Not long ago, the word manager described a specific kind of role. You managed people, their schedules, their output, their growth, their morale. It was a promotion you earned. It came with training, or at least the expectation of training. There were books, courses, frameworks. Situational leadership. Radical candor. First, break all the rules.
We built an entire industry around the question of how to effectively direct other humans toward a shared goal.
And then, very quietly, we handed that same responsibility to people who never asked for it.
Today, a customer service agent prompting an AI assistant to draft a response is, functionally, directing a worker. A claims adjuster using a copilot tool to pull relevant policy information is delegating a research task. A contact center team lead configuring an AI agent's escalation logic is setting performance expectations for a member of their team — one that happens to run on a server somewhere. We are calling AI agents our AI teammates.
We have made everyone a manager. We just haven't told them that.
The Invisible Promotion
When organizations deploy AI agents into the workflow, the conversation tends to focus on the technology. What can it do? How accurate is it? What does it cost? How do we measure the ROI?
These are fair questions. But they skip over a more fundamental one: who is actually responsible for how this AI performs on the ground, day to day?
The answer, in most organizations, is: the person using it. The individual contributor. The frontline employee who was hired to handle customer interactions, not to coach, calibrate, and quality-check an AI system.
Think about what effective AI use actually requires. You need to know how to give clear direction — how to frame a prompt so the output is useful, not generic. You need to recognize when the AI is confidently wrong, which is different from when it's obviously wrong. You need to know when to override it, when to escalate, and when to trust it. You need to make judgment calls about quality, tone, and appropriateness — in real time, under pressure, while still doing the rest of your job.
That is management. Every bit of it.
The Skills Gap Nobody Is Talking About
Organizations have poured significant resources into AI readiness: infrastructure, integration, change management communications, training on how to access the tool, how to log in, how to submit a ticket if something breaks.
What they have largely not done is train people on how to direct AI well.
This is not the same as prompt engineering — a term that often conjures images of technical specialists crafting elaborate instructions for large language models. I'm talking about something more practical and more human: the everyday skill of knowing what to ask for, how to evaluate what you get back, and how to course-correct when it's off.
New managers receive coaching on how to delegate. How to give feedback. How to set expectations and how to hold people accountable without micromanaging. These aren't instincts — they're learned behaviors, refined over time with support.
We expect people to develop equivalent instincts for AI through osmosis. And then we wonder why adoption plateaus.
The symptoms are familiar. Employees who use the AI tool for the easiest, lowest-stakes tasks and ignore it for everything else. Teams where one or two people have figured it out and quietly become the go-to resource for everyone around them. Agents who paste AI output directly into customer responses without reading it — and agents who rewrite everything the AI produces because they don't trust it at all. Both behaviors, by the way, are rational responses to inadequate training.
What Managing AI Actually Looks Like
Let's get concrete. A customer service agent working alongside an AI assistant is, whether they know it or not, engaging in a management loop: assign → review → correct → repeat.
The AI Management Loop
They give the AI a task. The quality of that assignment shapes everything that follows. A vague prompt produces a generic response. A well-framed one — with context about the customer, the issue, the desired tone — produces something genuinely useful.
They evaluate the output. This requires judgment about accuracy, appropriateness, and fit for the specific situation. It requires knowing what good looks like — and that knowledge has to come from somewhere.
They decide whether to use it, modify it, or discard it. This is a quality decision. Made dozens of times a day, by every agent on the floor.
They do it again, ideally getting a little better each time at knowing what works. But only if they're getting some signal about what good looks like. In the absence of feedback, people optimize for speed, not quality.
And here is something we don't say enough: this loop is tiring. Not in the way that handling a difficult customer call is tiring, but in a quieter, more insidious way. Workers are increasingly running up against the limits of their own cognitive capacity — not from the work itself, but from the constant overhead of managing the work. As Francesco Bonacci, founder of Cua AI, put it after a day of AI-assisted coding: he ended each day exhausted not from the code but from the managing of it — six worktrees open, four half-written features, two quick fixes that had spawned rabbit holes, and a growing sense that he was losing the plot entirely.
That's not a developer problem or a coding problem. That's what happens when people are handed a capable but undirected collaborator and left to figure out the relationship on their own. Cognitive saturation and mental fatigue are real reported side effects of unstructured AI use — and they are a signal that the management layer is missing, not that the technology is.
This loop is invisible in most organizations. There's no coaching around it. No measurement of it. No shared language for talking about it. It just happens, thousands of times a day, with widely varying results.
This Is a Leadership Problem, Not a Technology Problem
Here is where I want to push back gently on how we talk about AI adoption challenges. The conversation often settles on one of two explanations when things go slowly: the technology wasn't ready, or the people were resistant to change.
Both framings put the problem somewhere other than where it often actually lives — in leadership's failure to equip people for a genuinely new kind of work.
Imagine hiring a team of first-time managers, giving them a new employee to supervise, and then providing no onboarding, no coaching, no frameworks, and no feedback on how they're doing. You would expect inconsistent results. You would be right to expect them. And you would not describe those results as the new managers being resistant to management.
That is exactly the situation most frontline employees are in with AI today.
The employees who have figured it out — who use AI fluidly, confidently, and to genuine effect — almost always have one thing in common: they found a way to get feedback. From a peer. From a team lead who was curious enough to dig in. From their own experimentation with enough time and psychological safety to try things and fail without consequence. They built, informally, what should have been provided formally.
What Good Looks Like
Organizations that are getting this right are treating AI direction as a learnable, coachable skill — and building structures to support it.
They're running calibration sessions where teams review AI outputs together — not just to catch errors, but to build shared judgment about quality. They're creating simple prompt libraries — not rigid scripts, but starting points that encode institutional knowledge about what works. They're asking team leads to coach on AI interactions the way they coach on call handling: Walk me through your thinking on this one.
They're also being honest about the fact that managing AI is a skill that takes time to develop. They're not measuring adoption by login rates or feature usage. They're asking harder questions: Are our people getting better at this? Are they more confident than they were six months ago? Do they know when to trust the AI and when to override it?
These questions are uncomfortable because they are slow to answer. But they are the right questions.
A New Definition of Readiness
As AI agents become more capable — handling more complex tasks, operating with greater autonomy, taking actions on behalf of customers with less human review — the stakes of this gap get higher.
An agent who doesn't know how to direct a simple AI assistant today is not going to suddenly know how to oversee an autonomous AI agent tomorrow. The skills compound. And so do the gaps, if we don't address them.
AI readiness is not a technical threshold. It's not about having the right infrastructure or the right integrations or the right governance policies — though all of those matter. It's about whether the people doing the work have the judgment, the language, and the support to direct AI well.
We have given everyone a direct report they didn't ask for. The least we can do is help them become the manager that direct report deserves.
