November 10, 2026
November 10, 2026
The Judgment Premium: Why Expertise Matters More When Answers Are Free
Listen to this article
In September 1945, with Europe’s factories and production lines in pieces, Friedrich Hayek used the pages of the American Economic Review to write about spare stock. The knowledge that keeps an economy running, he argued, is seldom the scientific kind. It is knowledge of “the particular circumstances of time and place”: scattered, partial, often contradictory, owned by nobody in full. His examples are the person who knows where surplus stock can be found when supplies are interrupted, and the shipper who makes a living by filling cargo space on tramp steamers, ships without a fixed route, that would otherwise sail empty or half full. Hayek was describing the daily craft of procurement and logistics, eighty years before anyone needed the reminder.
We have since built machines and systems that come closer than anything before them to storing general knowledge in one place. A market overview, a supplier screening, a first should-cost model, a freight benchmark: an afternoon’s work where it once took a fortnight. From there it is a short thought to declaring expertise in procurement and supply chain a commodity. The thought leads the wrong way. As general answers approach a price of zero, the knowledge Hayek described is what remains worth paying for.
What has become cheap
Much of what is losing its value carried the consulting business for thirty years. Knowing who supplies what, at which cost structure and from which region, was an asset billed by the day. So was the pyramid: a partner sells the engagement, a team of analysts produces the benchmarks and the slides. Both lived off an information advantage that a well-briefed language model has largely erased. Inside companies the same erosion reaches the buyer whose staying put on knowing the supplier market by heart.
The loss carries a second cost that is harder to see. When everyone consults the same handful of models, the answers converge. In the BCG field experiment discussed further below, consultants working with GPT-4 produced ideas that were 41% less diverse than those of colleagues working without it. For anyone responsible for sourcing strategy that is an awkward finding. A category strategy drawn from the same model as the competitor’s may read convincingly, yet it is unlikely to give an edge over other players. The same applies for a dual/multi-sourcing strategy: if every business in an industry is moving towards the same three alternative suppliers, the alternative has reached capacity before the first purchase order is placed.
What the model forgets
In July 2024, researchers from Oxford, Cambridge, Imperial College, Toronto and Edinburgh published a study in Nature that business circles have largely passed over. They asked what happens when generative models are trained, generation after generation, on the output of earlier models. The fine details of the distribution go first. Rare events become under-represented and eventually vanish, while the common cases take over the model’s picture of the world. The authors call this model collapse, and they conclude that data from genuine human interaction will become more valuable as model-generated content spreads.
Every buyer and every planner knows the fine details by another name. It’s the part number nobody watched: the C-part with a single qualified source, the magnet buried in a bought-in motor, the export licence nobody needed until April 2025. Risk management in procurement and supply chain is the management of exactly these cases.
The procurement and supply chain literature written between 2000 and 2020 feeds a model for none of this. It describes a world of falling trade barriers, cheap energy, lean inventories and lowest-cost-country sourcing. A model trained on it will reproduce that world and with the greatest confidence on exactly the questions where that world in many areas no longer applies. Ask it for a safety stock policy or a sourcing region, and the answer will carry the logic of the years it learned from. Of course it all depends on the complexity and depth of the prompt which has been fed into the model.
Confidently wrong
The study most often quoted on AI in consulting is the field experiment that researchers from Harvard, Wharton, MIT and Warwick ran with 758 BCG consultants, now published in Organization Science. Its headline numbers are familiar: on tasks within the model’s capabilities, consultants using GPT-4 completed 12.2% more of their work and did so 25.1% faster.
For one task the researchers deliberately chose a business case in which the spreadsheet looked complete while the decisive clues were tucked away in interview notes with company insiders. Without AI, 84.5% of consultants reached the right recommendation. With GPT-4, between 60% and 70.6% did. Graders nevertheless rated the AI-assisted recommendations as more coherent and more persuasive, right or wrong. The group that had received prompt training did worst of all, which suggests that familiarity with the tool bred more reliance on it, not less.
AI is no less reliable than any other source. Every source gets things wrong now and then. The difference is that AI gets things wrong in flawless prose, and a polished report looks just as convincing when its conclusions are weak as when they are sound. Anyone who has been on a sourcing committee has seen the human version of this: the most polished supplier evaluation usually wins the room, whether or not it is the best one. With AI every evaluation is polished, so polish no longer tells you anything. The authors add that their participants were early in their careers and that results might differ for experienced managers. That caveat deserves attention: the study tested precisely the people who have not yet built the judgment this article is about.
The irony nobody budgets for
In 1983, the cognitive psychologist Lisanne Bainbridge published a five-page paper on automation in process plants and cockpits. Her point: automation doesn’t push people out of the system. It leaves them with the hardest part. The better the control system, the more depends on the operator in the rare moments it fails. And that operator has had almost no practice, because the system handled everything else.
