I don't trust AI with anything.
And honestly, you shouldn't either.
I was playing tennis the other day with another guy who just started dabbling with AI, and he asked me: "do you trust it?"
My response: "he11 no!"
That probably sounded weird coming from me, being pro-AI and all. I run so many of my business workflows through AI now.
But I think having AI do everything for you -- and not trust it at all -- is how you should do things.
LLMs are inherently unpredictable. They are prediction machines that guess the next block of letters.
The model isn't necessarily looking everything up. It's all a bit of a black box. You can ask it the same thing an hour apart and get a completely different result.
The way LLMs work are the opposite of computer scripts. These output the exact same result every time, and give us the ability to generate output from a consistent state. But LLMs are never going to work that way.
This is one of those things where when it clicks, everything starts to make a bit more sense.
And if you accept this, a lot of your hesitation around using AI starts to fade away.
So... you're probably wondering how the heck I work around this limitation.
Well when you stop waiting to trust it, you can start building things around the fact that you never will.
What does this look like? I have complete production AI workflows that carry out entire tasks for me.
This includes an n8n pipeline that handles all of the processing of my course sales and enrollments. It includes an email triage system that I have set up that runs and cleans over my inbox, basically acting like a VA. And it also runs a lesson pre-production pipeline that ensures consistent publishing and automation of translations for my courses.
But every single one of these workflows stops right before the important part.
The email gets drafted... but I send it.
The workflow gets built and wired up... but it sits around as disabled until I flip it on.
The lesson gets prepped... but nothing gets published until I do my own final review of it.
The AI does maybe 95% of the workflow work, ...but the last click is always mine.
You don't need to trust AI to get real value from it. You just need to own the last click.
So yeah, you should be a little fearful of the output. I am. I review everything it produces like it came from that new hire who's insanely fast at doing things, but just started guessing at a few things they didn't know before handing it off to me.
But being afraid of the output and refusing to use it in the first place are two very different things.
The devs sitting on the sidelines waiting for AI to become trustworthy are going to be waiting forever. That day isn't coming. It can't.
But the distrust isn't a reason to avoid AI, but a design constraint that you can use to help build a reliable workflow around a process that works for you.