The fastest way to be disappointed by AI is to treat it like an oracle — a box you ask a question and trust the answer. The fastest way to get real value is to treat it like an intern: bright, fast, eager, knows a little about everything, and will hand you confident, polished work that's sometimes flat wrong. You wouldn't forward an intern's first draft to a customer without reading it. Same rule here.

Once that clicks, two things follow. You stop expecting magic from a one-line question, and you start managing — which is where the value actually lives.

AI works by predicting the most likely next words. That has two consequences worth burning into memory.

First, its default answer is the average — the safe, middle-of-the-road take that offends no one and excites no one. Clean and confident reads as quality to a newcomer, but tidy isn't the same as right or useful.

Second, it will state a wrong fact, a fake date, or an invented number in exactly the same confident tone it uses for true ones. The polish is not evidence. This is the part that burns people: they assume confident means correct. It doesn't.

Neither flaw is a dealbreaker. They're just the two things a good manager compensates for.

Here's how I get genuinely useful output instead of average mush.

Give it the context, not just the question. An intern who knows your business beats a stranger every time. Hand it the relevant file — your standards, the actual document, last year's version — before you ask. (More on building those files in "Build the context first.")

Be specific, and show an example. Vague in, vague out. "Make this better" gets you nothing; "tighten this to under 150 words, plain language, keep the warranty line exactly as written" gets you something. Better yet, paste a sample of what good looks like and say "match this."

Make it interview you. If you're not sure how to frame the request, don't fake it — say "ask me what you need to know before you start." Let it pull the details out of your head instead of you guessing at the perfect prompt.

Don't stop at the first answer. The first reply is a rough draft, not the deliverable. Push: "argue the other side," "what did you leave out," "what's the weakest part of this." The good material usually shows up after a few rounds, not in the first one.

For anything that matters — a number, a claim, a quote, a spec — assume it might be invented until you've checked. Two habits make this easy.

Ground it in your own documents. Instead of asking from thin air, upload the real source — the policy, the contract, the manual — and tell it: "Use only what I gave you, and start each point with 'According to…' so I can trace it." This alone eliminates most of the made-up answers.

Keep a human in the loop on anything that leaves the building. Customer-facing, money, safety, legal — you read it before it ships. Use AI to get to a strong draft in a tenth of the time, then apply the judgment only you have.

There's a related trap worth naming: don't aim AI at the stuff you can't judge. If you don't understand the topic, you can't catch the mistakes, and you'll confidently approve garbage. Point it at your areas of real expertise, where your trained eye spots what's wrong in seconds. That's where it makes you faster without making you reckless.

Treat it like a sharp intern and the whole relationship makes sense: give clear direction, expect a confident first draft, check the work, and own the result. Do that and AI stops being a slot machine you hope pays out, and becomes what it actually is — the fastest junior teammate you've ever had, working under a boss who knows better than to ship the first draft.

Got an AI answer that was confidently wrong? Hit reply and tell me what happened — those stories are gold, and I read every one.