Double marginalization explains why deploying agents one function at a time can leave the enterprise worse off — with every agent performing correctly.
There is a stretch of the Rhine that tourists sail down for the castles. They are photographed from the deck as romantic ruins, and most people assume they were the country houses of the rich.
They were toll booths.
Every few miles a different prince held a different castle, and every one of them charged boats passing through. Each toll was set the way any rational owner would set it: high enough to maximize what that castle earned from the traffic it saw.
And here is the part that matters. The upstream prince had no reason to care what his toll did to the downstream prince’s revenue. Charging more meant fewer boats made the full journey, which cost the castles further along. But that cost landed on somebody else’s ledger, so nothing in the upstream prince’s decision accounted for it.
The downstream prince faced the identical situation in reverse.
So the tolls were collectively set too high. Not merely from the point of view of the traders — too high from the point of view of the princes themselves. A single owner holding the whole river would have charged less in total, moved far more traffic, and earned more than all of them earned separately.
That is double marginalization. When independent holders each add a margin to the same journey, the combined charge can exceed what a coordinated owner would set, leaving both total traffic and combined profit lower.
Nobody misbehaved
I want to hold on that, because it is the whole reason the problem survives.
No prince was greedy in any sense that a contemporary would have recognized. Each set a price that was correct given what he could see. Each was optimizing the thing he was accountable for. Each would have shown a defensible number if anyone had audited his castle in isolation.
The failure was not in any castle. It was in the fact that no one held the whole river — and the loss existed only in the relationship between the tolls, which appeared on no one’s accounts.
An audit of any single castle would have found nothing wrong. An audit of all of them, one at a time, would have found nothing wrong either.
Where I have run into a related structure
Before going further I want to be precise, because the next two examples are not textbook double marginalization and I am not going to claim they are. Neither involves successive monopolists adding markups along a single chain. What they share with the Rhine is the thing that matters here: locally rational commercial decisions whose combined effect left the enterprise paying more than the system required, with the loss appearing on nobody’s ledger.
The first time I met that structure I had no name for it.
In 2005 I published a paper on low-dollar, high-volume spend, and one line in it has aged better than anything else I wrote that year. It records the standing complaint of front-line buyers: that they can usually get better pricing with one phone call than by purchasing off a centrally negotiated contract.
Every category manager has heard that, and most treat it as an excuse for maverick spend. Read it as a report from the only person standing where both prices are visible.
Say the contract price is $95 and the local supplier is $80. Two things can happen, and neither of them puts that $15 into the record.
If the buyer complies, the organization pays $95. The contract was honored, the process was followed, and the reporting is clean. The $15 does not exist as data anywhere, because the alternative was never priced.
If the buyer picks up the phone, the organization saves $15 and the buyer is in violation. That saving cannot be recorded either — recording it means recording the breach. So it surfaces as a compliance exception, and the only number that reaches management is the violation count. The maverick spend gets measured. The saving that motivated it does not.
Either way the instrument reads the same: clean compliance, or a compliance problem. Never the contract is fifteen dollars above market on this item.
That is the shape of the whole problem in a single transaction. The person positioned to see the gap is the one person whose seeing it cannot be entered anywhere — so the organization stays machine-smart and process-blind. Its systems are correct, its reports are clean, and the thing determining the outcome never reaches the record.
The same year, a major US retailer engaged me to assess a procurement strategy that was not delivering. They had consolidated their supplier base under a leveraged-spend program. Suppliers cut price hard to win a slot, the terms were locked, and the savings reporting was clean.
They were paying 21 percent above market — because the commodities in question declined steadily in price while the contract held its number still. The suppliers who understood that said nothing, which was not misconduct. Their obligation was the negotiated price, not the real-time market price. Honoring the agreement and staying silent were the same act.
Each party performed exactly as agreed. The loss lived between them, on nobody’s ledger, and the reporting stayed clean throughout.
Why this gets worse the further upstream it sits
There is a corollary in the economics that is worth stating on its own, because it is the part that applies to what is being built right now.
A monopoly on a consumer good is a contained problem. A monopoly on an intermediate good — something used to produce other things — is not, because the inefficiency propagates into everything downstream of it. Every sector that touches it inherits the distortion. Nobody in those sectors did anything wrong; they simply bought their input at a price that was set without reference to what it would cost the rest of the chain.
The more layers a thing passes through, the more the small local rationalities compound.
Which brings me to the thing I actually want to write about.
Decisions are becoming intermediate goods
Enterprise AI is being deployed the way tolls were built on the Rhine: one function at a time, each with its own sponsor, its own objective, and its own definition of success.
