The New Hire Nobody Onboarded

The room is full of SMEs, subject matter experts, or at least people who believe they are, which is to say it is full of people paid to find the flaw. Someone is walking them through a new tool with an AI engine under the hood: a demo, some slides, a few confident sentences about how this changes everything.

Then one of them asks a question. Just one.

The tool answers. The process it describes is correct. The mathematical equation it recommends is not wrong, exactly. It’s just not the best one for the job.

And the room pounces. You can feel it happen. There it is. It’s wrong. It doesn’t even know. If we’d used that, we’d have been off. Within about ninety seconds, an entire tool has been tried, convicted, and sentenced to irrelevance.

I’ll admit I was sitting there fighting a smile. I have watched experts make exactly that kind of call, plenty of times. So have you. I’ve watched very smart, very public people make errors in the same neighborhood and keep their jobs, their reputations, and their parking spaces. Nobody called a meeting to declare that the human race could not be trusted with equations.

One slip from a person is called being human. One slip from a machine is called proof.

The second try

Here’s the part I love. Somebody ran it again.

This time, the tool was opened up to look more widely. And it came back with something a lot more interesting than a single answer. It found three different calculations. It noted that each one used different parameters. And then it did what the best experts do: it asked which one fit their scenario.

The tool wasn’t dumb. It had been boxed in. The first answer was only as good as the information it was handed, and the information it was handed had a gap in it. Take the box off, and it started behaving like a sharp colleague who says, “Well, it depends. What are you actually trying to do?”

That is not a machine that doesn’t know. That is a machine that finally had enough to ask a good question.

A footnote with a dead DOI

I have a smaller story in the same genre, and it’s my favorite kind: a detective story with a very boring crime scene.

When I research a topic with AI, I ask for sources and I ask for citations, MLA style, the whole formal outfit. Then I check every single link, because I’ve learned the hard way that “looks like a source” and “is a source” are different things.

On one occasion, a citation looked perfect. Author, title, journal, and a DOI, the little permanent address academics give their papers. I clicked it. Nothing. A dead end, in a very official-looking suit.

I dug a bit, though not all the way to the bottom. And here’s what I found, as far as I could tell: the AI may not have invented it at all. Someone, somewhere, had seemingly created that bogus citation, DOI and all, and put it out into the world, and the AI picked it up like anything else on the internet. I never ran it fully to ground, so I can’t swear to it. But it changed how I think about the problem. Fake information has always been out there. It just used to be harder to find, and now it’s faster to repeat.

I never used a word of it. That’s the point.

I’ve also learned to set rules before I start. I tell the tool plainly: I need a correct answer. If you’re not sure, tell me you’re not sure. It’s the same speech I’d give a new hire on day one, and it works for the same reasons. Does it take a little extra time? Sure. Won’t I always have time? Honestly, no. But checking is a skill, it’s easy to learn, and it’s the same old habit of asking why.

It’s the same speech I’d give a new hire on day one.

The dispatch office

Now the big one.

A logistics company brought in an AI tool for dispatching, the daily puzzle of which driver goes where, with what, and when. By every measure I could see, it was a good tool. Capable. Friendly. Well built. It even came with a digital dispatcher, a kind of avatar, which the team did not rush to embrace.

A friend inside the organization asked me to take a look from the leadership side, because something odd was happening. Four months in, the data said the tool was barely being used. A fraction of what it could do.

So we went looking for why.

The tool, it turned out, had been built around the process as someone understood it. Not the process as it actually happened. There’s a big difference between the two, and every organization on earth has a gap between them. On paper, dispatch was one thing. On the floor, it was another.

The most senior dispatcher, for example, understood something no manual had ever captured: different drivers do better in different situations. One driver simply doesn’t like driving at night. So, ideally, you put him on day work. Another driver was better at something else entirely. That knowledge, the most valuable in the building, lived in one person’s head. Not in the process. Not in the tool.

So the dispatchers did what people always do when a tool doesn’t match the work. They quietly stopped relying on it. They worked around it, passively ignoring the automation that was available, and since people watch each other, others followed. No one filed a complaint. No one held a meeting. The tool just got politely abandoned.

Nobody filed a complaint. The tool just got politely abandoned.

Here’s how it turned around. We stopped arguing about the software and started finding out how the work was really done. Once the real workflow was identified and captured, usage went up. And then came my favorite part: we began capturing how the work got done through conversations with the AI itself, and through interview sessions with the people who’d been doing it for years. The tool that had been ignored became the very thing that finally wrote down what the best dispatcher knew.

The machine hadn’t failed the people. We’d failed to teach the machine.

What a new hire needs

If you step back, all three stories have the same shape.

The first tool looked wrong because it lacked context. The citation went bad because of what it was fed. And the dispatch tool flopped because no one had given it the real job.

Not one of them was a story about a machine that couldn’t think. Each was a story about a machine that hadn’t been set up to succeed.

Think about how we treat people. Nobody expects a new hire to be brilliant from a process document on the first morning. We give them context. We pair them with someone who knows the ropes. We tell them how things really work, as opposed to how the binder says they work. We let them make a few mistakes and we correct them. Then, a few months later, we’re surprised that they’re good.

With AI, we hand over the process, expect perfection, and walk away at the first wrong answer.

Maybe some of that impatience is a little fear. A tool that talks in our language, answers in complete sentences, and sometimes out-answers the room is unsettling. It’s easier to catch it in one mistake and feel better. I understand the impulse. I’ve felt a flicker of it myself.

But the mistake isn’t the interesting part. The question is: why was it wrong, and what would make it right?

It’s worth remembering what an expert actually is. Not someone who never gets it wrong, but a master of the basics, someone who knows the fundamentals so well that they can tell when the standard answer doesn’t fit. That’s what the best dispatcher had. That’s what the tool was missing.

So when someone asks me whether AI can be trusted, I’ve started answering with another question.

What have you taught it?