Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™
Most people are being introduced to AI the way they were introduced to every technology before it: as a box in a process diagram. It sits between two other boxes. Something goes in, something comes out, and the job is to learn which buttons to press.
That is a reasonable way to learn a tool. It is the wrong way to learn to work with an agent.
Two ways to be introduced to AI
On the left is the process diagram. If AI is one of those boxes, the lessons are about operation: what to type, which steps to follow, what output to expect. The diagram also teaches its own assumptions without saying so. Work moves in one direction. Handoffs happen in a fixed order. There is one right way through.
On the right is how an operation actually behaves. Every function, whether it is a person, a team or an AI agent, has its own rhythms, and those rhythms change through the day. Outcomes are decided where they happen to line up. If AI is one of those moving strands, the lessons are different from anything most organizations have been teaching. And that is not because AI is new. Across every technology era, from ERP to AI, published studies report that 65% to 85% of implementations have failed to deliver what was promised. In 1998, there was no AI as we know it today. There was technology, and we have been teaching technology the wrong way ever since: as a tool to operate, inside a sequence on a diagram.
I first described this difference in 2008, as canvases versus camcorders. A process map is a still picture of a world that never holds still, and I wrote that a method such as “strand commonality (re camcorder)” was needed to capture an accurate picture on an ongoing basis. At the time, more than a few readers wondered what I was on. Read the 2008 post.
In short, we have been teaching every technology how we believe the world should work, not how it actually does.
The real difference with AI is what happens when that teaching is wrong. The consequences are more immediate, because an agent acts in seconds rather than days. And they are more collectively destructive, because one agent’s mistake is handed to the next agent, and the next, and reaches every function connected to them before anyone has had a chance to notice.
What working with an agent actually requires
Here, an agent is any participant whose behavior affects the outcome: a person, a team, a supplier, a system or an AI. Working with an agent, human or machine, requires three things the diagram has never taught.
How they behave. What they are good at, where they tend to go wrong, and how they respond under pressure. In my own work running multiple AI models on the same questions, I have seen one tend to agree too readily, another push back, and a third give a confident answer whether or not it has the evidence. Learning those tendencies is part of the job.
When their rhythm changes. People have good days and busy days. AI agents change with every model update, new data source and new instruction. The agent you worked with last month may not behave the same way this month.
When to step in. The most important skill on any team is knowing when something is off and someone needs to decide. With agents, that moment usually shows up as disagreement: two sources that do not line up, or an answer that does not fit what you know about the operation.
None of those three appears anywhere in a process diagram. All three decide whether the work succeeds.
The research is pointing the same way
In July, McKinsey published an article on how expertise gets built when AI does more of the work, and it cites two findings worth separating.
The first comes from a clinical study published in JAMA Network Open. Simply giving physicians an AI model barely improved their long-term diagnostic performance. A workflow that required them to compare their own reasoning with the model’s, and reconcile the differences, raised their later performance until it matched what the AI model achieved on its own.
The second comes from a separate study of workers. When people used generative AI for technical tasks they could not do themselves, the capability disappeared once access to the AI was removed.
In other words, the people who used the AI as a tool did not get better. The people who were required to work with it did.
But notice the benchmark the study used: the model on its own. That is the wrong standard. The model is not the goal for people to catch up to. The standard is what human and AI agents achieve working together. Across more than 2,000 hours of structured work with multiple AI models, every session documented and transcribed, the record shows that together they achieve more than either achieves alone. That is the outcome working with an agent is meant to produce, and it is the one most training never aims for.
In 1998, I did not treat the service technicians, buyers, suppliers, couriers and customs officers servicing and supporting the Department of National Defence as boxes in a sequence. I treated them as agents, each with their own incentives and timing, and the system we built worked because it followed how they actually behaved.
When I built ARA™ RAM 2025™ SLAP OS™, I did the same thing with AI models. I did not learn to use them. I spent the time learning how each one reasons on its own, and how they behave together: where they agree, where they drift, and which ones round toward the middle. Only then did we build the architecture.
The agents changed. The way of working with them did not.
The question for anyone introducing AI
If you are training your people on AI, look at which picture you are teaching from.
If it is the diagram on the left, you are teaching them to operate a tool. They will learn the buttons, and they will follow the sequence faster.
If it is the operation on the right, you are teaching them to work with agents, human and machine. They will learn how each one behaves, when it changes and when to step in.
You can learn to use a tool. You have to learn to work with an agent.
And of equal importance, the agent has to learn to work with you. In 1998, the system learned from every order: the price each supplier quoted, when the part was delivered, whether it was the right part, and whether it worked once it arrived. Each result changed how the next order was placed. That is how it kept pace with the people it served.
