Last time I ran through the strangest ways I use AI, and I ended on the question I get most often. How do you trust these tools with anything that matters? Fair question, because I do not trust them blindly, and you should not either. What I trust is the system I built around them. Here is what that system looks like.
The first rule is that I control where they are allowed to look. An AI left to roam the open internet will confidently hand you a beautiful, well-written wrong answer. So for anything that actually counts, I restrict the sources. When I ask a health question, the tools must pull from reliable medical sources, not whatever happens to rank well that day. Same logic for anything financial or legal. Narrowing where it looks is most of the battle.
The second rule is that every answer has to show its work. I make the tools cite where each claim came from, with links. No citation, no credibility. This sounds tedious, and it is, a little, but it is the difference between a research assistant and a slot machine.
The third rule is the one people skip: I read the links. Every time. The whole point was never to outsource my judgment; it was to do the homework faster. If I cannot trace an answer back to a source I would have trusted on my own, I throw the answer out. The AI does the legwork. I still make the call.
Then there is my favorite trick for when I want to crowdsource an answer: I make the models argue. I do not rely on one. I will ask the same question of Claude, ChatGPT, Co-Pilot, Gemini & Perplexity, then I tell each one what the other said and ask if it wants to change its answer. Once, standing in a bodega trying to pick the right medication for a friend's headache, I did exactly this. I photographed the options, asked two models, and then compared them. One of them stepped back, looked at the other's reasoning, and changed its recommendation. A model that can hold its position and concede when the other side is right is worth listening to. A model that just digs in is not.
And I want to be honest about the stakes, because this is not abstract for me. This year, I did my taxes with this kind of setup. It was faster and a real help, and I triple-checked every single number, because AI still hallucinates and it hallucinates badly. The guardrails are not me being paranoid. They are the only reason I can move quickly without getting burned.
Here is the part I want you to take away. The people who get hurt by AI are usually the ones who either trust it completely or refuse to touch it. The useful place is in the middle, where you treat it like a brilliant, fast, slightly unreliable intern. You give it good sources, you make it show its work, and then you check that work yourself. Garbage in, garbage out still holds. So put good things in, and verify what comes out.
Next, I will get practical and walk you through the workflows I have actually automated, the boring, repetitive tasks I never have to think about anymore.
For those of you warming up in the comments, yes, I obviously used AI to help me write this. That's my whole point: the ideas are mine, drawn from a five-page free-flowing brain dump and from real conversations I've had with people at all ends of the AI knowledge spectrum. AI helped me organize, tighten, and get the words on the page faster than I could on my own. I have been telling you throughout this series that AI is not here to replace you, it is here to make you more efficient. This article is the proof.