Before you hand work to AI, decide when it should come and find you

An independent developer hasn't looked at a line of code by hand in about a year; his development work runs through conversations with AI. He juggles four or five projects at once, a machine runs around the clock, and tasks are picked up by AI as they arrive. Once, while he was away, the AI worked through ninety-five tasks in two days. He had to step in perhaps three to five times.

It sounds like the future has arrived. But he spent more of his talk on the other side of it: how the setup runs away from you, and how it gets brought back under control.

The review that never ends

Someone in the audience asked a very concrete question. He had one AI write code and another review it; the reviewer raises issues, the writer fixes them; review again, more issues; fix again… some tasks just kept spinning and wouldn't stop.

The developer said he'd run into exactly the same thing recently. Their fix is nothing magical: after three consecutive failures, you're not allowed to keep going the same way. Escalate — try a more capable model; if that fails too, hand it to a person.

Three is the number they chose, not a law. The real principle is: give failure somewhere to go. Without it, automation spins in place, burning money, looking busy the whole time.

Make sure it knows who to call

The second principle: the AI must have a way to reach a person.

When it hits something it can't resolve, it needs to be able to say: I'm stuck, I need you to decide. And the reverse must hold — a person's instruction has to be able to interrupt it. After three tries, when someone types "stop, try a different approach", that message has to outrank everything currently executing.

Handing over a task is not the same as handing over the power to stop it, redirect it, or take it back.

Never let AI quote freely

Another entrepreneur, who runs an export business on AI, was asked where he still refuses to automate. His clearest answer was pricing.

The reason is plain: AI hallucinates, and it calculates with stale numbers — treating an exchange rate from two years ago as today's, for instance. Hundreds of expert-sounding words followed by one wrong figure, and the margin is gone.

So in his workflow a quote is never generated by one model and sent. One model calculates, a second checks, and then he verifies by hand. The AI prepares and reviews the quote; it doesn't carry the responsibility for it.

Three lines for a small team

Two very different businesses, nearly the same way of managing AI. We've reduced it to three lines that we'd advise any small team to draw before handing over work.

• The failure line. After a few consecutive failures, stop; escalate or call a person. Never let it retry forever.
• The takeover line. A person can interrupt, redirect or take back at any time; the AI has a clear channel to ask for help.
• The responsibility line. Any outward-facing amount, commitment or legal wording — AI may draft and check it, but a person signs it off before it goes out.

Draw the three lines and you can talk about letting go. Let go without them and it isn't automation; it's taking your hands off the wheel.

Right after describing ninety-five tasks in two days, the developer added: don't trust it completely. That's not cold water on AI. It's the other half of the act of handing over — you hand the work over, and you keep the power to stop it. Get that half right and the first half can grow.