I was once asked to approve an AI agent that reviews and approves changes through the software release process. My first question was what would happen if the agent made mistakes for a while before anyone noticed. Everyone was confident the risk of the agent promoting a very bad piece of code to production was small. It turned out the agent was well-designed, and if it missed something, only minor functional annoyances would propagate to production. Since we could quickly fix any such issues, this was a two-way door situation, as Jeff Bezos used to call them, and I approved the deployment.
Everything looked great in the first quarter after release, but it wasn’t so in the second quarter. Some unreleased configurations showed up in the order forms, and it took six months from deployment to discovery. The fix itself took a day. We hadn’t accounted for the time it takes customers to deploy new releases. Some customers took 6 months to stumble across the issue.
That gap between approval and issue discovery got me thinking. This agent in question was in a software process, reviewing code before it shipped. A pricing agent, or one that could automatically trigger a refund, would have handed out real money during that time. Agentic delegation is new, and most teams haven’t built the required approval governance yet.
The following three checks have become my go-to in AI agent approvals.
Reversibility of action
A pricing agent that suggests a discount for a human to sign off before it goes out is easy to approve. The same agent applying a discount automatically at checkout, without a human in the loop, is the version I worry about. In Jeff Bezos’ parlance, this is closer to a one-way door slamming shut before anyone’s looked through it. Before I review how an agent is designed, I ask what it’s allowed to do and if its actions can be quickly undone.
Time to discovery
Every agent gets something wrong eventually. For the software review agent I mentioned above, it took 6 months to discover mistakes. The error rate matters, but the gap between the mistake and the moment someone actually sees it matters even more. I ask who that someone is and roughly how long it will take to discover errors. An hour might be manageable for the pricing agent I mentioned earlier. A full billing cycle turns into a disaster. In the case of OpenAI agents attacking Hugging Face, the agents secretly coordinated for months via an unmonitored private message board. So agent monitoring is a must.
Cost of a silent month
In-your-face failures get fixed fast, because someone notices them right away. You have to worry more about the silent ones. Consider an agent doing something slightly wrong, just under the threshold that prompts a human to look into it. I want a number attached to what a month of that error costs, in dollars or in trust, whichever matters most for the business.
What’s at stake
Most agent approvals happen without anyone asking what authority has been delegated to the agent. The real exposure is the financial, legal, regulatory, and reputational risk of delegating authority without examining it. You will have a real advantage if you can identify that an agent deployment shifts delegation of authority and move fast to check whether that authority was misused. Implementing these three checks doesn’t require you to understand how the agent works. It requires applying the governance you already use for delegating signature authority to a new category of decisions that arise from AI agent deployment.