verifiedagents.ai
All posts

5 min readAI Agents · Cybersecurity

What evidence proves every AI agent has one accountable owner?

By Michelle Savage, Co-author, Agentic AI + Zero Trust

Hero: What evidence proves every AI agent has one accountable owner?

TL;DR: Proof of ownership is a record with a person's name on it, plus four things attached: a written scope, a shutdown plan, a review trigger, and an activity log. Test it with a spot check. Pick five AI agents at random and try to reach each owner. Five named people who can answer two questions is a pass.

Everyone agrees an AI agent needs an owner. Ask who it is, though, and the answers get vague fast. "The data team." "IT, I think." Neither one is a name.

So skip the debate and ask for evidence.

What counts as proof that an AI agent has an owner?

A row in a record with one person's name and title on it. A team doesn't count. Teams don't answer the phone.

The name needs four things attached to it, or it's just a name.

  1. A written scope. Plain words on what the agent can reach and what it's never allowed near.

  2. A shutdown plan. It says how the agent gets turned off and what breaks when it does.

  3. A review trigger. When the owner leaves or changes jobs, someone reviews the agent.

  4. An activity log. A record of what the agent did, for the day someone needs it.

The case for why it has to be one human is on MassiveScale.AI, in every AI agent needs one accountable human. This post is about checking that you have one.

How does the assessment score ownership?

Question 6 of the free assessment asks who's responsible for each AI agent. It's part of Identity Management, the first of the five ATF elements. Here's what each answer looks like when you go asking for evidence.

Answer

What it says

What you'll find when you look

A

There's no clear ownership defined for agents.

No record. Nobody to call.

B

Agents are generally owned by development or operations teams.

A team name. No person.

C

Each agent has a named technical owner.

Someone who can fix the agent, though maybe nobody who can say whether it should exist.

D

Each agent has both technical and business owners documented.

Two names per agent. One runs it and one answers for it.

E

A full RACI structure that adds a security owner to the other two.

A chart that shows who decides and who does the work.

Answer B is the sneaky one. It feels like ownership. Then something goes wrong at 2 a.m. and "the data team" turns out to be asleep.

How do you run the five-agent spot check?

It takes about an hour. You need your list of agents and a phone.

  1. Pick five agents at random. Don't pick your favorites.

  2. Find each owner's name. Give yourself five minutes per agent.

  3. Reach the person. A real reply counts. An out-of-office doesn't.

  4. Ask what the agent is allowed to do. They should answer in two sentences.

  5. Ask how they'd turn it off. They should know the path, even if someone else built it.

Five for five is a pass. Anything less is your finding. Write down which step failed for each agent, because that tells you what to fix.

If you can't build the list to pick from, that's a different problem. Start with checking whether your AI agent inventory is complete.

What does it look like when nobody owns the agents?

It looks like a very long list with an empty column.

In Josh's client calls, one security leader found around two thousand agents built in Copilot Studio inside his own company. Nobody had claimed a single one. With no owner to call, there was no way to decide which ones should stay.

That's the real cost. Without a name, nobody can make the call. So nobody does, and the list keeps growing.

Should the builder or IT own the agent?

The builder. The person who decided the agent should exist holds the name. IT owns the platform it runs on. Those are different jobs.

Question

The agent's owner

IT and security

Should this agent exist?

Decides, and answers for it

Advise on risk

What can it touch?

Writes the scope in plain words

Enforce the scope

Is it behaving?

Gets the call when it isn't

Run the watching

How does it stop?

Signs the shutdown plan

Build the shutdown path

Owners don't need to be engineers. They need to know what the agent is for and what it's allowed to touch. Checking that the limits hold is a separate test, covered in testing whether a new AI agent stays inside intern-level limits.

Frequently asked questions

Can a team own an AI agent?

A team can run it. A team can't own it. Name one person and let them hand out the work.

What if the owner doesn't understand the technology?

That's normal. The owner decides what the agent is for. Engineering decides how it works. If your owner has to understand how the model thinks, you've written a job nobody can fill.

What happens when the owner leaves the company?

Nothing, unless you built the trigger. Your identity team already has a process for people who leave. Ask them to add agents to it.

Do we need a tool to track ownership?

Not to start. A spreadsheet with the agent, the owner, the scope, and the shutdown plan beats an empty platform.

Key takeaways

  • Proof of ownership is one person's name and title in a record.

  • The name needs a scope, a shutdown plan, a review trigger, and a log attached.

  • A team name is the same as no owner.

  • The spot check is five random agents and two questions each.

  • Five for five is a pass.

Question 6 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.

Most teams already know who should own their agents. They just haven't written it down, and a name nobody wrote down won't pick up the phone.

See where your agents stand.

The free assessment takes ten minutes and scores you on the five elements of the Agentic Trust Framework.