A sourcing agent optimized for cost. A logistics agent optimized for on-time delivery. A finance agent optimized for working capital. A demand agent optimized for forecast accuracy. Each one procured separately, often from a different provider, each with its own consultant and its own view of what good looks like.
Assume every one of them performs exactly as designed. That is where this becomes interesting — because the problem does not require any of them to fail.
And the output of each is an input to the others. A sourcing decision constrains what logistics can achieve. A working capital target constrains what sourcing can commit to. A forecast shapes both. These are not independent optimizations sitting side by side. They are tolls on the same river.
Which means the structure from the castles carries over, even though the mechanism is not a price markup. An agent optimizing correctly against the objective it was given imposes costs that land on a function it cannot see and is not accountable for. The aggregate can be worse than a single coordinated decision would have produced — and worse for the functions individually, not only for the enterprise in the abstract.
And notice what has changed since the phone call. A buyer facing a $95 contract and an $80 local price at least had the option to break the rule and save the money. An agent given the contract as its reference has no phone. It complies, correctly, at $95 — and the fifteen dollars is now invisible by construction, with nobody left to be tempted into noticing it.
The more independently optimized layers a decision passes through, the greater the opportunity for local rationalities to compound. Which is why adding agents does not reliably improve the aggregate — and why a better local agent does not resolve a cross-agent conflict unless something in the architecture represents the shared outcome.
The reason it will not appear in any report
Here is what makes the agentic version harder than the medieval one.
Ask each agent how it performed and each will report success, accurately. Ask each function whether its objective was met and the answer will be yes. Run a governance review of any single deployment and it will pass, because nothing in it malfunctioned.
The loss exists only in the relationships between the decisions, and there is no instrument pointed at the relationships. Every instrument is pointed at a castle.
Which is a specific case of something I have written about recently: a condition where the reporting is clean because the mechanism capable of registering the problem was never built, not because the problem is absent. Objective Flatlining™ — no variance in the record, and no way to tell from the record whether that means alignment or blindness.
And unlike an overrun, it never runs out of money and forces a reckoning. It just quietly costs more every year while every dashboard stays green.
The answer is structural, and I wrote it down in 2007
The economics has a clean solution: put the river under one owner. Vertical integration of two monopolies genuinely improves the outcome, including for the monopolists — one of the few results in the field where consolidation helps everybody.
Enterprises cannot merge their functions, and shouldn’t. But the underlying requirement transfers, and it is not satisfied by a governance layer added over the top of the agents after deployment. Governance is necessary. It cannot create the cross-functional view it would need in order to govern.
The requirement is not one giant agent or one giant system. It is a shared view of the conditions that each decision changes for the others.
In August 2007 I published a post here on double marginalization. Its closing paragraph begins: “With the advent of agent-based modeling and the emergence of the meta-enterprise application, we are entering a period of discontinuous innovation.”
I want to be careful about what that does and does not establish. Agent-based modeling in 2007 was not agentic AI. It described the simulation of autonomous decision units acting on local conditions, out of complexity economics. The term is the same, the meaning has evolved, and the technology is not.
What carries across is the structural intuition, and that is the part worth having. Distributed units, each deciding against the conditions in front of it, require a synchronized rather than sequential architecture — otherwise the decisions stack the way the tolls stacked. That is what I described as a Metaprise™: an architecture linking the operating attributes of all transactional stakeholders on a real-world, real-time basis. Not integration in the systems sense. A buyer able to see the conditions facing suppliers, couriers and customs brokers at the same time, because a decision made against any one of them in isolation is a toll set without reference to the rest of the river.
The same body of work described a cross-verification mechanism: the buyer could check a decision against a second category of evidence rather than only against the contract they were working inside. That is the missing instrument in the retailer case, stated nineteen years before this post.
Neither of those is a governance framework. Both are structural answers to a structural problem, which is what double marginalization requires — because you cannot fix it by making any individual toll more careful.
So the two halves of this argument were sitting in the same 2007 document: the economics of stacked margins, and distributed decision units needing a synchronized architecture. What was missing then was the deployment. The agents have now arrived to fill a slot the analysis had already described.
What to ask before the next agent goes in
Not is this agent well governed? It will be.
Not does this agent meet its objective? It will.
Whose ledger absorbs the cost of this agent meeting its objective — and does anybody hold a view that spans both?
If the answer to the second half is no, you are not deploying an agent. You are building another castle on a river nobody owns, and the traffic will keep falling while every toll-keeper reports an excellent year.
A note on the sources. Double Marginalization and the Decentralized Supply Chain, 9 August 2007, and its companion the following day on the Point of Ideal Price Viability. I returned to the subject in 2023 in the context of data management and the Metaprise™. All remain as published.
