Imagine you run a small cat grooming business. You’d happily spend every hour brushing, shaving, and shampooing fur, but there’s also a business to run. Staff to pay. Leases to renew. Taxes to file. The list of worries gnawing at the back of your mind is endless. You want to focus on clients, but the business can’t be a black box.
So when you’re introduced to an AI assistant named Shirley — a six-foot gleaming slab of aluminum that promises to quietly handle the administrative load — it sounds like a dream. You set Shirley up in the back office with a computer, some sticky notes, and a coffee machine. Shirley gives a confident nod, tucks into a pile of invoices, and you close the door, relieved.
For a while, it’s great. You can hear aggressive typing from the next room — bam! bam! bam! — and a steaming mug mysteriously appears on your desk every so often. What a purchase.
Then, a few weeks in, things start to go sideways. Customers get double-booked. Your landlord mentions a late lease payment. Five thousand bottles of shampoo arrive at your door.
This isn’t just a hypothetical. Funny in retrospect, alarming in the moment. Both stories point to the same risk: giving AI too much autonomy, too soon.
Running a business is an art as well as a science. When you offload decisions without fully understanding what the AI is doing or why, you don’t eliminate the black box. You just move it. Everything looks fine until it isn’t — and by then, you’ve lost the context to course-correct.
The Framework
This is what led me to a framework I call Earned Autonomy.
The idea is simple: rather than handing AI the keys and hoping for the best, make it earn your trust incrementally — and only delegate what you genuinely understand yourself.
Instead of letting Shirley loose on your supply orders from day one, sit down with it first. Work through your expenses together. Understand where the shortfalls are and what problems are worth solving. Maybe Shirley spots a cheaper vendor for combs and brushes and can show you exactly how much you’d save over a year. Maybe it flags which services have become unprofitable and helps you rethink your offering. That’s genuinely useful — and you’re still in the loop.
AI should remove friction as a partner, not as a proxy — and it should help you understand why the friction existed in the first place.
The trap is the relief. Relief is seductive. Relief is also, very often, the first symptom of having given away the part of the job you most needed to keep.
For the Builders
What does this mean for those of us building AI into software?
The instinct is often to make AI do as much as possible, as automatically as possible. But there’s a better approach — one that builds user confidence alongside capability.
Take a personal finance app. Instead of having AI silently tag all of a user’s expenses in the background, surface changes before they’re executed. Let users see what’s happening, and why. The AI notices a laptop logged as a regular expense could be classified as a depreciating asset. Rather than reclassifying it unilaterally, it could say:
“I noticed this. Here’s what I’d suggest, and here’s why. Want me to make this change?”
At first, the user might adjust the action. Over time, as they see good results, they’ll naturally grant more autonomy. Trust compounds.
This matters especially when introducing AI into an existing product. A sudden paradigm shift raises expectations fast, and the higher the initial autonomy, the steeper the cost when something goes wrong. Easing AI in through small, transparent, approvable tasks is a much more durable path to adoption.
In Closing
So keep Shirley. The kids love giving it high fives, and it’s very good at data entry. But Shirley is a hire, not a silver bullet.
Invest in some quality time upfront. Help it understand what you need. Make sure you understand what it is doing. And hold onto your authority over the decisions that matter.
That way, when something does go sideways, you won’t wake up to a pile of cat toothbrushes and no idea how they got there.