He Fired the Employee. Then the AI Answered Exactly Like Him.
By Atiba de Souza

The room got quiet
Halfway through my four-day intensive in Vegas in June, a doctor in the room was building an AI employee to take over part of the job of a staff member he had let go the week before. The person had never seemed to understand what was needed. So he handed the work to an AI.
Then he looked up and said it, cleaned up for the cuss words:
"It's responding exactly like the employee I fired last week."
The room got quiet. Not the polite quiet. The other kind, where everyone in it has just heard something about themselves and needs a second.
If you have sat in an AI and leadership session this year, that is not the moment the session was built for. Those sessions are built to teach you how to lead the AI. The doctor was not leading anything in that moment. He was being shown something.
That is what most AI and leadership talks get backwards. They start with the machine and work backward toward you. The work runs the other direction.
He let the person go a week before the intensive. By the halfway point of four days, the AI was answering exactly like the person he fired.
The framework is the Agentic AI Diagnostic Sequence. Six moves:
- Constraint. Identify the binding constraint via Theory of Constraints.
- Role. Place the work in the AI-Empowered Org Chart.
- Work type. Classify it as Strategic, Tactical, or Analytical, then assign each slot Human or AI.
- AI employee. Define the role on the Diagnostic Canvas.
- PRD. Write the product requirements.
- Build.
Between step one and step two there is an un-numbered move I will not let you skip: triage. TOC asks where the leverage is, meaning whether this work sits at the constraint, feeds it, or sits near it. Eisenhower asks what the leverage of that work is, urgent against important. Then you eliminate. Work with no leverage gets deleted before a person or an agent is assigned to it.
Now rank the thing honestly, because not all six moves carry equal weight. Step one and the triage move are the whole game. Steps two through six are craft, and craft matters once you know you should be building at all. If you keep one idea from this piece, keep this: the sequence is a filter, not a checklist. Steps four, five, and six are where the money gets spent. Step one and triage decide whether you should spend it.
Which is exactly why the keynote skips it. The keynote starts at step four with a demo of an AI employee doing something impressive, or at step six with a build. Then the rest of the hour is change management. Step one never makes the slide, because step one asks a question about your business that you might not like the answer to, and a room that paid to feel ahead does not want that question.
Step one is the step with your name on it
Binding constraint: the one place in the system where the work is actually limited. TOC asks a narrow question about it: is this work at the constraint, feeding the constraint, or near the constraint?
And here is the rule that turns that definition into a decision: if the work is not at, feeding, or near the constraint, eliminate it before you assign it to a human or an AI. Not after. Before. If you build an AI employee for work that was never the thing holding you back, you have bought speed on the wrong road, and speed on the wrong road is the most expensive kind of progress there is. It looks like momentum and it moves nothing.
This is also the only step that can hurt, and that is the part the talks leave out.
Every other step returns an answer about a process, a role, or a tool. Step one can return an answer about the person who commissioned the work. If the constraint turns out to live in how work gets asked for and reviewed, the constraint has your name on it. And you cannot fire yourself.
Sit in that for a second, because it is the actual reason step one stays off the slide. Every leader I know would rather replace a person, a vendor, or a tool than rewrite how they ask. Replacing something outside you feels decisive. Rewriting how you ask feels like self-doubt. So the meeting goes to the tool.
Hear this clearly, though. This is not self-blame. Self-blame is cheap and it changes nothing. This is leverage. You cannot get out of the bottleneck while you are certain the bottleneck is everybody else.
One more thing, because a framework handed over as a universal is a framework you will misuse. When not to run this: if the work is plainly at the constraint, plainly urgent, and plainly important, stop diagnosing and build. The sequence earns its keep on the work you keep doing without ever asking why. It is a terrible excuse to spend a quarter writing requirements for something you already know matters.
The doctor was not running an AI project
He thought he was. Look at where he actually was. He was at step four, defining an AI employee, before he had run step one and triage on that role. The build was already half done when the diagnostic came back.
And the diagnostic was not the one he ordered.
The AI did not fail him. It responded exactly like the person he had fired the week before. Same gaps, same misses, same shape coming back over the desk. The common factor between the person who left and the employee he built was not intelligence, not effort, not which model he was using. It was how he gave instructions.
Which leaves one sentence he had to say to himself, and it is not a comfortable one. Maybe it was my fault. Not theirs.
That is the bill for this work, and it arrives right here, not in a footnote at the end. To take the mirror seriously, he had to look at the last person he let go, the last vendor he churned, and the team restructure he had been sizing up. And he had to consider that the constraint was upstream of all of them, sitting in the ask. Some of those decisions may well stand. He does not get to keep that assumption until he looks.
