The Biggest Skill in AI Isn't Coding — It's Management

The common assumption is that the people who will thrive in an AI-heavy world are the best engineers. Mike Taylor ran a 50-person growth marketing agency for five years before teaching himself to code. After four years of building AI products and writing the O’Reilly prompt engineering textbook, he has a different view: the skills that matter most with AI aren’t technical. They’re managerial.

Taylor is the CEO of Ask Rally, a synthetic focus group platform, and author of two O’Reilly books — one on prompt engineering, one forthcoming on context engineering with DSPy. His argument isn’t theoretical. It comes from noticing which of his own skills he actually uses when building AI systems.

The managerial toolkit

“I actually use my managerial skills more with AI than I use my data science skills,” Taylor says, “because most of the time it’s about did I provide a clear brief? Did I get it to plan first and accept the plan before I let it go ahead? Did I divide the labor up between different workers so that one of the workers doesn’t get too confused or overloaded?”

Every principle he lists maps directly to something he learned managing humans: clear briefs, planning before execution, division of labor, accountability systems, and logging. The translation from human management to AI management is more literal than metaphorical.

Taylor thinks that if prompt engineering gets renamed in the future, it might become part of business school curriculum. Agent management, synthetic corporations — the buzzwords are still forming, but the underlying discipline is recognizable to anyone who has run a team.

The Asana principle

At his agency, Taylor had an iron rule: if it’s not in Asana, it didn’t happen. A designer could claim they were waiting on a client, but Taylor would check the project management system and see the task was assigned yesterday. No Slack message, hallway conversation, or verbal agreement counted unless it was logged.

“That was my debugging of people,” he explains. “I’d look in Asana and say, this was assigned to this designer, the designer sat on it for a week, and then right before the deadline, they send it, and that’s why everything else is backed up.”

The exact same principle applies to AI systems. Taylor’s strongest recommendation for anyone building with AI frameworks is to look at the logs religiously. Sample what the model is actually sending. Most builders get lost in the abstraction — the framework calls it an agent, but what prompts is it actually firing? The answer, Taylor says, will shock you half the time.

You don’t need to manually replicate every step an optimizer takes. But you do need visibility into what your AI workers are doing, the same way you needed visibility into what your human workers were doing.

Why senior people struggle most

There’s a counterintuitive trap in the management-of-AI thesis. The people with the best management skills — senior professionals with years of experience — are often the worst AI adopters. Taylor has seen this pattern repeatedly.

“The senior people are the worst adopters of AI because they’re already pretty good. They already have great taste,” he says. “So they look at AI output and they go, it’s not good enough yet. But by the time it is good enough, it might be too late.”

Junior professionals don’t have that handicap. They look at AI output and think: this is already better than what I can do, so of course I’ll use it. The experience gap that used to be an advantage becomes a blindfold. The solution isn’t to lower your standards — it’s to force yourself through the discomfort of using AI even when the output doesn’t meet your bar, because the tool improves faster than you expect.

Three things Taylor did in 2020

Taylor offers a concrete playbook. First, force yourself to only do tasks with AI, even when it’s painful and slower than doing it yourself. Second, run actual Turing tests — send people AI-written content alongside human-written content and see if they can tell the difference. Taylor found that his own assessment of AI quality was colored by self-preservation instinct: he wanted the AI versions to be bad. The blind tests showed otherwise. Third, make space for play. Evenings and weekends count. The people who learn AI fastest are the ones who find it genuinely fun.

FAQ

Why are management skills more important than coding for AI?

Building with AI agents requires the same skills as managing human teams: writing clear briefs, planning before execution, dividing labor across agents, maintaining accountability through logging, and debugging when outputs go wrong. These managerial competencies determine whether AI systems produce reliable results more than raw technical ability does.

How do you debug AI agents when outputs go wrong?

Treat AI system logs the way a manager treats project management tools. Sample the actual prompts being sent — not the abstraction layer’s description of what it’s doing. Most AI frameworks obscure what’s happening at the prompt level. Regular log inspection reveals where agents are sending bad prompts, ignoring context, or producing inconsistent results.

Why are senior professionals the worst AI adopters?

Senior professionals with strong existing skills look at AI output and judge it against their own high standards, concluding it’s not good enough. By the time AI quality matches their bar, competitors who adopted early will have years of workflow optimization ahead. Junior professionals adopt faster because AI output already exceeds their current skill level.

What is the management-of-AI thesis?

The argument that managing AI agents draws more on organizational management skills — clear communication, task delegation, quality review, logging and accountability — than on data science or engineering. If prompt engineering gets renamed, it may become part of business school curriculum as a variant of team management theory.

How should you start using AI if you’re already experienced in your field?

Force yourself to complete every task with AI, even when it’s slower and worse than doing it manually. Run blind Turing tests to check whether your quality assessment of AI output is accurate or colored by self-preservation instinct. Make space for play — evenings and weekends count — because fun accelerates learning more than obligation.

What role does logging play in AI system reliability?

Logging is the accountability layer for AI agents, analogous to project management tools for human teams. Without logging, you can’t identify which agent failed, which prompt was ignored, or where the workflow broke down. Regular inspection of actual prompts sent and responses received is the primary debugging method for production AI systems.

Can non-technical people become effective AI builders?

Mike Taylor went from running a marketing agency to writing O’Reilly’s prompt engineering textbook by teaching himself to code through a self-paced bootcamp. Clients who hired him as a freelance developer never asked about his credentials — they only cared whether the code worked. Management experience, domain expertise, and clear communication skills often matter more than a traditional technical background.

How do you divide work between multiple AI agents effectively?

Apply the same labor division principles used for human teams: give each agent a focused task with a clear brief, prevent any single agent from getting overloaded with too much context, require planning before execution, and review outputs at checkpoints. Agent specialization prevents context confusion, the same way role specialization prevents scope creep in human organizations.