-30-
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Related
Double Marginalization and the Deployment of AI Agents
Posted on August 11, 2026
0
Double marginalization explains why deploying agents one function at a time can leave the enterprise worse off — with every agent performing correctly.
There is a stretch of the Rhine that tourists sail down for the castles. They are photographed from the deck as romantic ruins, and most people assume they were the country houses of the rich.
They were toll booths.
Every few miles a different prince held a different castle, and every one of them charged boats passing through. Each toll was set the way any rational owner would set it: high enough to maximize what that castle earned from the traffic it saw.
And here is the part that matters. The upstream prince had no reason to care what his toll did to the downstream prince’s revenue. Charging more meant fewer boats made the full journey, which cost the castles further along. But that cost landed on somebody else’s ledger, so nothing in the upstream prince’s decision accounted for it.
The downstream prince faced the identical situation in reverse.
So the tolls were collectively set too high. Not merely from the point of view of the traders — too high from the point of view of the princes themselves. A single owner holding the whole river would have charged less in total, moved far more traffic, and earned more than all of them earned separately.
That is double marginalization. When independent holders each add a margin to the same journey, the combined charge can exceed what a coordinated owner would set, leaving both total traffic and combined profit lower.
Nobody misbehaved
I want to hold on that, because it is the whole reason the problem survives.
No prince was greedy in any sense that a contemporary would have recognized. Each set a price that was correct given what he could see. Each was optimizing the thing he was accountable for. Each would have shown a defensible number if anyone had audited his castle in isolation.
The failure was not in any castle. It was in the fact that no one held the whole river — and the loss existed only in the relationship between the tolls, which appeared on no one’s accounts.
An audit of any single castle would have found nothing wrong. An audit of all of them, one at a time, would have found nothing wrong either.
Where I have run into a related structure
Before going further I want to be precise, because the next two examples are not textbook double marginalization and I am not going to claim they are. Neither involves successive monopolists adding markups along a single chain. What they share with the Rhine is the thing that matters here: locally rational commercial decisions whose combined effect left the enterprise paying more than the system required, with the loss appearing on nobody’s ledger.
The first time I met that structure I had no name for it.
In 2005 I published a paper on low-dollar, high-volume spend, and one line in it has aged better than anything else I wrote that year. It records the standing complaint of front-line buyers: that they can usually get better pricing with one phone call than by purchasing off a centrally negotiated contract.
Every category manager has heard that, and most treat it as an excuse for maverick spend. Read it as a report from the only person standing where both prices are visible.
Say the contract price is $95 and the local supplier is $80. Two things can happen, and neither of them puts that $15 into the record.
If the buyer complies, the organization pays $95. The contract was honored, the process was followed, and the reporting is clean. The $15 does not exist as data anywhere, because the alternative was never priced.
If the buyer picks up the phone, the organization saves $15 and the buyer is in violation. That saving cannot be recorded either — recording it means recording the breach. So it surfaces as a compliance exception, and the only number that reaches management is the violation count. The maverick spend gets measured. The saving that motivated it does not.
Either way the instrument reads the same: clean compliance, or a compliance problem. Never the contract is fifteen dollars above market on this item.
That is the shape of the whole problem in a single transaction. The person positioned to see the gap is the one person whose seeing it cannot be entered anywhere — so the organization stays machine-smart and process-blind. Its systems are correct, its reports are clean, and the thing determining the outcome never reaches the record.
The same year, a major US retailer engaged me to assess a procurement strategy that was not delivering. They had consolidated their supplier base under a leveraged-spend program. Suppliers cut price hard to win a slot, the terms were locked, and the savings reporting was clean.
They were paying 21 percent above market — because the commodities in question declined steadily in price while the contract held its number still. The suppliers who understood that said nothing, which was not misconduct. Their obligation was the negotiated price, not the real-time market price. Honoring the agreement and staying silent were the same act.
Each party performed exactly as agreed. The loss lived between them, on nobody’s ledger, and the reporting stayed clean throughout.
Why this gets worse the further upstream it sits
There is a corollary in the economics that is worth stating on its own, because it is the part that applies to what is being built right now.
A monopoly on a consumer good is a contained problem. A monopoly on an intermediate good — something used to produce other things — is not, because the inefficiency propagates into everything downstream of it. Every sector that touches it inherits the distortion. Nobody in those sectors did anything wrong; they simply bought their input at a price that was set without reference to what it would cost the rest of the chain.
The more layers a thing passes through, the more the small local rationalities compound.
Which brings me to the thing I actually want to write about.
Decisions are becoming intermediate goods
Enterprise AI is being deployed the way tolls were built on the Rhine: one function at a time, each with its own sponsor, its own objective, and its own definition of success.