Without the mirror, here is what repeats. He blames the person, hires again, hands over the same instructions, and gets the same result with a new name on the door. That is what the AI bought him that a person could not: a diagnostic, delivered clean, before he repeated the pattern with the next hire.
So the decision is small, and it is available Monday morning. Rewrite how you ask before you replace who you ask. Before the next hire, before the next vendor, before the next restructure, before the next agent.
And then, and only then, keep walking the sequence. Run the triage honestly. Is this work at, feeding, or near the constraint? Does it survive Eisenhower? Is there anything here to eliminate before it gets a slot? If yes, go on to role, work type, AI employee, PRD, build. If no, delete it, and notice what you just saved. You were about to build a faster version of something that was never holding you back.
The walk in Austin
The same lesson reached me a year and a half before that session. It did not arrive through an AI.
I was walking the streets of Austin during the last Get Scalable Live event there, talking with my chief of staff, Stephanie, about where the company was going. I was looking at my team. Graphic designers, video editors, artists. Really good at what they do. Struggling with AI and with what they were producing.
I said to myself: I would have to replace this entire team. Everyone. They are not going to make it.
Have you ever said something to yourself and known, in the same second you said it, that it was wrong? That was the moment. The bad feeling turned into a question. What if I accept that they are artistic and I am not, and we still get to the moon?
Look at what that question does not say. It does not say I will teach them to work the way I work. It gives up the assumption that my way was the standard they had to meet. That question became Organizational Independence, and it begins with building to replace yourself.
Had I gone the other way, I would have fired a team that is very good at what it does. I would have stayed the guy who is always superior at AI. And I would have stayed the bottleneck that everyone brings everything to, in a company that could not move without me standing in the middle of it.
Now look at the two moves side by side, because they are not the same move. The doctor's issue was the ask: what did I actually request, and did I request it clearly. Mine was the standard: who decided my way was the bar. Both start with catching yourself mid sentence and refusing to finish it the easy way.
And that is the rewiring cost, paid in reps, not insight. The reflex is years deep. When work comes back wrong, the fast question is who do I replace, and it feels like leadership. The slow question is what did I actually ask for, and it feels like hesitation. You do not get to feel fast on Monday. You get to ask the slow question once, while the frustration is still hot, and then again the next time it comes up. That is the only way the reflex changes.
Where the claim actually sits
Here is where the analogy earns its keep.
Ask a company how far along they are with AI and you will usually get a category, not a distance. We use AI agents. We are AI-forward. That is not a position. That is a badge.
Put it on a scale instead. Deploying an agent is the first two steps of a much longer walk. The constraint work is further down the road. The org chart work is further. Defining the AI employee properly, writing the requirements, building something that actually holds: further still, and mostly uphill.
AI-forward is not a category you join by switching an agent on. It is a distance, and most companies are calling the second step the destination.
That is why those talks feel so good and change so little. A category tells you that you have arrived. A scale tells you how far you have left and in which direction. Only one of those is arguable, and only the arguable one is useful.
The two versions, side by side
| Inconsistent | Consistent | |
|---|---|---|
| Where it starts | Capability: step four or step six, "look what the model can do" | Constraint and triage: step one, "what is actually limiting us" |
| What "AI-forward" means | A category, a badge you earn by switching an agent on | A distance, measured from the constraint outward |
| When work comes back wrong | Replace the person, then hand the same ask to the next person | Rewrite the ask before replacing anybody, yourself included |
| What the sequence is for | A build checklist you enter at the demo | A filter that decides whether to build at all |
| Where it ends | A launch announcement | A constraint that moved |
Before Monday
You do not need a new tool this week. You need one honest pass.
Take the thing that came back wrong most recently. Not the hardest thing, the most recent one, because the frustration is still warm enough to be useful. Then walk it through the front of the sequence before you touch the back of it.
Was this work ever at, feeding, or near the constraint? Be specific, because "it feels important" is not an answer. Then triage it. Urgent and important, or eliminable? And if it survives, ask the question that costs: what did I actually ask for, in writing, in a form another person could have followed.
If the answer is that your instructions were the constraint, the fix is not a new hire and it is not a new model. The fix is the ask. That one is yours, and that is good news, because it is the only part of this you can change without anyone's permission.
If you skip this pass, you will read the rest of the sequence, nod, and build a faster version of the wrong thing. I have watched a room go quiet watching a man discover that in real time. You can have the same discovery for free, on a walk, before the invoice arrives.
Questions