The AI Your Employees Leave Behind

TL;DR

The person everyone calls when an automation breaks has just given notice.

Their team depends on what they built: an agent that answers customer inquiries, a workflow that moves invoice data between systems, a handful of experiments. In two weeks, the laptop will be collected, the email account closed, and access revoked. The manager knows the team relies on this work. What the manager cannot say is exactly what was built, what each system can reach, or who will be responsible for it next.

But the company has something it rarely gets with a risk: notice.

Use the notice period

Companies already have an offboarding process. Add one question to it: what AI systems or automations did this person build, operate, or own?

Then decide, system by system, before the last day. Keep it with a new owner. Transfer it after a real handoff. Or shut it down, and remove the access it was given, not just its entry in the registry. Do it while the builder is still there to explain the exceptions, the dependencies, and what breaks when it fails. Confirm the handoff before their last day.

This is not really an AI problem. It is an offboarding problem. AI just raises the cost of ignoring it.

Reassignment is not a handoff

The platforms have started to help. Microsoft now lets administrators reassign some agents whose owners have left. Amazon and Google have launched registries to track the agents running in their clouds.

That is progress, but it can create the appearance of a completed handoff. A new name in the owner field does not tell the successor what the system does, what data it touches, or what happens when it goes wrong. It does not answer the most important question: should this system still exist at all? Some were experiments never meant to last. Some contain mistakes nobody has noticed. Some do important work that no one but their builder has ever reviewed.

Knowing that an agent exists is not the same as knowing who is responsible for it. Changing the owner's name in a system is not the same as transferring responsibility.

A registry also sees only what is registered with it. Plenty of automation lives outside every platform: a spreadsheet full of macros in finance, a script a contractor wrote two years ago that is still running. Don't ask whether something is technically an "agent." Ask whether it acts on the company's behalf in a way someone needs to answer for. If it does, it belongs on the list.

The software is looking for a technical definition. The company needs a business definition.

Someone has to own it

This is not a new management problem. Financial processes have owners. Customer relationships have account owners. Regulatory controls have control owners. Why should an AI system be different? If anything, it needs ownership more: an employee notices when a process stops making sense, while an automated system just keeps running.

Ownership is not something a tool can detect. It is a commitment a person makes. Software can record the commitment. It cannot make it.

Every system that can materially affect the business needs a named owner. Not a department. Not "the AI team." Not a distribution list. A person. Accountability spread across a team is accountability held by no one. It also needs a backup, because sole ownership simply schedules the next orphaning for the owner's last day, and a technical contact who knows where the system runs, what it depends on, and how to pause it. One person can hold more than one role, but every role should be named.

Oversight should match what the system can do. An agent that formats meeting notes is not a workflow that approves refunds or talks to customers. Ask what it can touch: sensitive data, customers, money, records, decisions. Then ask what happens if it gets something wrong. For a customer-facing or financial system, I would want monitoring, an audit trail, and regular reviews. A small internal automation needs much less. An experiment needs a date when someone decides whether it continues.

Someone also has to keep the record honest. Periodically, the owner should confirm that it is still accurate, that they still understand the system, and that it should continue. Companies already do this for financial controls: a name, a sign-off, a date. Without it, a registry becomes another database everyone trusts because it exists. A system can have a beautiful dashboard and still contain information nobody has looked at in two years.

A simple test

If I wanted to know whether a company had this under control, I wouldn't ask whether it had an AI governance policy. I wouldn't even ask whether it had an agent registry. I would ask the three questions at the top of this article.

If the answer is no, the company may have bought the software. It may have a registry. It may even have a very good governance policy. But it doesn't yet have accountability.

The fix can start small. Choose one team. List the automations it relies on, what each can do, and what each can access. Name an owner, a backup, and a technical contact for each. Decide what continues, what needs a real handoff, and what should be retired. Then put a date on the next review.

Companies have been quick to put AI to work. That is generally a good thing. But once these systems do real work on the company's behalf, we have to treat them more like employees than software.

Someone needs to know what they do.

Someone needs to know what they can access.

Someone needs to decide whether they should continue.

And, ultimately, someone needs to be responsible when they get it wrong.

The technology can help us find these systems. It cannot take responsibility for them. That part is still up to us.

TAKEAWAYS

Related guides