Procurement and supply chain are about to repeat this. AI now does the market scans, quote comparisons and price analyses. Those were the tasks that taught junior buyers their job. Manual forecast corrections and stock reviews did the same for young supply chain planners.
After a few years of comparing offers line by line, a buyer can tell when a price is too good, when a supplier is overstretched, when a deal doesn’t add up. A planner who has chased enough missing parts knows which lead times in the system are wrong. Take that training away, and we lose the experienced category managers and planners that will be needed in ten years. AI will handle more by then, but not everything.
This is the uncomfortable part. Some of the efficiency AI promises is borrowed from future capability. A CPO or Head of Supply Chain who automates entry-level work without rethinking how people learn is making a decision about 2035, whether they mean to or not.
What stays scarce
When general answers are cheap, three things gain value. Hayek would have recognised all of them.
Local knowledge. Which of the two approved suppliers really has spare capacity? Which plant needs nine months to qualify a new source? Which carrier still has space when a port is congested? What does it mean when a supplier’s managing director suddenly asks for shorter payment terms? None of this is in a database. Most of it is never written down.
Two weeks before a major plant turnaround, the main contractor came to see me. The shutdown had been planned for months, with hundreds of external contractors and a client waiting. He told me his company had filed for insolvency. Nobody had seen it coming. The firm the whole plan was built around was gone.
A model would have given me a list of alternative contractors in seconds. It could not have made a single one pick up the phone. What saved the turnaround were relationships built over years: people who took my call, trusted me and sent crews at a notice no tender process allows. We went ahead with other contractors. None of what made that possible was available in a system.
Accountability. A supervisory board asked to move tens of millions in production volume to another country wants to know who stands behind the recommendation, and what that person has seen first-hand. A model can make an argument persuasive, as the BCG study shows. It cannot be held accountable.
Execution. Transformations rarely fail for lack of analysis. They get stuck in internal politics, in budgets nobody wants to fund, in a plant that won’t give up its buffer stock, in a business unit defending a supplier of twenty years of good cooperation. Getting an organisation to move still happens in meeting rooms, on the phone and on the shop floor.
None of this means seniority wins. People can be as outdated as models. A category manager who spent twenty-five years moving volume to low-cost countries, or a supply chain planner trained on ever leaner inventories, carries most probably an old world view too. It just has no cut-off date printed on it. Experience can turn into habit as easily as into judgment. The difference is whether someone questions their own knowledge and patterns when a rule that worked for twenty plus years is going to stop. A model can’t do that between training laps. A person can. Not every experienced person does, and so the premium will be just accessed by those who do.
What this means for CPO's and Supply Chain leaders
Procurement teams are already drawing the line, mostly by instinct. EFESO’s 2026 CPO Pulse interviewed 50 CPOs. They see the most value for generative AI in contract analysis (69%), supplier and market intelligence (61%) and automating sourcing processes (55%). Only 35% see value in AI-assisted negotiation. Just 5% have rolled the technology out at scale.
A deliberate split could look like this:
Hand over to AI | Keep with experienced people |
|---|---|
Contract review and summaries | Deciding which risks the business is prepared to carry |
Supplier and market scans | Judging a supplier’s actual resilience, ideally on site |
Spend classification and first should-cost models | Negotiation strategy and the relationship behind it |
Forecast baselines and freight rate comparisons | Deciding where to hold buffer stock, and who will be supplied first when material is short |
Monitoring news, sanctions lists and risk signals | Interpreting what a signal means for business operations |
First drafts of category strategies | Final sourcing and location decisions |
One rule is missing from most AI roadmaps: let junior buyers and supply chain planners do some of the analysis by hand for a while, then compare their results with the model’s. It takes time. It is also the only reliable way to develop people who can tell when the model is wrong.
Apply the same test to external advisory and consulting, ours included. Three questions help:
Which parts of your work come from AI, and who checked them against reality?
Where is the model’s answer misleading for our situation, and why?
Will you personally sign off the recommendation, and what would change your mind?
Who can I call from your network to verify your success story?
An adviser who answers the last question with a name has the kind of knowledge Hayek wrote about.
The judgment premium
When a technology makes something cheap, whatever complements it becomes more valuable. Spreadsheets made calculation almost free and easy to do. Knowing what was worth considering for the calculating became the scarce skill. AI is now doing the same analysis and knowledge. The premium moves to judgment: knowing where AI’s limits are, taking responsibility for decisions beyond them, and motivating an organisation to act.
PROCUREMENT reWIRED is built on that conviction. It combines AI-supported intelligence for breadth and speed with decades of hands-on experience in manufacturing procurement and supply chains, for the decisions where the data runs out.
Hayek’s man with the surplus stock is still out there. In times of structural change, his world moves faster than any model can be retrained. That is the best reason I know to keep asking him.
