Most owners worry that an AI agent will give a wrong answer. A new survey points at a quieter problem. Even when the agent works fine, the company often cannot say what it did, undo it cleanly, or name who approved it.
That second problem is the one you can do something about this week, and you do not need a developer to start.
What the survey actually says
On June 17, 2026, Kore.ai published results from its 2026 Agent Productivity Index. Kore.ai is a software company in San Mateo, California that sells a platform for building and governing AI agents. So read these numbers the way you would read any survey run by a company that also sells the fix. The survey is a marketing instrument, fielded by a hired firm called Propeller Insights. There is no independent study backing it up yet. I am going to attribute every figure to Kore.ai for that reason.
Who answered
Here is who answered. Kore.ai says it surveyed over 400 IT business leaders, ages 34 to 50, at U.S. organizations with 2,000 or more employees, in May 2026. The stated margin of error is plus or minus 3.0 points at 95% confidence. These are big-company people. You are not running a 2,000-person IT shop. But the pattern they describe scales down to a one-person business faster than you would think.
The headline number, and the four behind it
Kore.ai reports that 72% of enterprises say their AI agents operate with unmanaged risk, including financial and compliance exposure. That is the line in the press release headline. The numbers underneath it tell the real story.
of these enterprises say they have had to reverse an action an AI agent took. The agent acted. A person had to walk it back.
Source: Kore.ai, 2026- 79% say they have had to reverse an action an AI agent took.
- 70% say they ran into a failure their teams could not trace.
- 53% say they are running agents they do not fully trust or understand.
- 42% report lost revenue tied to an agent failure.
Sit with that 79% for a second. Four out of five of these companies, by their own report, had to step in and undo something an agent did. The model is doing what it was told. The failure is authority handed to a tool before anyone defined the rules.
Why the agents had that much power in the first place
Kore.ai also asked what these agents are doing. The answers, again per their survey:
- 41% of agents are running data migrations and system updates.
- 26% are approving or denying decisions.
- 15% are acting on financial transactions.
So roughly one in four agents is saying yes or no to something. About one in seven is touching money. And more than half of the leaders running them admit they do not fully understand what they are running. That combination is the whole problem in one breath. Real authority, low visibility.
What this has to do with your business
This scales down. You do not need a 2,000-person company to recreate the problem. You need exactly one agent doing real work without anyone deciding the rules first. Picture a few small-business versions:
- An agent answers customer emails and quietly promises a refund you never approved.
- An agent updates your booking calendar and double-books a Saturday you blocked off.
- An agent posts to your social accounts, in your name, with a claim about your product that is not true.
None of those require the model to be broken. They require the model to do its job inside a process you never wrote down. That is the same shape as the survey: 79% reversing actions, 70% chasing failures they cannot trace. LaunchReady has one thesis under all of this. AI agents amplify whatever workflow they inherit. So the real work sits in the process you hand the agent.
Point a fast, tireless agent at a messy process and you get the mess faster, at scale, with your name on it.
The cheap version of governance
The survey is selling enterprise software. Peter Mullen, Kore.ai's chief marketing officer, put the pitch this way: "An agent that can be watched but not governed is still a liability." Fair point, and also exactly what a governance vendor would say. You do not need their platform to get most of the benefit. You need three sentences written down before you give an agent any real authority over money, data, or customer decisions.
- What it is allowed to do. Name the task. Draw the edges. Let it draft replies without sending them, or suggest a refund without issuing one.
- How you would undo it. If the agent does the wrong thing, what is your reverse button? If you cannot name one, the agent is not ready for that job yet.
- Who signs off. One human owns the outcome. Not the tool. A person.
That is the small-business version of "governance built in from the start." It costs you ten minutes and a notes app.
Where this fits in the climb
In The 7 Levels of AI Proficiency, the early levels are about using the tools well. The higher levels are about designing the systems those tools run inside, and keeping a human accountable for them. The leaders in this survey skipped a step. They handed agents authority that belongs to a higher level of proficiency before they built the habits that make that authority safe. The reversing and the untraceable failures are what that skipped step looks like in the wild.
You can climb in the right order. Start the agent on a task where the worst case is small. Watch what it does. Write down the rule it should have followed. Then widen its authority, one notch at a time, with a human still holding the pen. That is not slower. The companies in this survey that delayed deployments over governance concerns, 62% of them per Kore.ai, were not being timid. They were doing the cheap thinking up front instead of the expensive cleanup later.
Your next step
Pick the one AI tool or agent already doing work in your business. Today, write the three sentences for it: what it may do, how you would undo it, and who signs off. If you cannot answer the second one, that is your signal to pull its authority back a notch until you can.
Related reading: Level 5: The Captain (Design Thinker).
Sources
- New Kore.ai Survey: 72% of Enterprises Say Their AI Agents Operate With Unmanaged Risk and Create New Operational Burdens
- Kore.ai survey release (BusinessWire distribution)
Frequently Asked Questions
Should I trust these numbers?
Treat them as one company's survey, not settled fact. Kore.ai commissioned the study and sells the solution it describes. The figures are useful as a pattern, not as proof. No independent source has checked them.
Does this mean AI agents are too risky for a small business?
No. It means authority without a rule is risky. An agent doing a bounded task with a clear undo button and a human owner is a different thing than an agent quietly approving decisions nobody is watching.
I am not technical. Can I even govern an agent?
Yes. Governing an agent is mostly deciding what it may touch and who is responsible. Those are owner decisions, not coding decisions. The three sentences above are the start.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.