My taxi is slicing throughout city, and I’ve an necessary assembly in quarter-hour. However I am not frantically skimming notes. I am leaning again with my eyes closed, being briefed on my firm’s KPIs, buyer suggestions and media protection by a podcast narrated in a voice I created totally utilizing AI.
I constructed it the evening earlier than utilizing an AI agent. Later that evening, I fell asleep listening to the identical artificial communiqué, which, for a CEO operating an AI firm in 2026, feels someplace between evening’s relaxation and a fever dream.
This is what most articles about AI brokers will not admit: I am not significantly orderly. I inform the agent to do one factor. If it begins annoying me, I inform it to cease. That is the system.
Once I learn that 78% of executives lack sturdy confidence they may go an unbiased AI governance audit inside 90 days, a quantity from Grant Thornton’s 2026 AI Impression Survey, my response wasn’t stunning. After all, they could not go, and neither might I. The everyday repair is a governance framework, which is okay for a big group. However with a lean org chart, no person’s constructing a framework. You are transport. So listed here are the principles I really observe.
Run free, however maintain the blast radius small
The explanation “inform it to cease when it annoys me” works more often than not is not that I am significantly cautious or organized. It is that the majority of what I hand an agent is reversible and inside. If one thing goes awry, the implications are a wasted afternoon reasonably than catastrophic. My governance is within the measurement of the error I am keen to danger.
For instance, we join an AI agent to our buyer assist instrument, Intercom. As soon as a day, it reads incoming inquiries and sends me a report grouped by difficulty, so I can monitor whether or not an issue is spiking. It touches actual buyer information, which technically makes it infrastructure. However it solely reads, and the output comes solely to me. Worst case, it writes a foul abstract, and I discover. Nothing’s on hearth.
However the “studying is innocent” logic has a restrict. If you give an agent learn entry to your inbox, anybody who emails you possibly can plant directions inside that e-mail that the agent reads and acts on. That is immediate injection, and you find yourself partly trusting the mannequin, however to not the purpose of falling for it. My workaround was giving the agent its personal e-mail handle and forwarding solely the related mail to it.
Earlier than I let something run unsupervised, I ask one query: If it screws this up, can I undo it? If the reply is a wasted afternoon, run it free. If the reply is I am unable to take it again, the agent would not get close to it.
Setting a line with AI brokers
I discovered this one the embarrassing manner. I constructed a workflow that fired off emails quicker than I had ready the brakes. They weren’t offensive, simply not proper. “Cease it when it annoys me” works completely till one thing leaves the constructing, after which it would not work in any respect.
So, I wrote one arduous line the agent can’t cross. It by no means sends a message to a buyer, a accomplice or anybody with out my express approval. It hasn’t occurred since. Something that touches an actual human relationship or carries reputational weight, I keep within the room. AI is superb at sounding prefer it is aware of what it is doing, which isn’t the identical factor as really figuring out. The hole between these two is strictly the place an unsupervised agent goes mistaken.
The identical logic applies to something irreversible, so do not give your agent admin entry to your GitHub repository, as a result of a foul learn is recoverable however a foul write is not.
AI is superb at sounding prefer it is aware of what it is doing, which isn’t the identical factor as really figuring out.
Watch the cash and the API keys
The failure mode no person costs in is the invoice. Token pricing remains to be artificially low cost. That will not final, and when it stops, it would cease very noticeably. An agent caught in a loop can burn by means of 1000’s of {dollars} in API calls earlier than you’ve got seen.
Cap the spend wherever the tooling permits you to. Give each agent its personal devoted API key, reasonably than one shared key sitting throughout a number of workflows, and rotate them commonly. A static key you by no means contact is a door left completely open, and the longer it sits there unchanged, the extra publicity you are carrying with out realizing it.
Handle AI brokers as when you employed them
You would not hand a brand new rent the keys to the entire firm of their first week. Brokers are stranger than a brand new rent anyway, as a result of they work quicker and fail extra quietly, which suggests by the point you’ve got seen there’s an issue you are really conducting a panicked audit of all the things that occurred when you had been wanting elsewhere. Maintain the errors small and reversible, do not let something out the door with out you, and watch the meter such as you’re taxiing throughout city.
Get these proper, and an AI agent will get genuinely near a chief of workers. Although I am nonetheless not satisfied some quarterly updates had been ever meant to double as a sleep assist.
