Kill your ghost agents.

Somewhere in your organization, an automation is still running that nobody remembers building. It sends emails, moves files or updates records every morning, and no one could tell you who owns it.

You already have ghosts.

A ghost agent is any AI agent, workflow or automation that is still switched on but no longer has an owner. Someone built it to solve a real problem. Then they changed roles, the process changed around it, or the tool it depended on was replaced, and the agent kept going anyway.

Organizations collect them faster than they think. A sales rep wires up a lead-scoring flow. An office manager connects the inbox to a spreadsheet so new intake forms land in a tracker. A contractor leaves behind three scheduled jobs and a shared login. Each one made sense on the day it was built, and each one still looks reasonable if you happen to stumble across it.

The trouble is that nobody stumbles across them. A school district office might have a flow that copies enrollment questions into a shared folder for a coordinator who left in the spring. A clinic might have a reminder job texting patients about a schedule that changed months ago. Both are doing exactly what they were told, which is the problem.

Why they pile up.

Building an agent has never been easier, and that is mostly good news. The tools make starting things effortless, though, and stopping things invisible. There is no reminder when an automation outlives its purpose, and nobody gets credit for switching one off.

Starting an agent takes a minute. Nothing ever reminds you to stop one.

So the count only goes up. Every new agent adds a little more noise to the system, a few more permissions in the wrong hands and one more thing that can quietly break. In our work with teams, the people who build the most automations are usually the most curious and capable people in the building, which is exactly why their old experiments outlive their attention. They have moved on to the next thing.

What a ghost costs you.

The obvious cost is money: subscriptions, usage fees and seats for tools that do nothing useful. The bigger costs are harder to see. A ghost that still has access to your client list is a security gap. A ghost that writes to a shared record can overwrite good data with stale rules, so the invoice tracker shows a vendor's old terms or the proposal template keeps reverting to last year's pricing. And a ghost that sends messages in your name can say things your team would never approve today.

Ghosts also make it harder to trust the agents you do want. When nobody knows which automations are live, every odd result turns into a long hunt through settings pages and old email threads. People start to blame the whole idea of AI for what turns out to be one forgotten flow.

Find them before you judge them.

You can't review what you can't see, so the first job is a search. Start with the admin panels of the tools you already pay for: most automation platforms, email systems and document suites have a page listing connected apps, scheduled jobs and third-party access. Export or screenshot each list.

Then ask people directly. A short message to each team asking "what have you set up that runs on its own?" tends to surface more than the admin panels do, because people remember the spreadsheet script they wrote even when it lives under a personal account. Check the accounts of anyone who left in the last year and any shared logins a contractor used, since that is where most ghosts turn up.

Give every agent a keeper.

The fix is simple and a little boring. Keep one list of every agent and automation in the organization, and put one person's name in a column called keeper. That person doesn't have to have built it. They just have to answer for it.

Each row in the register answers four things. Who keeps it, with a named backup for when they're out. What job it does, in one sentence, along with what starts it and what it produces. What it can reach: which inboxes, records and files it reads, which systems it changes, and anything it sends in your name. And when a person last looked at it, when the next review falls, and how to stop it in under a minute. If a row can't be filled in, that is your first ghost.

Put the review on the calendar every quarter. A spreadsheet is enough; the register only has to exist in one place everyone can find.

Five questions for every agent.

Put each automation through these at the quarterly review. Anything without a good answer to the first two gets switched off without a debate.

1. Who owns this agent today, by name?
2. What job does it do, and does that job still exist?
3. When did a person last check what it produced?
4. What can it read, write or send on our behalf?
5. What would break if we turned it off for a week?

When something fails, switch it off and revoke its access the same day; removing the agent while leaving its login active only creates a quieter ghost. If nobody can answer the fifth question, turn it off for a week and see who notices. Someone usually will within a day if the agent mattered, and that person becomes its keeper. Retiring agents this way keeps room for the ones that matter, and the organizations that get the most from AI tend to be the ones that can say what every agent they run is doing.