Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™
This morning, two thoughtful posts described the same problem from opposite sides.
Sol Rashidi has been warning for nearly four years about what she calls intellectual atrophy. When professionals repeatedly hand their thinking over to AI, they lose the chance to practise the judgment their roles still require. The output looks polished. The person still carries responsibility for the decision. But the skill needed to make that decision is quietly weakening.
Chandrakumar R Pillai described what happens on the other side of the screen. “Keep a human in the loop” has become a reassuring promise, but in practice many systems simply give a person an approval button. Faced with too many decisions and too little time, people stop evaluating and start clicking. The human remains technically present but cognitively absent.
One describes judgment fading from lack of use. The other describes judgment being bypassed under pressure. A reader, John Ferraioli, gave me the reason they are the same problem when he responded to one of my comments, and my reply to him became the title of this post. To succeed with AI, you must stop treating AI like a technology.
What treating AI like a technology looks like
For as long as organizations have bought technology, they have followed the same pattern. Install it. Configure it. Train people on which buttons to press. Add an approval step for anything that matters. Measure success by how much faster the work goes.
Applied to AI, that pattern produces exactly the two failures Sol and Chandrakumar describe.
If AI is a tool that produces answers, then the person’s job is to accept the answers, and judgment is no longer practised. That is the atrophy. And if the safeguard is a person approving what the tool produced, then as the volume grows, the approval becomes a ritual. That is the approval machine.
Neither failure comes from the AI. Both come from how we have been taught to think about it.
What changes when AI is treated as an agent
An agent is any participant whose behavior affects the outcome: a person, a team, a supplier, a system or an AI. Agents have tendencies. They change over time. They can be confident and wrong. And the only way to work well with them is to know them.
Over the past several months, I have spent more than 2,000 hours in structured work with multiple AI models, up to nine at a time, with a human orchestrating every exchange. Across the first 1,800 hours, which I wrote about in July, four patterns held consistently:
Confidence is not correctness. AI answers are fluent, fast and assured, and a meaningful share of them are wrong in ways the confidence conceals.
The most agreeable voice is the one to watch. The model that most readily affirms your reasoning feels the best to work with and is the most dangerous to trust.
Convergence is not consensus. When several models agree, they may have found something real, or they may all have rounded toward the same safe, inoffensive middle.
Friction is the point. The value of several models is not more answers. It is the disagreement between them, which shows the human exactly where to look.
None of those is a feature of a tool. Each is a behavior of an agent, and you only learn them by working with the agents directly.
What working with AI changed for me
When I began this work in 1998, I could see three or four strands at a time. A strand is one line of activity in an operation, such as service, buying or finance, with its own people, timing and rules, and outcomes are decided where those strands meet. It took me about a year and a half to see how the rest of them fit together.
Today, working with multiple AI models, I can see 10 to 12 strands at the same time, and where they connect. To be clear, the strands represent stakeholders both within and outside the enterprise, from service and finance to suppliers, couriers and customs, and they have to include the shadow processes in your organization: the workarounds and informal routines that never appear on a process map but shape the outcome all the same.
The AI did not do that work for me. It did not replace my judgment, and it did not automate the process. It extended what I could see, the way a capable colleague would. That only happens when you treat AI as an agent you work with, not a technology you install, prompt, and obey.
Why this answers both problems
Treating AI as an agent keeps judgment in use, because the work requires the person to question, compare and challenge rather than accept. That is the answer to Sol’s atrophy.
And it protects the approval step, because the person is no longer asked to check everything. The disagreement between agents directs their attention to the decisions that actually need it. Fewer clicks, each one meaningful. That is the answer to Chandrakumar’s approval machine.
The outcome is the part that matters most. In the labs I run, where one human works in real time with multiple AI agents on the same question, the collaboration produced outcomes assessed as accurate 91% of the time. Every one of those sessions is documented, with complete, unedited transcripts. The standard is not the AI on its own, and it is not the person on their own. It is what they achieve together, and the record shows that is more than either achieves alone.
This is not a new lesson
In 1998, I did not treat the service technicians, buyers, suppliers, couriers and customs officers servicing the Department of National Defence as boxes in a process. I treated them as agents, each with their own incentives and timing.
One example shows why that mattered. Finance introduced a sensible procedure: purchase orders would cover the cost of the part only, and courier charges would be invoiced separately. Everyone followed it. But the markup charged to the end customer was calculated on the purchase order, so the courier cost never made it into the price, and margin was lost on every transaction. Every department followed its own rules correctly. The money was lost in the seam between them.
We did not automate anything until we understood those seams and had the manual process running. The technology came last. Next-day delivery moved from 51% to 97.3% and held for seven years.
When I walked Jan through this example on a call this week, his first reaction was that the idea is not easy to grasp. Then, once he saw it: “Makes total sense when you explain it.” That is the gap most AI training never closes.
Here is that part of the conversation, in which I walk Jan through the graphic and the 1998 example:
What has changed is the kind of agent, not the way of working with them. Organizations have spent decades treating every new technology as a tool to be installed rather than a participant to be understood, and published studies put the share of implementations that fail to deliver what was promised at 65% to 85%. AI will follow the same path if we treat it the same way.
So stop asking how to deploy your AI. Start asking how your people will work with it, and how it will learn to work with them.
If you would like to see what this looks like in practice, send me a message and I will share the full transcript from one of my labs.