Listen to this article
In September 1945, with Europe’s factories and production lines in pieces, Friedrich Hayek used the pages of the American Economic Review to write about spare stock. The knowledge that keeps an economy running, he argued, is seldom the scientific kind. It is knowledge of “the particular circumstances of time and place”: scattered, partial, often contradictory, owned by nobody in full. His examples are the person who knows where surplus stock can be found when supplies are interrupted, and the shipper who makes a living by filling cargo space on tramp steamers, ships without a fixed route, that would otherwise sail empty or half full. Hayek was describing the daily craft of procurement and logistics, eighty years before anyone needed the reminder.
We have since built machines and systems that come closer than anything before them to storing general knowledge in one place. A market overview, a supplier screening, a first should-cost model, a freight benchmark: an afternoon’s work where it once took a fortnight. From there it is a short thought to declaring expertise in procurement and supply chain a commodity. The thought leads the wrong way. As general answers approach a price of zero, the knowledge Hayek described is what remains worth paying for.
What has become cheap
Much of what is losing its value carried the consulting business for thirty years. Knowing who supplies what, at which cost structure and from which region, was an asset billed by the day. So was the pyramid: a partner sells the engagement, a team of analysts produces the benchmarks and the slides. Both lived off an information advantage that a well-briefed language model has largely erased. Inside companies the same erosion reaches the buyer whose staying put on knowing the supplier market by heart.
The loss carries a second cost that is harder to see. When everyone consults the same handful of models, the answers converge. In the BCG field experiment discussed further below, consultants working with GPT-4 produced ideas that were 41% less diverse than those of colleagues working without it. For anyone responsible for sourcing strategy that is an awkward finding. A category strategy drawn from the same model as the competitor’s may read convincingly, yet it is unlikely to give an edge over other players. The same applies for a dual/multi-sourcing strategy: if every business in an industry is moving towards the same three alternative suppliers, the alternative has reached capacity before the first purchase order is placed.
What the model forgets
In July 2024, researchers from Oxford, Cambridge, Imperial College, Toronto and Edinburgh published a study in Nature that business circles have largely passed over. They asked what happens when generative models are trained, generation after generation, on the output of earlier models. The fine details of the distribution go first. Rare events become under-represented and eventually vanish, while the common cases take over the model’s picture of the world. The authors call this model collapse, and they conclude that data from genuine human interaction will become more valuable as model-generated content spreads.
Every buyer and every planner knows the fine details by another name. It’s the part number nobody watched: the C-part with a single qualified source, the magnet buried in a bought-in motor, the export licence nobody needed until April 2025. Risk management in procurement and supply chain is the management of exactly these cases.
The procurement and supply chain literature written between 2000 and 2020 feeds a model for none of this. It describes a world of falling trade barriers, cheap energy, lean inventories and lowest-cost-country sourcing. A model trained on it will reproduce that world and with the greatest confidence on exactly the questions where that world in many areas no longer applies. Ask it for a safety stock policy or a sourcing region, and the answer will carry the logic of the years it learned from. Of course it all depends on the complexity and depth of the prompt which has been fed into the model.
Confidently wrong
The study most often quoted on AI in consulting is the field experiment that researchers from Harvard, Wharton, MIT and Warwick ran with 758 BCG consultants, now published in Organization Science. Its headline numbers are familiar: on tasks within the model’s capabilities, consultants using GPT-4 completed 12.2% more of their work and did so 25.1% faster.
For one task the researchers deliberately chose a business case in which the spreadsheet looked complete while the decisive clues were tucked away in interview notes with company insiders. Without AI, 84.5% of consultants reached the right recommendation. With GPT-4, between 60% and 70.6% did. Graders nevertheless rated the AI-assisted recommendations as more coherent and more persuasive, right or wrong. The group that had received prompt training did worst of all, which suggests that familiarity with the tool bred more reliance on it, not less.
AI is no less reliable than any other source. Every source gets things wrong now and then. The difference is that AI gets things wrong in flawless prose, and a polished report looks just as convincing when its conclusions are weak as when they are sound. Anyone who has been on a sourcing committee has seen the human version of this: the most polished supplier evaluation usually wins the room, whether or not it is the best one. With AI every evaluation is polished, so polish no longer tells you anything. The authors add that their participants were early in their careers and that results might differ for experienced managers. That caveat deserves attention: the study tested precisely the people who have not yet built the judgment this article is about.
The irony nobody budgets for
In 1983, the cognitive psychologist Lisanne Bainbridge published a five-page paper on automation in process plants and cockpits. Her point: automation doesn’t push people out of the system. It leaves them with the hardest part. The better the control system, the more depends on the operator in the rare moments it fails. And that operator has had almost no practice, because the system handled everything else.
