96% Say They Know Every AI Agent They Run. Two Thirds Have Had One Cause a Problem.

Almost every IT leader in a new survey said they have a complete list of their AI agents. Two thirds of companies running agents also had one cause a problem in the past year. The gap between those two numbers is where the risk sits.

•5 min read
Chris Fitkin
By Chris Fitkin•Partner & Co-Founder

Ask an IT leader whether they know which AI agents are running in their company, and the answer is almost always yes. A survey published in September put a number on that confidence: 96.4% of IT decision makers said their organization has a complete and accurate inventory of its AI agents.

The same survey found that 66.7% of organizations running agents had an agent-related operational problem in the past 12 months.

Both numbers can be true at once. The list is complete for the agents somebody remembered to put on it. The problems tend to come from the rest.

What the survey found

The research was run by Morning Consult and published by Guild AI, a company that sells agent management software. Morning Consult interviewed 362 IT decision makers at US companies with 100 or more employees between August 4 and 9, 2026. The margin of error is plus or minus 5 percentage points.

Finding Share of respondents
Confident their AI agent inventory is complete and accurate 96.4%
Had an agent-related operational problem in the past 12 months (companies with agents) 66.7%
Think it’s likely employees are deploying agents without formal approval 60.6%
Have a central dashboard or monitoring tool for agents 42.7%
Have logging or audit trails for agents 39.8%
Require agents to be registered before they go live 35.9%
Can stop a malfunctioning agent with an automated kill switch 31%

The rows lower in the table explain the top two. Fewer than four in ten companies require an agent to be registered before it goes live, and six in ten think employees are already deploying agents without approval. An inventory built that way only covers the agents that went through the front door.

A second survey points the same way

In April, the Cloud Security Alliance published a survey of 445 IT and security professionals, fielded in late 2025 and commissioned by Zenity, an agent security company. It found that:

  • 53% of organizations have had AI agents go beyond the permissions they were meant to have.
  • 47% had a security incident involving an AI agent in the past year.
  • 54% reported between 1 and 100 unsanctioned AI agents.
  • Only 15% said that three quarters or more of their agents have a defined owner.

Different sample, different sponsor, a year apart, and the same picture. Agents are spreading faster than the lists, owners and permissions meant to keep track of them.

Why the list matters

An agent is software that takes actions on its own: it reads email, updates records, sends messages and moves data between systems. When it does something wrong, three questions come up right away. Which agent was it? Who owns it? Who can stop it?

If the agent isn’t on a list, none of those questions has a quick answer. We wrote about what that delay looks like in The Alarm Worked. Nobody Was Sure Who Could Stop It., and about a case in our own production where the stop came before anyone knew in The Stop Came First. We Found Out Four Hours Later.

The fix starts with a rule, before any new software: no agent touches live company data until it’s written down, with an owner, a list of what it’s allowed to do and a way to turn it off.

How we keep the list for the agents we build

We hold our own agents to this rule first. Metacto’s own revenue, finance and delivery teams work with the same agents we build for clients, and every team has a named person accountable for what its agents do.

For clients, every agent goes on the list before it touches live data, with four things written down:

  1. A named owner. Someone on the client’s team answers for what the agent does, and a named Metacto lead runs it day to day.
  2. The systems it can touch, and what it’s allowed to do in each, set to the minimum the job needs. The agents run in the client’s own cloud.
  3. A person approving what matters. Until an agent earns more autonomy, people approve every consequential step. The rule for earning it is written down too: 20 decisions, 90% agreement with the reviewer, and no rejections in 14 days.
  4. A record of what it did and a way to stop it, so any action can be traced back to the agent and the request behind it.

Each item takes minutes to set up before launch and much longer to piece together after an agent has caused a problem.

A ten-minute check you can run this week

You don’t need a project to find out where you stand. Get the people who run your systems in a room and do two things.

  1. List every AI tool that touches live company data. Include agents your teams built, agents that came inside software you already pay for, and anything an employee connected on their own.
  2. Ask which of them has done something you had to explain to somebody. A wrong email, a changed record, a report with numbers nobody could trace.

If the second list has anything on it that the first list missed, you’ve found your inventory gap.

Five questions to ask about the agents you run

  1. Is there one list of every agent touching company data, and who keeps it?
  2. Does every agent on it have a named owner?
  3. What is each agent allowed to do, and who approved that?
  4. Can you see what an agent did last Tuesday?
  5. If one of them misbehaves right now, who can stop it, and how long would it take?

If there’s recurring work you’d like an agent to take on, with these controls in place from day one, show us the work. Describe it in about three minutes and we’ll send back a plan for your first agent: what it would do, what we’d measure, and how it goes live. If you’d rather talk it through, talk to an expert.

Sources

Share this article

LinkedIn
Chris Fitkin

Chris Fitkin

Partner & Co-Founder

Chris Fitkin is a Partner and Co-Founder at Metacto, where he leads the firm's Operational AI practice. He works with private equity sponsors and operating teams to find the workflows worth funding, build the business case, and ship governed AI systems that create measurable value. His background spans engineering leadership, internal operations automation, and technical due diligence, including sell-side diligence for a mid-nine-figure private equity transaction.

View full profile

Find the first piece of work worth giving back to your team.

45 minutes with our AI experts to find your highest-value opportunity and a practical next step. No preparation required.

We'll use your details to respond to your inquiry. See our Privacy Policy.

45 minutes
No prep required
A practical next step

Subscribe to our newsletter

Be the first to get insights on putting AI to work, engineering quality, and building systems that move real business metrics.

By subscribing you agree to ourPrivacy Policy.