AI agents can do more of the work. What are you still unwilling to let them decide?

I’ve been building AI agents into my own marketing workflow, and the hard part has not been getting them to produce things. They can research a market, synthesize customer calls, draft a campaign brief, write content, analyze results, and propose the next task.

The hard part is deciding where their authority should stop.

A plausible but wrong draft is fixable. A plausible but wrong “fact” that becomes shared company context is much more dangerous. So is an unsupported claim that gets published, a customer message that gets sent, or a campaign change that starts spending money.

The boundary I’m using right now looks like this:

• Agents can read approved context and evidence.

• They can research, draft, analyze, and propose changes.

• Their work keeps its source and version history.

• A person decides what becomes company truth.

• External actions—publishing, sending, deleting, or spending—need approval for the exact action.

This has made me less interested in the question, “How autonomous is your agent?” I’m more interested in: What authority has it earned?

For other founders and makers working with agents:

What is one decision you still refuse to automate, even though the technology could probably do it?

And what would need to become observable, reversible, or provable before you changed your mind?