Procurement and supply chain are about to repeat this. AI now does the market scans, quote comparisons and price analyses. Those were the tasks that taught junior buyers their job. Manual forecast corrections and stock reviews did the same for young supply chain planners.
After a few years of comparing offers line by line, a buyer can tell when a price is too good, when a supplier is overstretched, when a deal doesn’t add up. A planner who has chased enough missing parts knows which lead times in the system are wrong. Take that training away, and we lose the experienced category managers and planners that will be needed in ten years. AI will handle more by then, but not everything.
This is the uncomfortable part. Some of the efficiency AI promises is borrowed from future capability. A CPO or Head of Supply Chain who automates entry-level work without rethinking how people learn is making a decision about 2035, whether they mean to or not.
What stays scarce
When general answers are cheap, three things gain value. Hayek would have recognised all of them.
Local knowledge. Which of the two approved suppliers really has spare capacity? Which plant needs nine months to qualify a new source? Which carrier still has space when a port is congested? What does it mean when a supplier’s managing director suddenly asks for shorter payment terms? None of this is in a database. Most of it is never written down.
Two weeks before a major plant turnaround, the main contractor came to see me. The shutdown had been planned for months, with hundreds of external contractors and a client waiting. He told me his company had filed for insolvency. Nobody had seen it coming. The firm the whole plan was built around was gone.
A model would have given me a list of alternative contractors in seconds. It could not have made a single one pick up the phone. What saved the turnaround were relationships built over years: people who took my call, trusted me and sent crews at a notice no tender process allows. We went ahead with other contractors. None of what made that possible was available in a system.
Accountability. A supervisory board asked to move tens of millions in production volume to another country wants to know who stands behind the recommendation, and what that person has seen first-hand. A model can make an argument persuasive, as the BCG study shows. It cannot be held accountable.
Execution. Transformations rarely fail for lack of analysis. They get stuck in internal politics, in budgets nobody wants to fund, in a plant that won’t give up its buffer stock, in a business unit defending a supplier of twenty years of good cooperation. Getting an organisation to move still happens in meeting rooms, on the phone and on the shop floor.
None of this means seniority wins. People can be as outdated as models. A category manager who spent twenty-five years moving volume to low-cost countries, or a supply chain planner trained on ever leaner inventories, carries most probably an old world view too. It just has no cut-off date printed on it. Experience can turn into habit as easily as into judgment. The difference is whether someone questions their own knowledge and patterns when a rule that worked for twenty plus years is going to stop. A model can’t do that between training laps. A person can. Not every experienced person does, and so the premium will be just accessed by those who do.
What this means for CPO's and Supply Chain leaders
Procurement teams are already drawing the line, mostly by instinct. EFESO’s 2026 CPO Pulse interviewed 50 CPOs. They see the most value for generative AI in contract analysis (69%), supplier and market intelligence (61%) and automating sourcing processes (55%). Only 35% see value in AI-assisted negotiation. Just 5% have rolled the technology out at scale.
A deliberate split could look like this:
Hand over to AI | Keep with experienced people |
|---|---|
Contract review and summaries | Deciding which risks the business is prepared to carry |
Supplier and market scans | Judging a supplier’s actual resilience, ideally on site |
Spend classification and first should-cost models | Negotiation strategy and the relationship behind it |
Forecast baselines and freight rate comparisons | Deciding where to hold buffer stock, and who will be supplied first when material is short |
Monitoring news, sanctions lists and risk signals | Interpreting what a signal means for business operations |
First drafts of category strategies | Final sourcing and location decisions |
One rule is missing from most AI roadmaps: let junior buyers and supply chain planners do some of the analysis by hand for a while, then compare their results with the model’s. It takes time. It is also the only reliable way to develop people who can tell when the model is wrong.
Apply the same test to external advisory and consulting, ours included. Three questions help:
Which parts of your work come from AI, and who checked them against reality?
Where is the model’s answer misleading for our situation, and why?
Will you personally sign off the recommendation, and what would change your mind?
Who can I call from your network to verify your success story?
An adviser who answers the last question with a name has the kind of knowledge Hayek wrote about.
The judgment premium
When a technology makes something cheap, whatever complements it becomes more valuable. Spreadsheets made calculation almost free and easy to do. Knowing what was worth considering for the calculating became the scarce skill. AI is now doing the same analysis and knowledge. The premium moves to judgment: knowing where AI’s limits are, taking responsibility for decisions beyond them, and motivating an organisation to act.
PROCUREMENT reWIRED is built on that conviction. It combines AI-supported intelligence for breadth and speed with decades of hands-on experience in manufacturing procurement and supply chains, for the decisions where the data runs out.
Hayek’s man with the surplus stock is still out there. In times of structural change, his world moves faster than any model can be retrained. That is the best reason I know to keep asking him.
