What should an AI agent never do on its own?
Two questions decide it. Does the job need judgment, and how big a deal is it if it goes wrong.
People worry that an AI agent will do something on its own they'd never have agreed to. Fair enough. You don't want to check every single thing it does, that's just a slow way of doing the work yourself. And you don't want it loose on the things that matter. So you need a simple way to tell those two apart. Two questions do it.
Does the job need judgment, or is it new?
Anything that needs judgment needs a person. Anything being done for the first time needs a person. But a job we've done hundreds of times, and got right hundreds of times, can run on its own. The more often the agent does something and gets it right, the less it needs watching. It has to get a job right over and over, with someone checking, before it does that job on its own. It doesn't get trusted on day one.
How big a deal is it if it goes wrong?
Sending a few emails to a list is not a big deal, so the agent just does it. Putting something live in front of millions of people, or spending a big budget, is a big deal, so a person checks it first. It isn't about whether it's an email or a web page. It's about the size of the consequences if it's wrong. Small consequences, let it run. Big consequences, a person looks first.
The rule
An agent runs the jobs it has done many times where a mistake would be no big deal. A person stays in the loop for anything that needs judgment, anything new, and anything where a mistake would cost you.
It's built in
This isn't a policy we ask you to take on trust. It's how the agents are built. The agent runs the routine work on its own, and stops before anything with real consequences: something public, something that spends serious money, something it hasn't done before. The bigger the deal, the surer we are that a person has looked.
Where this leaves you
The agent does the work. You keep the decisions that would cost you something. Get in touch to see where those lines sit on your own campaigns.
Questions people ask
Can an AI agent send emails to my customers on its own?
For routine sends it has done many times, yes. A mistake there is no big deal, so it just does it. For anything with bigger consequences, a large campaign, a big spend, or something new, a person checks it first.
What decides whether the agent needs a person?
Two things. Whether the job needs judgment or is being done for the first time, and how big a deal it would be if it went wrong. Routine, low-stakes jobs run on their own. Judgment calls, new things, and anything with real consequences get a person.
Does it do more on its own over time?
Yes. It has to get a job right many times, with someone checking, before it runs that job on its own. It doesn't get trusted on day one.
What stops it from doing something with big consequences?
The bigger the consequence, the surer we are a person has looked before it happens. Public things and serious spends always get a person, and the rules about who can be contacted, and how, are written into the code.