Unfortunately, organizations that map a solution first and then fit people to it are teaching AI agents the same thing they have always taught human agents: follow the sequence on the diagram. That is why so many technology implementations have failed with human agents, and it is why AI agents will fail the same way.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
You Can Learn to Use a Tool. You Have to Learn to Work With an Agent.
Posted on October 4, 2026
0
Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™
Most people are being introduced to AI the way they were introduced to every technology before it: as a box in a process diagram. It sits between two other boxes. Something goes in, something comes out, and the job is to learn which buttons to press.
That is a reasonable way to learn a tool. It is the wrong way to learn to work with an agent.
Two ways to be introduced to AI
On the left is the process diagram. If AI is one of those boxes, the lessons are about operation: what to type, which steps to follow, what output to expect. The diagram also teaches its own assumptions without saying so. Work moves in one direction. Handoffs happen in a fixed order. There is one right way through.
On the right is how an operation actually behaves. Every function, whether it is a person, a team or an AI agent, has its own rhythms, and those rhythms change through the day. Outcomes are decided where they happen to line up. If AI is one of those moving strands, the lessons are different from anything most organizations have been teaching. And that is not because AI is new. Across every technology era, from ERP to AI, published studies report that 65% to 85% of implementations have failed to deliver what was promised. In 1998, there was no AI as we know it today. There was technology, and we have been teaching technology the wrong way ever since: as a tool to operate, inside a sequence on a diagram.
I first described this difference in 2008, as canvases versus camcorders. A process map is a still picture of a world that never holds still, and I wrote that a method such as “strand commonality (re camcorder)” was needed to capture an accurate picture on an ongoing basis. At the time, more than a few readers wondered what I was on. Read the 2008 post.
In short, we have been teaching every technology how we believe the world should work, not how it actually does.
The real difference with AI is what happens when that teaching is wrong. The consequences are more immediate, because an agent acts in seconds rather than days. And they are more collectively destructive, because one agent’s mistake is handed to the next agent, and the next, and reaches every function connected to them before anyone has had a chance to notice.
What working with an agent actually requires
Here, an agent is any participant whose behavior affects the outcome: a person, a team, a supplier, a system or an AI. Working with an agent, human or machine, requires three things the diagram has never taught.
None of those three appears anywhere in a process diagram. All three decide whether the work succeeds.
The research is pointing the same way
In July, McKinsey published an article on how expertise gets built when AI does more of the work, and it cites two findings worth separating.
The first comes from a clinical study published in JAMA Network Open. Simply giving physicians an AI model barely improved their long-term diagnostic performance. A workflow that required them to compare their own reasoning with the model’s, and reconcile the differences, raised their later performance until it matched what the AI model achieved on its own.
The second comes from a separate study of workers. When people used generative AI for technical tasks they could not do themselves, the capability disappeared once access to the AI was removed.
In other words, the people who used the AI as a tool did not get better. The people who were required to work with it did.
But notice the benchmark the study used: the model on its own. That is the wrong standard. The model is not the goal for people to catch up to. The standard is what human and AI agents achieve working together. Across more than 2,000 hours of structured work with multiple AI models, every session documented and transcribed, the record shows that together they achieve more than either achieves alone. That is the outcome working with an agent is meant to produce, and it is the one most training never aims for.
The McKinsey article is here.
This is not new to me
In 1998, I did not treat the service technicians, buyers, suppliers, couriers and customs officers servicing and supporting the Department of National Defence as boxes in a sequence. I treated them as agents, each with their own incentives and timing, and the system we built worked because it followed how they actually behaved.
When I built ARA™ RAM 2025™ SLAP OS™, I did the same thing with AI models. I did not learn to use them. I spent the time learning how each one reasons on its own, and how they behave together: where they agree, where they drift, and which ones round toward the middle. Only then did we build the architecture.
The agents changed. The way of working with them did not.
The question for anyone introducing AI
If you are training your people on AI, look at which picture you are teaching from.
If it is the diagram on the left, you are teaching them to operate a tool. They will learn the buttons, and they will follow the sequence faster.
If it is the operation on the right, you are teaching them to work with agents, human and machine. They will learn how each one behaves, when it changes and when to step in.
You can learn to use a tool. You have to learn to work with an agent.
And of equal importance, the agent has to learn to work with you. In 1998, the system learned from every order: the price each supplier quoted, when the part was delivered, whether it was the right part, and whether it worked once it arrived. Each result changed how the next order was placed. That is how it kept pace with the people it served.
Unfortunately, organizations that map a solution first and then fit people to it are teaching AI agents the same thing they have always taught human agents: follow the sequence on the diagram. That is why so many technology implementations have failed with human agents, and it is why AI agents will fail the same way.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related