A sourcing agent optimized for cost. A logistics agent optimized for on-time delivery. A finance agent optimized for working capital. A demand agent optimized for forecast accuracy. Each one procured separately, often from a different provider, each with its own consultant and its own view of what good looks like.
Assume every one of them performs exactly as designed. That is where this becomes interesting — because the problem does not require any of them to fail.
And the output of each is an input to the others. A sourcing decision constrains what logistics can achieve. A working capital target constrains what sourcing can commit to. A forecast shapes both. These are not independent optimizations sitting side by side. They are tolls on the same river.
Which means the structure from the castles carries over, even though the mechanism is not a price markup. An agent optimizing correctly against the objective it was given imposes costs that land on a function it cannot see and is not accountable for. The aggregate can be worse than a single coordinated decision would have produced — and worse for the functions individually, not only for the enterprise in the abstract.
And notice what has changed since the phone call. A buyer facing a $95 contract and an $80 local price at least had the option to break the rule and save the money. An agent given the contract as its reference has no phone. It complies, correctly, at $95 — and the fifteen dollars is now invisible by construction, with nobody left to be tempted into noticing it.
The more independently optimized layers a decision passes through, the greater the opportunity for local rationalities to compound. Which is why adding agents does not reliably improve the aggregate — and why a better local agent does not resolve a cross-agent conflict unless something in the architecture represents the shared outcome.
The reason it will not appear in any report
Here is what makes the agentic version harder than the medieval one.
Ask each agent how it performed and each will report success, accurately. Ask each function whether its objective was met and the answer will be yes. Run a governance review of any single deployment and it will pass, because nothing in it malfunctioned.
The loss exists only in the relationships between the decisions, and there is no instrument pointed at the relationships. Every instrument is pointed at a castle.
Which is a specific case of something I have written about recently: a condition where the reporting is clean because the mechanism capable of registering the problem was never built, not because the problem is absent. Objective Flatlining™ — no variance in the record, and no way to tell from the record whether that means alignment or blindness.
And unlike an overrun, it never runs out of money and forces a reckoning. It just quietly costs more every year while every dashboard stays green.
The answer is structural, and I wrote it down in 2007
The economics has a clean solution: put the river under one owner. Vertical integration of two monopolies genuinely improves the outcome, including for the monopolists — one of the few results in the field where consolidation helps everybody.
Enterprises cannot merge their functions, and shouldn’t. But the underlying requirement transfers, and it is not satisfied by a governance layer added over the top of the agents after deployment. Governance is necessary. It cannot create the cross-functional view it would need in order to govern.
The requirement is not one giant agent or one giant system. It is a shared view of the conditions that each decision changes for the others.
In August 2007 I published a post here on double marginalization. Its closing paragraph begins: “With the advent of agent-based modeling and the emergence of the meta-enterprise application, we are entering a period of discontinuous innovation.”
I want to be careful about what that does and does not establish. Agent-based modeling in 2007 was not agentic AI. It described the simulation of autonomous decision units acting on local conditions, out of complexity economics. The term is the same, the meaning has evolved, and the technology is not.
What carries across is the structural intuition, and that is the part worth having. Distributed units, each deciding against the conditions in front of it, require a synchronized rather than sequential architecture — otherwise the decisions stack the way the tolls stacked. That is what I described as a Metaprise™: an architecture linking the operating attributes of all transactional stakeholders on a real-world, real-time basis. Not integration in the systems sense. A buyer able to see the conditions facing suppliers, couriers and customs brokers at the same time, because a decision made against any one of them in isolation is a toll set without reference to the rest of the river.
The same body of work described a cross-verification mechanism: the buyer could check a decision against a second category of evidence rather than only against the contract they were working inside. That is the missing instrument in the retailer case, stated nineteen years before this post.
Neither of those is a governance framework. Both are structural answers to a structural problem, which is what double marginalization requires — because you cannot fix it by making any individual toll more careful.
So the two halves of this argument were sitting in the same 2007 document: the economics of stacked margins, and distributed decision units needing a synchronized architecture. What was missing then was the deployment. The agents have now arrived to fill a slot the analysis had already described.
What to ask before the next agent goes in
Not is this agent well governed? It will be.
Not does this agent meet its objective? It will.
Whose ledger absorbs the cost of this agent meeting its objective — and does anybody hold a view that spans both?
If the answer to the second half is no, you are not deploying an agent. You are building another castle on a river nobody owns, and the traffic will keep falling while every toll-keeper reports an excellent year.
A note on the sources. Double Marginalization and the Decentralized Supply Chain, 9 August 2007, and its companion the following day on the Point of Ideal Price Viability. I returned to the subject in 2023 in the context of data management and the Metaprise™. All remain as published.
-30-
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Share this:
Like this:
Related