There is a common assumption buried in most conversations about AI in the workplace: that AI is a tool, and tools need operators. The person using AI is the operator. The AI does the work. The distinction between doing and directing stays where it has always been.
That assumption is wrong in ways that matter for how organizations develop people.
When you direct AI toward a task, you are not pressing a button. You are doing leadership work. You decide what the goal is. You explain the context. You define what good looks like. You allocate the task. You review the result. You take responsibility for what gets sent out. These are not peripheral activities that happen around the real work. They are the real work, when AI is handling execution.
The consequence is that leadership capability - previously the domain of people with a certain title or a certain position in the hierarchy - is now a functional requirement for almost anyone who uses AI to accomplish something meaningful.
What Leadership Work Actually Consists Of
Leadership gets discussed in vague terms so often that it is worth being specific about what it actually involves at the operational level.
At its core, leadership work is the work of directing others. That includes: articulating a goal clearly enough that someone else can pursue it without constant guidance; providing the context that allows someone to make good judgments in situations you have not anticipated; setting a standard for quality and helping others develop a shared understanding of what that standard means; delegating tasks with enough specificity that the person executing understands what to do and what to decide on their own; reviewing results against the standard and feeding back in ways that improve future performance; and being accountable for the outcome, not just the effort.
Every one of these activities is required when working effectively with AI. The person who can articulate what they want, provide the right context, set clear quality standards, delegate appropriately, and review results critically will consistently get better output from AI than someone who types a vague instruction and hopes for the best.
The difference between those two people is not technical literacy. It is leadership capability.
Why This is Not Just About Prompt Engineering
The rise of "prompt engineering" as a concept has captured some of what matters here, but it also obscures it. The framing suggests that better AI output is primarily a matter of knowing the right techniques: how to structure an instruction, which modifiers to add, which format to specify. This is real, but it is secondary.
The primary skill is knowing what you want clearly enough to explain it. No prompt technique substitutes for actually understanding your goal, your constraints, your quality standard, and your context. And that understanding is not a technical skill. It is the kind of clarity that comes from thinking seriously about what you are trying to accomplish.
Consider two people drafting a customer proposal with AI assistance. The first knows their product well and understands what this particular customer cares about; they provide context, specify the audience, and describe what a useful proposal looks like for this relationship. The second asks AI to "write a proposal for a potential client." The gap between their outputs has almost nothing to do with prompt engineering and almost everything to do with whether they understood the task and the audience well enough to direct someone else toward it.
That understanding - of what the task actually is, why it matters, and what a good result looks like - is leadership knowledge. People develop it through experience, through feedback, through responsibility. It cannot be learned by reading prompt collections.
The Hierarchy Problem
For decades, organizations have concentrated certain types of clarity at the top of the hierarchy. The senior person knows the strategy. The manager knows the priorities. The director knows what the client actually needs. People further down the hierarchy are given tasks rather than goals, and are often not expected to need the full context.
This model has costs that organizations have always absorbed: slow decision-making, reliance on key individuals, difficulty adapting to change. But it also had a certain logic. Information was scarce, decisions had high coordination costs, and there were real reasons to centralize judgment.
AI disrupts this logic. When AI can execute at the level of a skilled individual, the bottleneck is no longer execution. The bottleneck is the quality of direction. And direction that flows only from the top of the hierarchy, translated through several layers before reaching execution, is both slow and degraded. Organizations that want to use AI well cannot afford to have goal clarity, context, and quality standards locked up with a small number of people.
This creates pressure to distribute leadership work. Not to flatten the hierarchy entirely - there are still good reasons for coordination and oversight at senior levels. But to ensure that people at every level can articulate goals, provide context, and evaluate results for their own domains.
That is a different capability profile than what most organizations have historically developed in their people. And it requires different conditions than most organizations have historically created.
The Conditions That Make It Possible
Research on what allows people to work well and develop professionally has produced fairly consistent findings over several decades. Autonomy, competence, and connection to meaningful outcomes are the conditions that support both motivation and growth (Deci and Ryan's self-determination theory). Psychological safety - the ability to raise problems, admit uncertainty, and take risks without fear of punishment - is the condition that allows teams to learn (Edmondson's research on team learning).
Both of these frameworks map directly onto what AI-assisted leadership development requires.
Autonomy is necessary because developing leadership judgment requires actually making judgments - about what to ask for, how to direct, what to evaluate. An employee who needs approval for every interaction with AI cannot develop these capabilities. They are executing, not leading. The organizations where AI capability develops fastest are those where people have genuine latitude to try things, adjust their approach, and own the outcomes.
Psychological safety is necessary because directing AI well involves getting things wrong in visible ways. A poor prompt, a misspecified goal, a quality standard that turns out to be wrong - these are the normal materials of learning. Teams where people cannot admit that their AI interaction produced garbage, or where trying something novel carries social risk, will develop their AI capability slowly if at all.
Competence development - in the self-determination theory sense, not just skill accumulation - requires real tasks with real stakes and real feedback. This is not served by one-time AI training sessions. It is served by giving people actual work to do with AI, reviewing the results together, and treating the experience as material for learning rather than as a performance evaluation.
The Manager's New Job
If leadership work is being distributed to more people, what does this mean for people who are formally in leadership roles?
It does not reduce the importance of leadership. It increases the importance of certain kinds of leadership.
The work of clarifying direction - what are we actually trying to accomplish, and why does it matter - becomes more important, not less, when AI can accelerate execution. Unclear direction produces worse outcomes at higher speed with AI than it does with purely human execution. The cost of ambiguity rises.
The work of setting and maintaining quality standards becomes more important when AI makes it easy to produce large volumes of output quickly. Speed without standards produces a lot of bad outputs fast. The manager who can articulate what good looks like, and develop that understanding in their team, is more valuable than the manager who can produce quickly themselves.
The work of creating the conditions for team learning - psychological safety, meaningful feedback, genuine autonomy - becomes more important when capability development is continuous and distributed rather than periodic and centralized.
None of this is new management theory. What is new is the urgency. Organizations that have always known, abstractly, that these conditions matter now face direct consequences for not having them, because AI adoption tests them immediately and visibly.
Developing the Capability
Organizations that want their people to work well with AI need to invest in leadership capability, not just AI literacy.
That means giving people responsibility for outcomes, not just tasks. It means making quality standards explicit and teachable rather than leaving them in the heads of senior people. It means creating space to experiment, fail, and learn without treating every poor result as a performance problem. And it means managers doing the actual work of providing context, setting direction, and reviewing results - not delegating that entirely to the people they manage.
The most important shift is probably treating this as a capability that develops through real work rather than through training. You do not become better at directing work by learning about directing work. You become better at it by directing work, seeing what happens, and adjusting.
That is also, not coincidentally, exactly how people develop any other form of leadership capability. AI has not changed that. It has simply made it visible and urgent in a new way.