The New Power User Is a Manager of AI Workers

The biggest productivity gap is no longer between people who use AI and people who don’t. It’s between people who use AI as a tool and people who are learning to manage it like a workforce.

Most professionals still sit in the middle of every workflow. That was me a few months ago. I’d ask AI to research something, review the answer, give it another instruction, move the result somewhere else, and decide what happened next. AI made each step faster, but I was still coordinating the work.

Power users are working differently. Instead of directing every step, they are building systems that can research, write, code, analyze, monitor, and prepare work with less intervention. Their job becomes deciding what to delegate, setting standards, designing the workflow, and stepping back in when judgment is required.

Once you start working this way, you will never look back.

You can already see this in practice. Silicon Valley Girl has shown a system where agents turn long-form videos into clips, generate titles and descriptions, publish them, and use performance data to inform future editorial decisions. The interesting part isn’t that AI can make a short video. It’s that dozens of small decisions can increasingly be turned into a repeatable system.

We started seeing this shift inside Forward Future several months ago. We now have always-on workflows that take published content and turn it into platform-specific distribution and repurposing opportunities across channels. We’ve built systems that continuously monitor parts of our website, identify performance bottlenecks, make targeted improvements, and retest the results. And on the engineering side, we can hand work off from Slack to coding agents that operate across projects, open pull requests, monitor CI, and keep working through failures until the build is green.

The same pattern is appearing inside much larger organizations. McKinsey describes a bank modernizing a legacy core made up of roughly 400 pieces of software with humans supervising squads of agents that documented applications, wrote code, reviewed one another’s work, and tested the results.

KPMG found that 54% of organizations were already actively deploying AI agents in early 2026, while 57% of leaders expected employees to manage and direct them.

That creates a new productivity gap: one person uses AI to move through work faster. Another builds systems that increasingly do the work on their behalf.

Both people are “using AI.” But they are not the same. They are operating at completely different levels.

The distinction becomes clearer when you look at a simple recurring task.

Imagine two people who both conduct competitive research every week.

The first asks Claude to research five companies, reviews the results, follows up where needed, and manually turns the output into a briefing for the team. AI makes each step faster, but the person still coordinates the entire process.

The second has built a workflow that collects relevant information, checks it against approved sources, compares it with previous findings, identifies meaningful changes, prepares the briefing, and flags anything uncertain or strategically important for review.

They may be using the same underlying model.

The difference is that one person is using AI to move through a workflow faster. The other has built a system that moves through the workflow for them.

Multiply that difference across research, email, CRM maintenance, coding, campaign analysis, content repurposing, scheduling, reporting, or any other recurring workflow, and the gap starts to get large very quickly.

Getting there requires more than good prompts. In practice, managing AI workers comes down to four skills: delegation, orchestration, evaluation, and improvement.

Good managers know that assigning work is not the same as defining success.

“Research our competitors” leaves too much unresolved. Which companies matter? What changes are worth surfacing? Which sources should the agent trust? What should it do when the evidence is incomplete or contradictory?

A strong agent assignment answers those questions up front. It defines the outcome, gives the system the relevant context, sets constraints, provides access to the right tools and information, establishes what good work looks like, and makes clear when the agent should stop and escalate to a person.

That is why expert agent use goes far beyond prompting. The prompt matters, but so do the system’s data, tools, permissions, memory, examples, environment, and rules of operation.

The useful question is no longer simply, What should I ask the model?

It becomes: What does this system need to know, access, and understand in order to succeed without me constantly intervening?

The second challenge appears when work starts moving between agents.

Consider a content workflow. One agent researches a topic. Another verifies the important claims. A third writes the first draft. Another turns that draft into social posts or video scripts. A final system prepares everything for publishing.

None of those steps is especially difficult on its own. The hard part is deciding how work should move between them.

What information should each agent receive? What happens when one agent finds a problem? Which actions can happen automatically, and which ones require human approval?

That is where agent management starts to look less like prompting and more like operations. The important pieces are responsibilities, inputs, outputs, permissions, dependencies, and escalation paths.

A workflow with five capable agents can still be worse than one good agent if the handoffs are poorly designed.

More autonomy creates another problem: AI is extremely good at producing work that looks complete.

