Treat AI like a new hire

The trap nobody warns you about

There are two ways business owners get their first AI worker wrong, and they're opposites.

The first is treating it like a wizard. You dump your entire business on it on day one ... every file, every login, every password ... and expect it to run the place by lunch. It chokes. It doesn't have the memory or the guardrails to hold all of that yet, so it does a mediocre job of everything, and you walk away thinking AI is overhyped.

The second is treating it like a threat. You're so nervous about what it might touch that you give it nothing. One folder, read-only, and you never let it actually do the work. That's safer, but you never get past the demo.

Both mistakes come from the same place: you forgot you already know how to do this. You've onboarded a new employee before.

You're not hiring a wizard. You're hiring an intern with a keyboard.

That line came from an AI that sat in on one of our sessions, and it's the whole shift in one sentence.

You don't hand a new hire the master keys on their first morning. You don't expect them to know where everything is, who every client is, and how you like things done before they've even started. You give them one job. You give them access to what they need and nothing they don't. You watch how they handle it. And as they prove themselves, you widen the keys.

An AI agent is the same. It isn't a genius and it isn't a robot. It's a capable new hire who's fast, eager, a little too confident, and waiting for you to tell it exactly what the job is.

Start with one job, not the crown jewels

Here's the move, and it's the opposite of what most people do. Don't start with the biggest, scariest, most valuable thing. Start with one low-risk job that you do over and over.

Invoicing. Drafting a proposal. Watching your website forms. Writing the Monday update. Pick one box. Get the agent doing that one thing well. Once that box is checked, the next door opens on its own, and the second agent is ten times easier, because you've already learned how to onboard one.

The counterintuitive part: the best agents aren't the smartest ones. They're the most bounded ones. An agent with one clear job, the right data, and clear guardrails beats a vague do-everything agent every time. Narrow wins. Prove value in a small space, then earn more room.

The prompt I actually use to build a first agent

You don't build an agent by knowing all the answers. You build it by letting AI interview you. This is the exact move, and it works even if you've never touched code:

Read what that does. It flips the burden. Instead of you trying to explain a job you've never said out loud, AI pulls it out of you one good question at a time. The multiple-choice-with-rationale part means you're choosing, not composing, so you move fast and you catch the wrong turn before it's taken. And making it confirm the job back to you before building is exactly what you'd do with a real hire: tell me what you think I just asked you to do.

Give it a name and a job description

Name the agent. It sounds small. It isn't.

When you give an agent a name and a one-line job description, you can hold a real team in your head. You know who to ask for what. You get repeatability instead of a pile of one-off chats. I run a handful of named agents, not fifty, because a handful is a team I can actually manage. When a new task shows up, I ask which of them is the right hire, or whether I need to make a new one.

That's the manager's mindset, and it's the real unlock. You're not prompting a tool. You're building a team, and you're the one deciding what each member does and what they're trusted with.

Hand over the keys one at a time

Trust is earned in steps. So is an agent's access. Think of it as a ladder.

1. 1Read-only firstGive it access to one folder and let it look, not touch. Watch what it would do before it does anything.
2. 2Draft, don't sendLet it write the email, the invoice, the post, but you approve before anything leaves the building. Anything with a wide blast radius gets your eyes first. A new hire doesn't email your biggest client unsupervised on day one, and neither should your agent.
3. 3Act on the small stuffOnce it's earned it, let it run the low-risk, repetitive work on its own. That's when it stops being a demo and starts being leverage.

You decide what each rung looks like, the same way you'd decide what a new employee can touch in week one versus month three. The keys are yours. Hand them over on purpose.

The part most people miss

The goal was never one clever agent. It's a team you trust enough to leave alone with the boring work, so you can spend your hours on what your customers actually pay you for.

That only happens if you onboard on purpose. One job. The right access. A name. A trust ladder. Do that with the first one, and you'll do it in your sleep with the fifth.

Second lap: only tell me what I need to decide

Here's what the first lap doesn't warn you about. You onboard one agent properly, it works, so you build another. Then six. And somewhere in there the thing quietly inverts.

Every one of them writes to you. Here's what I did. Here's what I found. Here's a decision I'd like you to make. None of it is wrong, exactly, and all of it lands in the same place at the same time. By Thursday you're spending your morning reading reports about work you delegated so you'd stop having to think about it. That isn't a workforce. It's micromanagement with extra steps, and you built it yourself.

Nobody warns you because the failure looks like diligence. Your agents are being thorough. They're keeping you informed. They're checking with you. The same behavior in a human hire would have you sitting them down in week two.

The fix isn't a settings change, it's a conversation at the top level about how the whole team reports. Why are you sending me this? What would happen if you just decided it? And of everything you've sent me this month, which ones did I ever actually change? That last question does the work, because the answer is almost always none of them. A message you always approve isn't a decision. It's a notification wearing a decision's clothes.

1. 1Things I decideMoney moving, something going out under your name to a customer, anything you can't take back. These stay yours, and there are fewer of them than you think.
2. 2Things they decide and tell me laterThe bulk of it. Give the rule, not the ruling. A hire who has to ask about everything is a hire who was never told what good looks like.
3. 3Things I never need to seeThe work happening as designed. A worker who reports that they did their job is a worker who doesn't trust you to assume they did.

Then set the reporting rule: exceptions and decisions only, batched into one message a day. Not one per agent. One, for the whole team. And name the default, because it's the piece everyone skips ... if you don't answer by a certain hour, the recommended action happens anyway. An agent that freezes while you're on a plane isn't safe, it's fragile.

Run that where your agents are managed, not inside one agent's own session. An agent asked to justify its own emails will defend them.

One more move, once a month and it takes five minutes. Look at the decisions you escalated and find the ones that got approved every single time without a change. Those are promotions waiting to happen. Widen the authority, write down the rule they'll follow, and take them off your desk permanently. That's how a good hire gets promoted, and there's no reason it should work differently here.

Then the part that makes it worth doing. The reports you stop reading are hours, and hours are the thing you were buying. Spend them where a customer can feel it before they've paid you a dollar ... research somebody's business, their market, what their competitors claim that they don't, and hand it over for free. Gifts are reciprocal. Before AI you'd never have done that work for a stranger. Now it costs you a prompt and capacity you were going to let expire anyway.