By John Wheeler, Founder of Wheeler Intelligent Systems and Builder of a One-Person AI-Native Company
I think I accidentally recreated a very old management problem with very new technology.
I have been building Wheeler Intelligent Systems as an experiment in what I call an AI-native company. I am the human, but more of the company’s work is being performed by AI.
I now have AI general managers working in different parts of the company. Research has a GM. Sales and Opportunities has one. Product development has one. Courses and Education has one. AI workforce management has one. Other AI roles help manage and evaluate work across the company.
The idea is simple.
Give each AI manager a job. Give it goals and the information it needs. Define what it can and cannot do. Let it learn from prior work. Then allow it to operate without asking me about every small decision.
It is beginning to work.
And that has created a problem I didn’t expect.
My AI company can create work faster than it can finish it.
We have created the AI version of a very old, very non-AI management problem:
Too much work-in-process.
Seven AI Managers Can Create a Lot of Work
When I first started using AI, I was still the source of most of the work.
I thought of a question and asked AI to research it. I thought of a software feature and asked AI to help build it. I thought of an article and asked AI to help write it.
I controlled the amount of work simply because I could only think about so many things at once.
That changes when AI managers can identify their own useful work.
A Research GM can discover something that needs more study. Sales can find a business problem worth exploring. Product can turn an idea into an experiment. Education can turn research into a course. One AI manager can hand work to another.
Suppose seven AI managers each identify three worthwhile next actions.
That’s 21 pieces of work.
Tomorrow they can do it again.
And AI is fast.
It can research while I eat dinner. It can analyze an opportunity while I sleep. It can draft a course while I’m doing something else.
But all those outputs have somewhere to go.
Research needs to be evaluated. Product experiments need conclusions. Software needs testing. Courses need review. Opportunities need decisions.
Moving work is not the same as finishing work.
Manufacturing Has Known This for Years
There is nothing new about work-in-process.
Imagine a factory where the first machine can produce 1,000 parts per hour, but the second machine can process only 100.
Making the first machine twice as fast doesn’t necessarily increase finished production.
It may simply create a bigger pile in front of the second machine.
The first machine looks incredibly productive.
But the factory isn’t producing more finished goods.
The bottleneck moved.
AI can do exactly the same thing to knowledge work.
We can make research 10 times faster. Writing can become dramatically faster. Software can be produced faster. We can generate more ideas than we could ever investigate.
But if our ability to review, verify, decide, approve, integrate, and act doesn’t increase too, we create piles.
They just don’t look like factory inventory.
They look like reports, documents, recommendations, experiments, software pull requests, emails, research findings, and tasks.
AI Can Make a Bad Management System Look Productive
This is where businesses need to be careful.
AI produces visible activity.
An agent completed 30 tasks. Another generated 15 reports. A coding agent wrote thousands of lines of code. A research agent reviewed hundreds of sources.
Those numbers can look like enormous productivity gains.
Maybe they are.
But I am becoming more interested in a different question:
How much useful work reached a finished and accepted state?
If AI writes ten reports and nobody uses them, what did we gain?
If AI writes software that never passes review, what did we produce?
If AI identifies 20 business opportunities but nobody tests them, did we create opportunities—or 20 more pieces of work-in-process?
And if AI creates ten times as much material but people spend all day reviewing it, did we remove work?
Or did we simply move it?
Starting Work Has Become Cheap
Historically, starting a project had a natural cost.
A market study required someone’s time. Software required development capacity. A course required research, writing, editing, and production.
Those costs forced some discipline.
You couldn’t start everything.
AI changes that.
“Research this.”
“Analyze this.”
“Build this.”
“Write this.”
“Test this idea.”
Each instruction can now start substantial work at very little cost.
That’s an incredible capability.
But it removes one of the natural brakes on work-in-process.
AI makes starting cheap. It does not automatically make finishing cheap.
The New Bottleneck May Be Judgment
Someone—or some properly governed system—still needs to decide:
Is this research good enough?
Should we act on it?
Does this product deserve more investment?
Is this software correct?
Is this course ready?
Can this AI worker be trusted with more authority?
Some of that judgment can also be assigned to AI. I am experimenting with independent AI evaluators and AI managers handing work to other managers.
But that creates another old management problem:
handoffs.
A report being sent isn’t the same as someone accepting it.
A task being assigned isn’t the same as someone taking responsibility for it.
A draft being completed isn’t the same as an accepted result.
I am starting to think about AI work as moving through a chain:
Created → Assigned → Acknowledged → Worked → Evaluated → Accepted → Used → Outcome measured
That’s very different from:
Agent says, “Done.”
AI Companies May Need WIP Limits
This brings me back to an old operations idea: work-in-process limits.
Maybe the smartest AI company isn’t the one running the most agents.
Maybe it’s the one that knows when not to start more work.
Research might finish existing handoffs before beginning another major investigation. Product might complete its current experiment before launching five new ones. A software team might stop creating new code while finished work is waiting for evaluation.
That can look inefficient.
The AI is available. Why not keep it busy?
Because keeping AI busy isn’t the goal.
Producing valuable finished outcomes is the goal.
I think our AI metrics will need to reflect that.
Instead of only measuring agent runs, tokens, documents, or tasks, we should also ask:
How much work is currently in process?
Where is it stuck?
How many handoffs are waiting?
How much requires human review?
How much work was accepted?
How much accepted work was actually used?
And how much human attention did it take?
One measure I am especially interested in is:
Accepted useful output per hour of human attention.
The Machines Are New. The Management Problem Isn’t.
This is one of the most interesting lessons from my AI-native company experiment.
The technology is new.
Many of the management problems aren’t.
We still have bottlenecks. We still have queues. We still have bad handoffs. We still start too many things. We still confuse activity with results.
AI doesn’t eliminate those problems.
AI can amplify them at machine speed.
I started this experiment asking:
How much work can AI do?
I’m beginning to think the better question is:
How much AI work can the company successfully absorb, finish, verify, and turn into useful results?
That brings us back to a very old management principle that may become even more important in the AI age:
Stop starting. Start finishing.