The 3 questions I ask myself before giving AI another job

Everyone is creating agents and skills right now. Every week someone posts their new AI operating system, with an agent for every job in the business, and it looks incredible.

What nobody tells you is how you know when it's the right time to build one.

I learned this the expensive way. When I first mapped out my AI team, I mapped 204 AI jobs. It was way too much, and very quickly I realised most of them didn't need to exist. Now I run a small team, and before I give AI another job, I ask myself three questions.

The three questions

1. DO I HATE IT, OR DO I DO IT ON REPEAT?
   Yes → I'm going to try & give it to AI.

2. CAN AI ACTUALLY DO IT?
   Yes → it's a job for AI. No → it stays with me, or it waits.

3. WHAT PART DO I HAND OVER, & WHERE DO I SIGN OFF ON IT?
   AI gets everything up to my sign-off. Nothing goes out without it.

That's all of it. There's no scoring system and no spreadsheet. The rest of this post is how I actually apply each one, with the jobs that passed and the one that didn't.

Question 1 - Do I hate it, or do I do it on repeat?

This is where I start, because it's how I actually started.

The first real AI employee I built was my inbox assistant. I didn't want to read emails anymore. I had several email accounts, I was getting overwhelmed by all of them, and reading every single one was taking a chunk of my day I wasn't getting back. So that was the first job I handed over.

The second one was my morning brief, and then a research assistant. Every one of them was something I didn't want to do anymore.

If I didn't want to do it, I made an assistant for it.

The other half of the question is repetition. Some jobs I don't mind at all, but I do them every day or every week in exactly the same way. Those are just as good to hand over, because the time adds up even when the job doesn't feel heavy.

If the answer to both halves is no, I stop there. A job I enjoy and only do once a quarter isn't worth building anything for, however clever the build would be.

Question 2 - Can AI actually do it?

This is the question most people skip, and it's where most of my 204 jobs fell over.

Wanting to hand something over doesn't mean AI can do it. So before I build anything, I check three things.

Does it happen on a screen? Email, documents, notes, research, data, writing, your calendar, your task list. AI is very good at work that happens on a computer. If the job needs your hands, your face or your voice in the room, it doesn't pass.

Can I explain what good looks like? Either the steps I follow, or two or three examples of the job done well. If I can't put it into words, AI can't copy it.

Can it reach what it needs? The inbox, the calendar, the files, the notes. If the information only lives in my head, I have to write it down first, or the job doesn't pass yet.

One more thing to check here. If the job is the exact same steps every single time, with no reading or deciding in it, it doesn't need AI at all. It needs an automation. When someone comments a keyword on one of my posts, ManyChat sends them the link. There's no thinking in that, so there's no AI in it either.

The one that should have stayed with me

Writing hooks for my content is something I tried to hand to AI, and it should have stayed with me.

AI is really good at a lot of things, but hooks are one I couldn't get it to do well. I know a good hook when I see one, but I couldn't map out why it's good. So I kept getting generic hooks back, and I'd rewrite them anyway.

That's question 2 failing on "can I explain what good looks like?" It isn't that AI can't write hooks. It's that I hadn't taken the time to map out what makes mine work, and AI is only as good as what you can explain to it.

What Anthropic says about this

This question lines up with what Anthropic, the company that makes Claude, teaches its own builders.

In a talk called How We Build Effective Agents at the AI Engineer Summit in 2025, Barry Zhang from Anthropic gave a four-point checklist for deciding whether a job should go to an agent: how complex the task is, whether it's valuable enough to be worth the cost, whether the model can actually do the critical parts of it, and how costly a mistake would be and how easily you'd catch it. If the steps are easy to map out in advance, his advice is to build it as a simple, fixed workflow instead.

Anthropic's written guide, Building effective agents, says the same thing in one line: find "the simplest solution possible", and only add more when it clearly helps.

So my "can AI actually do it?" is the plain-English version of their viability check, and the automation line is their advice to keep it simple when the steps never change.

Question 3 - What part do I hand over, and where do I sign off on it?

AI doesn't get the whole job. It gets the part before my approval.

This is my rule, and it isn't flexible. None of my AI employees can contact a client, send anything or publish anything without me approving it. There are hard rules in place. Nothing goes out into the world, onto the internet or to my clients, without going through me first.

So for every job, I ask where the sign-off goes. If I can put a check in, AI gets everything up to the check. If a mistake would reach a client or the internet before I could catch it, it stays with me.

Here's how that looks in my business.

My inbox assistant reads three inboxes every 30 minutes, from 7am to 9pm. It files everything that isn't from a real person, and for the emails that need a reply, it drafts one. It never sends. Every draft waits for me.

My morning brief runs at 06:06. It reads my calendar, my tasks and where every client is up to, and tells me what today is. It doesn't change anything. It only reads and reports, so the sign-off is me reading it.

My blog posts and scripts are written by AI employees and come to me for review in ClickUp. Nothing is published until I move the task to publish. This post went through exactly that.

That's also where Anthropic's last checkpoint sits. Zhang's point is that when a mistake would be high-stakes and hard to spot, you limit what the agent can do or keep a person in the loop. For a business owner, the person in the loop is you.

How the three work together

Here's one job run through all three, the way I'd do it.

Writing up client meeting notes.

1. Do I hate it, or do it on repeat? Both. Every call ends with notes to file and tasks to create, and I'd put it off.
2. Can AI actually do it? It happens on a screen. I can explain what good looks like: the notes in the client's folder, and the actions as tasks with owners. And it can reach what it needs: the transcript, the client folder and my task manager.
3. What part do I hand over, and where do I sign off? AI files the notes and creates the tasks. Anything that goes back to the client, like a follow-up email, comes to me first.

It passes, so it became an AI employee. If you want to build that one, it's in The routine that turns every client call into tasks.

And here's one that didn't pass: my hooks. Yes to question 1, no to question 2, so it stays with me until I sit down and map out what makes a good one.

What changes when you ask these first

You build fewer things, and the things you build get used.

Most of the 204 jobs I mapped never needed their own employee. Some were automations. Some were things I didn't actually mind doing. Some needed me to write down how I do them before AI could touch them. The ones that passed all three questions are the ones running in my business today.

When a job does pass, the next question is whether it needs its own AI employee or whether it's a skill for one you already have. That's a different test, and it's in Do you need another AI employee?

What this looks like as a business

The question business owners ask me most is some version of "what can AI actually do in my business?" Sitting down with someone, going through their week with these three questions and telling them what to hand over, what to automate and what to keep is a strategy job. You're paid for knowing what to build and what not to build, and the second half is where most people are wasting their money.

Start with one

Write down everything you did last week, in whatever order it comes out. Then take the first job that gets a yes to question 1 and run it through questions 2 and 3.

If it passes, that's your next AI job. If it doesn't, you've just saved yourself a build.

If you want to see every job in your business laid out before you build anything, that's the next step: How to map your AI org chart before you build any AI employees.

If you haven't got your AI brain set up, that's the foundation for all of this: start here.