to a vendor, your ability to retire or replace it cleanly depends on terms you agreed at the start. If the contract does not give you the logs, instructions and test cases, you may lose them on the day you leave.
Vendor exits are likely to become common. Gartner predicted in September 2026 that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering.6 Our guide to the contract terms for vendor-built agents sets out the clauses in detail.
For retirement, five things matter most. You need an export of the full audit log in a standard format. You need a copy of the instructions, prompts and tool definitions you paid for. You need the test cases and their expected results. You need a notice period long enough for a parallel run. And you need written confirmation that the vendor deleted your data and revoked its access to your systems after exit. Our list of due-diligence questions for agent platforms covers how to ask for these before you sign.
Our view: if a vendor will not commit to exporting the logs and instructions, treat the agent as one you can only switch off, never hand over.
Should you retire the agent, replace it or fix it?
Decide by two questions: is the task still needed, and is the agent still the right way to do it? Those answers separate a retirement from a replacement, and both from a repair that would be cheaper than either.
Our guide to AI agents bundled into software plans covers the common replacement case, where a vendor's included agent now does what a separate one was built for.
Retire, replace or fix: decide by the task and the agent
Task still needed →
Agent still the right tool →
How do you check the agent is really off?
Check that the agent's accounts show no actions for a full business cycle after switch-off, and that every report and process that depended on it still works. Silence in the first week proves little if the agent ran a monthly job.
Look in three places. First, each system's own change history, filtered by the agent's accounts, should be empty after the switch-off date. Second, the logs of whatever hosted the agent should show no runs, no failed attempts and no calls to its tools. Failed attempts matter: they show a trigger you missed that is still firing. Third, the dependants in your inventory should confirm that their reports and queues still fill, now from the successor.
Repeat the check after a month and after the next quarter end. Then close the agent's row in your inventory with the date, the successor and where its records are kept.
Replacing an agent with a newer one: carry the tests across
When a newer agent replaces an old one, its first test set should be the old agent's test set plus every incident and correction from the old agent's life. That history is the most specific evidence you have of where the work goes wrong.
Run the successor against those cases before the parallel run starts, then add every difference found during the parallel run. Our guide to regression-testing an agent before a change covers how to build and score that set.
Carry the supervision across too. The digests, samples and approval rules people built around the old agent were added for reasons, often after a mistake. Start the successor with the same rules and relax them only when its own record earns it.
The switch-off checklist on one page
Retiring or replacing an agent follows the same order every time: decide, inventory, assign successors, run in parallel, drain, revoke, keep records, verify. Skipping the inventory is the most common way to break a process that seemed unrelated.
Decide with the owner and write down the trigger. Build the inventory from a full business cycle of logs, with triggers, systems and dependants. Name a successor for every action, run it in parallel and drain the open work on cutover day. Revoke every account, key, token, schedule and inbox rule, and disable before you delete. Move the logs, instructions and test set into retention, and check the accounts stay silent for a full cycle.
A disclosure: Sigzen AI builds agents for clients, so weigh our view with that in mind. Our AI automation page sets out how we scope that work.
Sources
- Microsoft, FY26 Q4 earnings, 29 Jul 2026: registered agents, not active ones
- KPMG, AI Quarterly Pulse Q3 2026: 314 US leaders at $1bn+ firms, 24 Jul–25 Aug 2026 (Sep 2026)
- Gartner, survey of 297 cybersecurity leaders, Q2 2026; geography not stated (Sep 2026)
- AWS, Well-Architected Agentic AI Lens, AGENTSUS03-BP04: Decommission unused agents and prevent agent sprawl (2026)
- TrueFoundry, AI agents retire too: a decommissioning playbook for enterprise AI (Aug 2026)
- Gartner prediction (Sep 2026): 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering by 2028
- European Union, General Data Protection Regulation (EU) 2016/679, Article 5


