I’ve been wrestling with a question lately as I build more AI agents into my own business.

How do I know when an AI agent is ready to work without me?

I can give an agent access to my calendar, email, business systems, and even the ability to execute tasks. The technology is getting remarkably good at figuring out what needs to happen next.

But just because an agent can do something doesn’t mean I’m comfortable letting it.

I started thinking about how we’ve handled this problem throughout history. Surely we’re not the first generation to figure out how to give someone or something more independence.

And there’s a pretty good model sitting right in front of us.

Think about how someone learns to fly.

They start by studying the rules, understanding the equipment, and watching an instructor. Eventually, they take the controls with someone experienced sitting beside them.

Over time, the instructor intervenes less. The student demonstrates competency, handles unexpected situations, and eventually earns the privilege of flying solo.

Even then, they don’t have unlimited authority. They’re certified for specific aircraft, conditions, and operations. More freedom requires more demonstrated capability.

Why wouldn’t we approach AI agents the same way?

I’ve started thinking about AI autonomy as something that should be earned, rather than something we enable with a checkbox.

Back in 1978, researchers Thomas Sheridan and William Verplank developed a framework describing different levels of human and machine control.

At one end, the human makes every decision. At the other, the machine operates entirely independently.

Between those extremes are several interesting arrangements. A computer might recommend an action, execute it only after approval, or complete the work and simply report what happened.

Nearly 50 years later, we’re confronting the same question with AI agents.

The technology has changed dramatically. The fundamental problem hasn’t.

How much authority should we delegate, and what evidence should we require before increasing it?

Imagine introducing an AI agent into your business with a progression that looks something like this.

Notice something important here. We’re authorizing the agent to perform specific jobs, rather than declaring the agent generally trustworthy.

An agent might be ready to manage my calendar independently while still requiring approval to send an email to a client.

This is where I think we need to become more disciplined.

Before giving an agent more independence, I’d want answers to five questions.

I’d also want the agent tested against situations it hasn’t encountered before. Repeating a familiar task successfully isn’t the same as exercising good judgment when circumstances change.

I suspect we’re going to make two mistakes with AI autonomy.

Some businesses will keep their agents permanently supervised, approving every little action until the technology becomes more work than it’s worth.

Others will give agents broad authority because they performed well in a handful of demonstrations.

Neither approach makes much sense to me.

I’m leaning toward a system where autonomy expands as an agent demonstrates reliability, but authority remains limited by the consequences of getting something wrong.

And just like a pilot, an agent should be able to lose its privileges when performance deteriorates.

Here’s something worth trying this week.

Pick one business process you’d like an AI agent to own. Write down what it would need to demonstrate before you’d let it operate without approval.

If you can’t define that standard, you’re probably not ready to delegate the responsibility.

I think the businesses that figure this out will have a significant advantage. They’ll be able to give AI meaningful work without gambling on whether it can be trusted.

And perhaps that’s the real question we should be asking.

Instead of asking when AI will be smart enough to take over, maybe we should be asking when we’ll be disciplined enough to let it?

Until next week,

—Jared

Text Me: 314.806.3912

The professionals pulling ahead aren't working more. They're using Claude.

Our free guide will show you how to:

Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.

Get The Free Guide