Management is derivative. Leadership is Human

AI can automate almost any knowledge task. It cannot automate most knowledge roles. Nearly every argument I see about AI in organizations lives in the gap between those two sentences.

Sometime in the past year, model intelligence stopped being the constraint. The leading models are already smart enough to do the digital work most companies run on: coding, integrations, reporting, analysis, and a large share of what we call management. So why hasn't every org transformed overnight?

Because delegating work to an agent safely and competently is real work. And it is not the model's work. It's ours.

The good news is that you don't need to be an AI expert to do it. You have to ask for what you want.

The loop is simpler than the strategy deck

I run the internal AI team at Stord. Here is the entire loop we ask people to run. Someone asks Claude for what they want, in plain language. One of three things happens.

It works. Good, move on.

It can't reach the data or take the action. That's a capability gap, and it belongs to engineering. My team's job is connectors and tools, not domain knowledge.

It can do the work, but only after a lot of back-and-forth guidance. If it's a process that will come up again, turn that session into a skill: codified instructions the agent can invoke later. Now the tenth person to need it gets the result without knowing the skill exists.

That's it. No AI strategy document required to start.

What this arrangement gets right is the division of labor. Our engineers don't need to learn the ins and outs of every function in the company. They provide generic capabilities that agents can use. The people closest to the work provide the judgment, the instructions, and the definition of success. Both halves are necessary and neither can do the other's job.

Non-technical teams have to own the instructions layer for their own domain. That part doesn't delegate upward to the AI team. The taste and judgment in a controller's head, or a warehouse ops lead's head, has to get written down by them.

Your users are already writing your backlog

Product teams are sleeping on the most direct consequence of this.

If your users are working with AI, you no longer have to guess what they want. You don't need to write speculative user stories, run a focus group, or send a survey. You can look at what they tried to do with an agent and failed to finish, and build the tooling that lets them succeed.

They're asking for what they need in their own words, and every attempt the agent made is sitting in the thread. That's not a replacement for product judgment. It's a direct record of unmet demand, timestamped, unfiltered, in the user's own language.

Start collecting your traces. It may be the highest-signal backlog you have.

The skill is context management, not prompting

The common advice is that prompting is the skill. Asking for what you want in specific language does matter, and it doesn't come naturally to most people. But it misses the pattern that separates advanced users from casual ones.

The real skill is context management. And good managers are already doing it.

When you hand work to one of your reports, you don't stream-of-consciousness brain-dump at them and call it an assignment. You put yourself in their seat. What do they already know from their vantage point? Where does the rest of the information live? Who should they ask? How does this fit the company's goals? What constraints can't they violate?

You also don't hand them every scrap of context you have. That's overwhelming, and it sends people in directions you didn't want. You wouldn't dump context for twenty projects on someone at once, either. You keep each person focused on one thing.

All of that transfers directly.

What's changed recently is that agents now gather much of their own context. We've been connecting Claude to the systems our work actually lives in: Atlassian, Google Drive, our order and warehouse management systems, and the rest of the stack. Once company information is findable by the agent, a good prompt only has to specify what needs to happen and where the context lives.

You also don't need the perfect prompt. Work with agents is collaborative, the same way it is with people. You don't send one definitive email and pray the output matches what you pictured. You have a conversation, clarify as you go, and watch for whether your intent is landing.

Why it's worth the effort

This is more work than typing a question into a box. Six properties make it pay off.

Agents can have effectively perfect memory. The context window is finite, so any single session can only hold so much. But skills persist across sessions. After you guide an agent through a job, tell it to turn the process into a skill. The next session follows the same path without you re-explaining.

You can run them in parallel. Once a skill does what you need, you can run a hundred instances of it at once.

You can run them in the background. A tuned skill can run in the cloud on a schedule, at whatever cadence the work deserves.

Skills compound as company advantage. Share a skill with your department and your team can be completely unaware it exists while still getting the benefit. They ask for what they need, and the agent invokes it. Everyone inherits the version you workshopped.

The process clarifies the work itself. A good agent will interrogate you until the implicit becomes explicit. The side effect is that you just wrote the SOP you never got around to writing.

This is the worst the technology will ever be. Longer context, continual learning, and custom enterprise models are all in motion. Getting fluent now is how you stay ahead of the next capability jump instead of starting over at it.

What to rethink

Take agents to their logical extent and a few organizational defaults stop making sense.

Cut handoffs. When one person can drive work across several domains, does a process that requires four handoffs still earn its cost? Is there a single owner who could just finish it? If the handoff exists for security or governance, the question becomes how to make the approval itself cheap rather than how to route around it.

Stay flat. Every layer between the strategy and the person doing the work is a layer of translation loss. If employees can do meaningfully more, give them meaningfully more ownership and get out of the way.

Reporting doesn't generate revenue. Your dashboards, email summaries, and weekly reviews do not make money. They give visibility and they support decisions, which is real. But reporting should increasingly be a side effect of doing the work rather than a separate job. It's derivative, and derivative work is what AI is best at.

There are no diminishing returns on customer satisfaction. Speed is not a license to do less. AI capability is becoming ubiquitous, which means it isn't a moat, and the rules of competition haven't changed. So what do you do with the time you get back? If you're genuinely customer-obsessed you shouldn't have to ask. You should already have a long list of things you'd do for customers if you had the hours.

Where management ends and leadership starts

Here's the line I keep coming back to.

Coordinating, reporting, reviewing, translating, and moving information between people is management. Choosing a direction, setting standards, making tradeoffs, developing people, and owning the consequences is leadership.

The first list is largely automatable. The second isn't, and I don't expect it to be soon. If a role's center of gravity is routing information between humans, that role is going to change shape, and the people in it are better served by hearing that early than by discovering it late.

This cuts the other way too. AI can do the work. Humans still own the outcome. Scale your review to what's at stake, and treat delegation as a transfer of effort, never a transfer of judgment. If your agent sends a customer the wrong number, exposes data it shouldn't, or makes a bad call, "Claude did it" is not a defense that exists.

What this asks of you

None of this requires being intimidated by the tools, or by the bigger problems they let you take on. It does require accepting that there's a skill here, and that skill takes intentional practice.

The employees who are thriving with this look similar across functions. They're T-shaped: deep in the domain you hired them for, plus a horizontal working knowledge of basic tech, context management, and where AI is strong and where it falls over. That combination lets them take on more, and it effectively makes them managers, of agents rather than people.

So optimize for less rigid job descriptions and more accountability for outcomes. When one person can carry a multi-domain task without waiting for another team's backlog, they can own the delivery end to end, on both quality and timeline.

Stop waiting for an AI strategy. Ask for ambitious outcomes. Remove the barriers your agents hit. Turn what works into skills. Hold humans accountable for the result.

Business as usual is a choice now, and it's getting more expensive.