It will happen in a meeting you didn’t prepare for. Someone more senior than you will turn and ask, more or less, “what do you make of all this AI stuff?” — and you will have about four seconds.

There are three answers available. Two of them are safe and forgettable. The third is the one worth being able to give.

The two forgettable answers

“It’s going to change everything.” True in the way that horoscopes are true. It commits to nothing, so it tells the room nothing, and it’s what everyone else says. The person who asked already knew it was going to change things — that’s why they asked.

“Honestly, I think it’s overblown.” This one feels safer because scepticism reads as rigour. It isn’t, quite. It ages badly, and more importantly it signals that you haven’t looked. Scepticism is only impressive when it’s specific.

Both answers have the same problem: they’re about AI. The person asking doesn’t actually want your view on AI. They want to know whether you’ve thought about your work.

The third answer

The third answer is bounded, specific, and slightly boring:

“For the thing I do, it’s genuinely useful for X. It’s unreliable for Y, and here’s how I check.”

That’s it. Not a prediction. Not a position. A report from someone who tried it.

It works because it’s falsifiable. Anyone can say “it’ll change everything.” Almost nobody in the room can say “I used it for this, it saved me a day, and here’s the specific way it was wrong the first time.” That sentence is rare enough to be memorable, and it’s the one that gets you invited to the next conversation.

Building the answer before you need it

You don’t need to understand how the models are built. You need three things, and you can get them in an afternoon:

One task where it clearly helped. Not a hypothetical. Something you actually did, where you can say roughly how much time it saved and why the output was good enough.

One task where it clearly failed. This is the more valuable half, and most people skip it. Try it on something in your own domain — something where you’d immediately notice a wrong answer. Watch what happens when it doesn’t know. The failure is rarely a blank; it’s usually something fluent, well-formatted, and wrong. Once you’ve seen that once, you’ll never fully trust the format again, which is exactly the instinct you want.

A check you’d actually run. “I verify every reference against the source” is a real answer. “I only use it where I’d catch the error myself” is a real answer. “I’d trust it” is not.

Why this matters more than it should

Being good at your job and being known to be good at your job are two different skills, and the second one is not automatic.

What’s changed is the cost of not having it. When the visible output of competent work gets cheaper to produce — the clean draft, the tidy summary, the well-organised deck — being the person who produces it quickly stops being evidence of anything. What’s left as evidence is judgment: knowing which question to ask, knowing an answer is wrong even when it reads well, knowing which constraint exists for a reason.

That judgment is the scarce thing now. The meeting where someone asks you about AI is, whether or not they intend it, a test of whether you have any.

It’s worth being ready for.

If you’d rather not start from zero, The AI Primer is the plain-English version of all of this — what these systems are, where they fail, and the vocabulary to follow the conversation. $23.97.