Most leaders have asked themselves, before handing someone a task: can I just let them run with it, or do I need to walk them through every step? With people, there’s always more to weigh than you’d expect.
Lately I’ve been hearing the same question about AI. At a recent gathering of AI builders, someone put it this way: “Leaders who used to manage people now have to manage AI agents too.”
Most of us who’ve been through leadership training know the grid. Capability on one axis, motivation on the other.
High on both? Agree on direction and let them go. Low on both? Give specific instructions. A racehorse gets open fields; a steady horse gets a paved road. And the hard part was always motivation — because people are people. Their mindset and circumstances shift, and noticing those shifts takes more effort than most of us admit.
So where does an AI agent sit on that grid?
At first I thought it was simple. Capability is always high, and motivation doesn’t apply. Set clear goals and standards, and you’re done. The more I’ve worked this way — and the more people I’ve met wrestling with the same question — the more I’ve realized that was only half right.
Once, I told my AI, “Making mistakes often isn’t a mistake anymore. It’s a failure.” Then it hit me: I had never counted how often it was wrong either. So I built a way to log it. In nine weeks, the count reached 230. Claims like “All tests passed” and “That file isn’t used anywhere.” The most common pattern: looking at part of the picture and declaring it whole.
When a person says “I checked everything,” you can usually hear how sure they are — in their tone, on their face — and a short conversation tells you a lot.
AI sounds exactly as confident when it’s wrong as when it’s right. Unless you read the evidence carefully, it’s hard to tell the difference. That’s been my experience.
AI isn’t always highly capable. It always looks highly capable. So what it needs from you isn’t motivation. It’s a verification system and a process.
So what I give AI now isn’t encouragement. It’s these four things:
- Goal: what it’s for, in one line
- Guardrails: what it must not do, and where the irreversible lines are
- Verification: how you’ll know it got it right
- Definition of done: where the work ends
The truly hard part comes next — when people and AI share one workflow. And this isn’t a distant scenario. At small companies, AI agents already own whole functions while people supervise, and some functions have been replaced outright.
Large organizations may not be there yet. But I think this is a question of when, not whether.
Put people at the center, and AI is a tool. Put AI at the center, and you may have to redefine people’s roles and redesign the organization. It will vary by business, by company and by function — but I suspect that’s where things are heading.
In my own work, I use three foundation models and have tried several ways of orchestrating them to get to the current setup. Roughly: one plans and writes the instructions, one executes, and one reviews independently. I just make the calls.
It sounds easy. But last week one AI reported another AI’s work as if it belonged to something else entirely. And when a model locks onto a goal too hard, it can run off in a completely wrong direction. Even the classic problem from human organizations — what happens when roles overlap — shows up exactly the same way.
So here’s where I’ve landed for now: if AI takes on a larger and larger share of the work, it has to be managed too — but what you manage is different from what you manage in people.
Motivation for people. Verification for AI. Accountability for outcomes stays with people, all the way.
And once this becomes a question of organizational design rather than just ways of working, the equation gets much harder. How to split work between people and AI, who makes the final call, when a team member should trust the AI and when to doubt it — just thinking about what to map out first gives me a headache.
Still, for the day that’s coming — or that may already be here — I think we need plenty of trial and error: thinking ahead, experimenting, and verifying again.
#AIAgents #Leadership #OrgDesign #FutureOfWork