verifiedagents.ai
All posts

5 min readAI Agents · Cybersecurity

How do you know if your AI agent inventory is complete?

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

Hero: How do you know if your AI agent inventory is complete?

TL;DR: You know an AI agent inventory is complete when four outside sources stop finding agents that aren't on it. Check your list against what leaders say, your network logs, your finance charges, and your sign-in grants. Each agent found off the list is a miss. Zero misses on two checks in a row is a pass.

Having a list of your AI agents is one thing. Knowing the list is right is another.

A list you can't check is a guess with nice formatting.

Why is an AI agent inventory so hard to trust?

Because the agents didn't come through IT. Marketing built one to draft emails. Sales built one to sort leads. Nobody filed a ticket.

These aren't bad actors. They're your most eager people, moving fast. So you're not hunting attackers. You're counting enthusiasm.

And the count runs short almost everywhere. In Gravitee's State of AI Agent Security 2026 report, teams said only 47.1% of their AI agents were actively monitored or secured, on average. The story of how this happens is on MassiveScale.AI, in how many AI agents are already running in your company.

How does the assessment score an inventory?

Question 1 of the free assessment asks how you identify AI agents in your organization. It's part of Identity Management, the first of the five ATF elements. Each answer holds up differently when you check it against outside sources.

Answer

What it says

What the check turns up

A

We have no formal process. We're not sure which agents exist.

No list to test. Every agent you find is new.

B

We track agents informally, in spreadsheets or team knowledge.

Several lists that don't agree with each other.

C

We keep a basic inventory that lists each agent with its purpose and owner.

One list, with misses from teams that never registered.

D

Each agent has a unique ID in one trusted system, with its permissions written down.

Few misses, and each one is easy to trace.

E

Every agent is registered with cryptographic credentials and an ownership chain, plus a record of what it can do.

An unregistered agent can't get a credential, so there's little left to miss.

How do you test whether the list is complete?

Check it against four sources that don't depend on anyone remembering to register.

  1. Ask your leaders, one at a time. Ask each one which AI agents run in their area and what those agents can reach. Do it before they compare notes.

  2. Scan your network logs. Look back 90 days for calls to the big AI providers. A call from a service account is an agent at work.

  3. Read your finance charges. Pull six months of vendor bills. An AI platform charge with no security review is an agent.

  4. Pull your sign-in grants. Every time someone clicked "allow" to connect an AI app, your identity system kept a record. Josh walks through it in testing your inventory against your OAuth consent grants.

Now count the misses. A miss is any agent a source found that isn't on your list.

The pass bar is zero misses, two checks in a row. One clean check could be luck.

What does a complete record look like?

Finding every agent is half the test. Each row has to be complete too. A full row has six fields.

Field

The question it answers on a bad day

Agent name and version

Which one was it?

Named human owner

Who do we call?

What it can read

What did it touch?

What it can do

What could it have changed?

How to shut it off

How fast can we stop it?

Last review date

Is any of this still true?

A spreadsheet works fine. The tool isn't the point. The point is that every agent has a row and every row has all six fields.

The owner field is the one teams skip most. It has its own check, in what evidence proves every AI agent has one accountable owner.

How do you keep the list complete after you pass?

Make signing up easier than sneaking around.

If it takes three weeks to register an agent, people will skip it. I'd skip it too. A short form that takes ten minutes gets filled out.

Then rerun the four-source check every quarter. New agents show up between checks, and the misses tell you which teams the form isn't reaching.

Frequently asked questions

What counts as an AI agent for the list?

Software that decides its own next step. A fixed automation gives the same output for the same input. An agent can choose differently each time, so it needs a full row.

How many misses are normal on the first check?

More than you'd like. Don't treat the number as a failure. Treat it as the first honest count you've had.

How often should we rerun the check?

Every quarter at least. Rerun it after any big rollout of a new AI platform too.

What if a team won't register its agents?

Most refusals are confusion, so start by explaining. The company answers for whatever the agent does, registered or not. That usually ends the argument.

Key takeaways

  • A list you can't check is a guess.

  • Test the list against leaders, network logs, finance charges, and sign-in grants.

  • A miss is any agent a source finds that your list doesn't have.

  • Zero misses on two checks in a row is a pass.

  • Every row needs all six fields.

Question 1 is one of 30 in the free assessment. It takes about ten minutes and scores you on all five ATF elements.

Somewhere in your company, an agent is working right now that nobody wrote down. The test is how you find out how many.

See where your agents stand.

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