My first real use of AI was planning a vacation.

Not a work project, not some grand transformation. I opened ChatGPT and Copilot, used the plain chat box, and asked for help mapping out a trip. Then some personal projects. Then a bit of planning at work. For a while it was a research assistant, and a good one.

The shift happened when I started asking it to make things, actual deliverables I’d use. That’s when I noticed the pattern that changed how I think about all of this: the quality of what came back was almost entirely a function of what I put in. A vague ask got a vague, generic answer. A clear one, with the context and the constraints spelled out, got something I could actually use. The better my input, the better the output. Every time.

When I moved to Claude Code, where the output really is the whole point, that lesson went into overdrive. I started refining my instructions, saving the good ones, building a small library of them. And somewhere in there it clicked: this is exactly what it feels like to ask a member of my team for something. The clearer I am about what I want and why, the better the result. I wasn’t prompting a tool. I was delegating to a report who happened to be a machine.

That’s a small realization with a long tail. Because if working well with AI is really an act of delegation, then the skill it demands is leadership. And nearly everyone is about to need it.

TL;DR

The skill isn’t prompting. It’s delegation.

There’s a lot of talk about prompt engineering, as if the trick is a magic phrase. That’s not what I’ve experienced. What actually moves the quality is the same thing that moves the quality of work I hand to a person: a clear outcome, enough context, the constraints that matter, and a sense of what “good” looks like.

When I ask a colleague for a report, I don’t hand them three words and hope. I tell them what it’s for, who’s going to read it, when I need it, what to leave out. Do that with AI and the output changes just as much. Skip it and you get something that’s technically responsive and practically useless. The people I see getting real value from these tools aren’t the ones who found a clever prompt. They’re the ones who are good at defining what they actually want.

That’s a leadership skill. It always has been.

The other half is judgment

Delegation doesn’t end when you hand off the work. The harder half comes when the work comes back.

Early on, I got fooled. The deliverables AI produced looked great. They were polished, well structured, and delivered with total confidence. So I trusted them. And they were wrong. Not always dramatically. Sometimes the error was buried a few layers down, the kind of thing you’d only catch if you already knew the subject. A small number of times, early on, I didn’t check carefully enough and shipped something I shouldn’t have.

I learned fast that everything AI made with me needed a genuine, in-depth review before it went anywhere. Ethan Mollick, who studies this closely at Wharton, describes today’s AI as a brilliant, eager intern: fast and capable, occasionally confidently wrong, and always in need of someone who checks the work. That matches my experience exactly. The polish is the trap. Confidence isn’t correctness, and the tool can’t tell the difference. You have to.

So leading an AI takes both halves of leading a person: a clear ask going in, and real judgment coming out. Neither of those is a technical skill.

Leadership used to be rationed

Here’s the part I keep coming back to.

For as long as I’ve worked, leadership has been a scarce role. In any organization, only some people lead. Part of a senior leader’s job was to find them: to watch the team, spot the few best suited to set direction and own outcomes, and invest in pulling them up. Leadership was rationed by design, handed to the few who earned it.

AI breaks that model. When everyone on the team is directing an eager, unreliable intern every single day, everyone is doing the thing leaders do. Setting the outcome. Giving direction. Judging the result. Owning what ships. An individual contributor with no direct reports is now, in practice, leading many times a day.

I want to be precise about the word, because the popular version of this idea is that we’re all becoming managers of AI agents. Managing is real, but it’s the smaller claim: assigning tasks, tracking outputs, keeping the workflow moving. What these tools actually demand is leadership, which is deciding what’s worth doing, knowing what good looks like, and taking responsibility for the outcome even when a machine did the typing.

Management is mostly process. Leadership is judgment and ownership. AI automates a lot of the first. It raises the price of the second.

What this looks like on a real team

My team is early in this, like everybody. I’ve written before that the AI skills gap on a team is a leadership gap first; this is the same idea one level down. And I’ve already watched it play out in individual people.

One team member started experimenting and checking in with me. Some days they were genuinely excited. They’d seen what their own direction could pull out of the tool, and it lit them up. Other days they were frustrated: the output wasn’t what they wanted and they weren’t sure how to move it, other than to give up and do it by hand. That gap, between the excitement and the frustration, was almost entirely a directing gap. Learning to lead the tool was the difference.

Another team member turned in a thirty-page AI-generated report as their own work, essentially untouched. It became a coaching moment. I asked them, honestly: did you read this yourself? If I asked you to stand up and present it right now, could you? The answer was no. The lesson landed: AI can do the heavy lifting, but you still have to be the domain expert. You still have to own the thing with your name on it. Handing over raw output isn’t delegation. It’s abdication.

Self-check

Are you an AI user or an AI leader?

Rate how often each is true of how you actually work. Six quick reads, then a result that stays in this browser.

Directionthe ask going in
Before I prompt, I get clear on the outcome I actually want.
I give the AI the context and constraints I'd give a new teammate.
I tell it what “good” looks like — the audience, the format, an example.
Judgmentthe review coming out
I review the output closely, assuming it may be confidently wrong.
I know the subject well enough to catch what's subtly off.
I could stand up and defend every output as my own work.

Runs in your browser. Nothing is saved or sent.

So what does a leader do now?

If leadership is becoming everyone’s job, then a senior leader’s own job inverts. For years mine was mostly to find the few best suited to lead and pull them up. Now the work is to grow that capability across the whole team.

That’s changed concrete things for me. In hiring, comfort with AI is now a core part of what we screen for. When a candidate balks at using these tools to do better work, I read it as a real signal about fit. That part is the easy part, honestly, because you’re choosing who comes in.

The existing team is the harder and more important work. We’re standing up our own internal coaching program, because the range of skill is enormous, with both ends of the spectrum sitting on the same team. We’re pairing people new to AI with a buddy who’s further along. And we’re moving toward cross-functional cohorts, small groups drawn from across the different marketing-technology disciplines, working a real problem together. None of this is really prompt training. It’s the judgment and directing instinct we used to develop only in people we’d already picked for a leadership track.

The part that worries me

I learned judgment by making mistakes. Everyone I respect did. You do the entry-level work, you get some of it wrong, and over years you build the instinct for what good looks like. But AI is very good at exactly that entry-level work: the first draft, the starter analysis, the grunt tasks people used to cut their teeth on.

So if the machine does the reps, where does the next generation get the judgment to lead the machine?

I worry about it. The fear that AI makes us lazy isn’t unfounded. My hope is that people early in their careers understand the effort didn’t vanish, it moved, into direction and review and judgment instead of raw production. Mistakes will still get made. The real question is whether we still treat them as how people learn.

It also points at the hardest part of my job now: letting my team fail. If I catch and fix everything before it can go wrong, they never learn the lesson that keeps it from happening twice. So I have to let some things break. Not everything, which is what guardrails are for. We put clear ones in place: what data never goes into a tool, what always gets a human review before it leaves the building, which decisions get escalated instead of automated. Staying inside those lines is exactly what lets me give people room to experiment, because marketing leadership is comfortable sponsoring the work when the downside is capped and the kinds of risks IT and security care about are handled up front. The guardrails aren’t there to prevent failure. They’re there to make failure survivable, so it can still teach.

I built this very website with AI as my partner, with no engineering background, by doing all of this: directing, reviewing, catching confident mistakes, and failing my way to something that works. It’s the clearest proof I have of the argument. I’m not an engineer, but I could lead one very eager intern to build the thing.

We spent a long time treating leadership as a title for the few. AI just handed a version of it to everyone. The leaders worth following now are the ones who see that, and spend less energy guarding the role than building it in the people around them.