It's tempting to judge an AI agent by how much it can do without you.
Research a company. Update a CRM. Draft an email. Publish a page. Trigger a workflow. The more it handles, the more impressive it looks.
But the question I keep coming back to isn't just, "Can it do this?" It's, "Should it be allowed to do this on its own?"
Working across product design, websites, CRM and delivery has made me conscious of how decisions travel between systems. A change to a contact record might affect who gets contacted. A campaign update can affect a landing page, the message and the experience that follows. An action can be technically correct in isolation and still be wrong for the wider business.
Autonomy is not an all-or-nothing setting
That's why I don't think autonomy should be an all-or-nothing setting.
If an AI system is gathering information, organising research or preparing recommendations, it can usually have more freedom, provided there's a way to check the results.
If it's about to email a customer, change a live website or update important business records, the consequences are different. That might call for a review step, tighter permissions or a clear way to reverse the action.
And there's another distinction worth making. Undoing a database edit is one thing. Undoing a message that's already reached a customer is another.
But I don't think every little task needs approval either. If something is low-risk and easily reversible, stopping for permission at every step just moves the workload into an approval queue.
The real skill is knowing the difference.
Start with the business process
For me, that means understanding the business process before deciding what to hand over to an agent. Who owns the decision? What happens if it gets something wrong? Will anyone notice? Can the change be undone?
Those aren't just technical questions. They're product, operational and business questions, and experience across those areas changes how you approach them.
AI can take a lot of work off people's plates. But giving it more authority than the task warrants doesn't make the system more useful.
Good automation isn't maximum autonomy. It's the right autonomy for the job.
Where would you draw the line between an AI agent recommending an action and actually taking it?