On Automation

Automation is having its moment. Everyone wants to automate everything with AI, yesterday. And most of them are about to automate the wrong things, badly.

So let's talk about what you should automate, what you shouldn't, and the rules almost nobody follows.

Automate the $10 tasks, not the $100 ones

Start with a simple filter: some tasks have to get done but aren't worth much. Those are your candidates.

Here's a real one. I don't run my company's social media. At all. Our internal AI system generates the posts and the descriptions, and a human just approves them before they go out. Drafting a post and captioning it in Canva is a ten-dollar task. If your time as an owner is worth a hundred bucks an hour, why are you doing the ten-dollar job by hand?

That's the point of automation: get the low-value, must-get-done busywork off your plate so you can spend your hours on the things only you can do.

Only automate what you've done yourself

Here's where I'll lose some people. The best things to automate are the things you've already done by hand.

When we started, I wrote the social posts myself. I learned what landed, what flopped, and where the weird little failure modes were hiding. Then, and only then, I built the automation, because by that point I knew what "good" actually looked like and what could go wrong.

You can't automate a process you don't understand. You'll just build a machine that produces garbage faster. Do it yourself first, learn where the bodies are buried, then hand it to the robot.

Let the machine do 80%. You own the last 20%.

This is the hill I'll die on: you are not removing the human from the loop. Not with AI, not ever.

Picture Henry Ford in the 1920s. The assembly line did the heavy, repetitive work, that was the whole revolution. But Ford still had a foreman walking the floor, eyeballing the cars, making sure they came out right. The line handled the volume. A human guaranteed the quality. A hundred years later nothing's really changed, except the machines got smarter.

So that's the shape you want. Let the AI, the code, whatever you built, do about 80% of the grunt work, the part you don't want to touch. Then a human takes the last 20%: the quality check, the judgment call, the "actually, not that one." A few things you can automate end to end. But most of the valuable stuff still wants human eyes on the finish line, same as it did on Ford's floor.

It will break. That's the plan.

Something nobody tells you: every automation you build will hit bugs. Weird edge cases, unique failures, stuff you never would have predicted. That's not a sign you did it wrong. That's just the process.

When you first flip it on, you're finding out where it breaks. Then you harden it. API throwing errors? Add automatic retries. Something slipping through? Add a check. I tell my clients this constantly: if we're running into a different problem than last week, good, that means we're making progress.

The real unlock is speed

And this is why automating with real code and AI is so powerful: you can test and iterate stupidly fast. That speed is the whole point. The Altmans and Musks of the world are obsessed with iteration speed for a reason, fast iteration is how you make real progress. But that's a whole other post.

So before you automate everything in sight, ask two questions. Is this a ten-dollar task eating hundred-dollar hours? And have I done it myself enough to know what good looks like? If it's yes and yes, automate it, keep a human on the last mile, and get ready to fix the weird stuff. That's how you actually win with this.

Want automation that actually holds up?

Real systems, built by people who did the work first, with a human where it counts. That's what we do at HyppoAI.

Build it right