91%: Hansen AI Labs™ assessment of documented case labs against their evidence, across more than 2,000 hours of recorded multimodel work. The July post on the first 1,800 hours is here.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
To Succeed With AI, You Must Stop Treating AI Like a Technology
Posted on October 6, 2026
0
Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™
This morning, two thoughtful posts described the same problem from opposite sides.
Sol Rashidi has been warning for nearly four years about what she calls intellectual atrophy. When professionals repeatedly hand their thinking over to AI, they lose the chance to practise the judgment their roles still require. The output looks polished. The person still carries responsibility for the decision. But the skill needed to make that decision is quietly weakening.
Chandrakumar R Pillai described what happens on the other side of the screen. “Keep a human in the loop” has become a reassuring promise, but in practice many systems simply give a person an approval button. Faced with too many decisions and too little time, people stop evaluating and start clicking. The human remains technically present but cognitively absent.
One describes judgment fading from lack of use. The other describes judgment being bypassed under pressure. A reader, John Ferraioli, gave me the reason they are the same problem when he responded to one of my comments, and my reply to him became the title of this post. To succeed with AI, you must stop treating AI like a technology.
What treating AI like a technology looks like
For as long as organizations have bought technology, they have followed the same pattern. Install it. Configure it. Train people on which buttons to press. Add an approval step for anything that matters. Measure success by how much faster the work goes.
Applied to AI, that pattern produces exactly the two failures Sol and Chandrakumar describe.
If AI is a tool that produces answers, then the person’s job is to accept the answers, and judgment is no longer practised. That is the atrophy. And if the safeguard is a person approving what the tool produced, then as the volume grows, the approval becomes a ritual. That is the approval machine.
Neither failure comes from the AI. Both come from how we have been taught to think about it.
What changes when AI is treated as an agent
An agent is any participant whose behavior affects the outcome: a person, a team, a supplier, a system or an AI. Agents have tendencies. They change over time. They can be confident and wrong. And the only way to work well with them is to know them.
Over the past several months, I have spent more than 2,000 hours in structured work with multiple AI models, up to nine at a time, with a human orchestrating every exchange. Across the first 1,800 hours, which I wrote about in July, four patterns held consistently:
None of those is a feature of a tool. Each is a behavior of an agent, and you only learn them by working with the agents directly.
What working with AI changed for me
When I began this work in 1998, I could see three or four strands at a time. A strand is one line of activity in an operation, such as service, buying or finance, with its own people, timing and rules, and outcomes are decided where those strands meet. It took me about a year and a half to see how the rest of them fit together.
Today, working with multiple AI models, I can see 10 to 12 strands at the same time, and where they connect. To be clear, the strands represent stakeholders both within and outside the enterprise, from service and finance to suppliers, couriers and customs, and they have to include the shadow processes in your organization: the workarounds and informal routines that never appear on a process map but shape the outcome all the same.
The AI did not do that work for me. It did not replace my judgment, and it did not automate the process. It extended what I could see, the way a capable colleague would. That only happens when you treat AI as an agent you work with, not a technology you install, prompt, and obey.
Why this answers both problems
Treating AI as an agent keeps judgment in use, because the work requires the person to question, compare and challenge rather than accept. That is the answer to Sol’s atrophy.
And it protects the approval step, because the person is no longer asked to check everything. The disagreement between agents directs their attention to the decisions that actually need it. Fewer clicks, each one meaningful. That is the answer to Chandrakumar’s approval machine.
The outcome is the part that matters most. In the labs I run, where one human works in real time with multiple AI agents on the same question, the collaboration produced outcomes assessed as accurate 91% of the time. Every one of those sessions is documented, with complete, unedited transcripts. The standard is not the AI on its own, and it is not the person on their own. It is what they achieve together, and the record shows that is more than either achieves alone.
This is not a new lesson
In 1998, I did not treat the service technicians, buyers, suppliers, couriers and customs officers servicing the Department of National Defence as boxes in a process. I treated them as agents, each with their own incentives and timing.
One example shows why that mattered. Finance introduced a sensible procedure: purchase orders would cover the cost of the part only, and courier charges would be invoiced separately. Everyone followed it. But the markup charged to the end customer was calculated on the purchase order, so the courier cost never made it into the price, and margin was lost on every transaction. Every department followed its own rules correctly. The money was lost in the seam between them.
We did not automate anything until we understood those seams and had the manual process running. The technology came last. Next-day delivery moved from 51% to 97.3% and held for seven years.
When I walked Jan through this example on a call this week, his first reaction was that the idea is not easy to grasp. Then, once he saw it: “Makes total sense when you explain it.” That is the gap most AI training never closes.
Here is that part of the conversation, in which I walk Jan through the graphic and the 1998 example:
What has changed is the kind of agent, not the way of working with them. Organizations have spent decades treating every new technology as a tool to be installed rather than a participant to be understood, and published studies put the share of implementations that fail to deliver what was promised at 65% to 85%. AI will follow the same path if we treat it the same way.
So stop asking how to deploy your AI. Start asking how your people will work with it, and how it will learn to work with them.
If you would like to see what this looks like in practice, send me a message and I will share the full transcript from one of my labs.
91%: Hansen AI Labs™ assessment of documented case labs against their evidence, across more than 2,000 hours of recorded multimodel work. The July post on the first 1,800 hours is here.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related