A polished research memo can rely on a bad source. Code can run while introducing a failure elsewhere. A confident recommendation can be based on missing information. If you are delegating more work, “this looks right” quickly becomes an inadequate quality-control system.

This is where human judgment still matters, and where the best AI users separate themselves from everyone else.

KPMG and the McCombs School of Business at the University of Texas at Austin studied more than 500 junior workers and how they worked with AI. The researchers identified three broad behaviors.

Delegators largely accepted whatever the AI returned. Apprentices critiqued the work but often failed to improve the system effectively. Amplifiers iterated with the AI over multiple rounds and ultimately produced the strongest work.

The lesson isn’t that you should “double-check AI.” That doesn’t scale.

The more useful skill is learning to define quality in advance.

For a research agent, that might mean every material claim needs a primary source, statistics must be traced back to the original report, conflicting evidence must be surfaced, and uncertainty must be explicitly labeled. For a coding agent, tests must pass, performance cannot deteriorate beyond a specified threshold, and security-sensitive changes always require review. A sales agent might be prohibited from contacting an account with an unresolved support issue or required to escalate deals above a certain value.

This is where evals become useful far beyond AI labs. An eval is simply a repeatable way of testing whether an AI system is performing the task the way you expect.

Good managers don't inspect every keystroke from their employees. They establish standards for what good work looks like and pay attention when results fall outside them. AI systems need the same thing.

This is where the productivity gap starts to compound.

When most people get a bad result from AI, they correct it and move on. If a summary is weak, they rewrite it. If an agent misses an important source, they add it themselves.

That fixes today’s output, but it does nothing to improve tomorrow’s.

When a recurring AI workflow fails, the more valuable question is: Why did the system fail this way?

Was the instruction unclear? Was important context missing? Did the agent have the wrong tool, source, or level of autonomy? Or should that decision never have been delegated in the first place?

One executive interviewed by McKinsey compared bringing an agent into a workflow with onboarding a new employee. Both need feedback before they become effective.

The difference is that feedback to an AI system can often be encoded permanently. Change the instruction, add context, tighten a permission, improve an eval, or redesign a handoff once, and every future run can benefit.

That is where the leverage compounds.

The person using AI one conversation at a time starts from roughly the same place tomorrow.

The person improving the system can wake up with a better worker than they had yesterday.

There is an obvious ceiling to all of this: if you still need to approve every action an AI system takes, you are still the bottleneck.

The goal is not constant supervision. It is management by exception.

Routine work should happen automatically inside clearly defined boundaries. The system pulls you back in only when something falls outside them: two credible sources disagree, confidence drops below a threshold, required information is missing, an irreversible action is about to happen, financial stakes exceed a limit, or the task falls outside an established policy.

Instead of reviewing 100 ordinary decisions, you review the five that actually require judgment.

That is where agentic AI creates a different kind of leverage than a chatbot. The value is not just saving a few minutes on individual tasks. It is removing yourself from large portions of routine execution while preserving your attention for the decisions where you actually add value.

As access to powerful AI becomes widespread, the model itself will matter less than what people build around it.

The advantage will increasingly come from the systems they create: what they delegate, how they structure the work, how they evaluate the results, and how quickly those systems improve.

You do not need to build a ten-agent swarm to start developing this skill. In fact, that is probably the wrong place to begin. I made that mistake early on and wasted a bunch of time building elaborate systems before I had figured out which workflows were actually worth automating.

Take one recurring task you already use AI for and ask what it would take to remove yourself from most of the execution.

Competitive research is a simple example:

Collect sources → verify claims → compare against previous findings → identify meaningful changes → explain the implications → prepare a briefing → escalate uncertainty.

Then manage it like a system.

What does the agent need to know before starting? Which sources can it trust? Which decisions can it make autonomously? What does a correct result look like?

Run it several times. Watch where it breaks. When something goes wrong, resist the instinct to simply repair the output. Change the workflow so the same failure is less likely next time.

Once that works, find the next recurring process.

That is how AI stops being something you constantly operate and starts becoming something you manage.

The most capable knowledge workers won't be the people who can personally do the most with AI. They'll be the people who can reliably manage the most useful work being done on their behalf.