By Jon W. Hansen, FCIPS | Procurement Insights | Hansen Models™
In a LinkedIn post this week, Sangeet Paul Choudary challenged one of the AI industry’s most reassuring promises: there will still be a human in the loop.
His argument, developed in the Harvard Business Review article “The Human-in-the-Loop Fallacy”, is that organizations are designing an increasingly marginal role for people. AI performs the work. Humans review its output, manage exceptions and retain whatever tasks the technology cannot yet perform.
As AI improves, that frontier contracts.
Choudary is right about the problem. A person employed as the machine’s permanent proofreader does not represent a compelling future of work. Reviewers may also lose the knowledge and attention needed to recognize errors as they become increasingly removed from performing the work themselves.
But the human in the loop is not the fallacy.
Putting the human at the end of the loop is.
The human should not receive a finished decision
Choudary offers four examples of organizations moving beyond task automation.
IKEA reportedly used a chatbot to handle millions of customer enquiries while retraining employees for remote interior-design work. Virtual production on The Mandalorian moved visual-effects decisions from post-production onto the set. Garfield Law used AI to make previously uneconomic small-debt cases commercially viable. John Deere reorganized around agricultural production systems rather than individual equipment lines.
These examples demonstrate that AI can change capabilities, markets, organizational structures and where decisions are made.
They do not demonstrate that the human has left the loop.
At IKEA, people moved into more complex customer work. On The Mandalorian, specialists influenced decisions earlier in production. Garfield Law still requires people to define acceptable legal outcomes, risk and accountability. John Deere’s leaders changed organizational authority and decision rights.
The humans did not disappear. Their position changed.
They moved from reviewing completed work to helping determine what work should be done, how decisions should be formed and which outcomes matter.
That is not the elimination of the human in the loop. It is the elevation of the human from reviewer to orchestrator.
Starting with the constraints AI removes is still starting with AI
Choudary contrasts two transformation approaches.
One begins with today’s work, automates tasks and assigns the remainder to people. The other begins with the constraints AI removes and asks what new decisions, capabilities, markets and structures become possible.
The second is unquestionably more ambitious. It still begins one step too late.
Before asking what constraints AI removes, the organization should establish what is producing its current outcomes.
That does not mean documenting the official process, mapping the workflow or asking people how work is supposed to happen. It means starting with the result being produced today and tracing backward through the people, incentives, decisions, informal handoffs, external agents and workarounds producing it.
Only then can the organization answer two questions:
What in this operation should an agent take over, and why?
What could an agent enhance or extend that the operation could not do before?
Without that grounding, organizations can automate the wrong work or build an entirely new capability on a false understanding of how the enterprise operates.
Every task succeeded. The operation failed.
In 1998, a Department of National Defence maintenance operation was achieving 51% next-day parts delivery against a 90% contractual requirement.
Every function appeared to be doing its job.
Technicians met their call-volume targets. Buyers processed the orders they received. Suppliers shipped. Couriers transported the parts. Finance followed its controls.
The operation still failed.
We began with the 51% outcome and traced backward. That revealed that technicians were holding parts orders until late afternoon. The practice helped them meet their own performance target, but it narrowed procurement’s supplier and transportation options, collided with courier cutoffs and customs clearance, and made a part costing $100 in the morning cost as much as $900 later in the day.
That condition did not appear in the documented process. No one had asked when the orders arrived. A task-level AI assessment would not have found it because each task, viewed independently, was being performed correctly.
The failure existed in the relationships between the tasks.
The Map You Design vs. The Map You Trace: the same system, mapped from two starting points. The designed map runs cleanly to its destination. The traced map starts from the failure, leaves the mapped process, and surfaces the obstacles the plan could not contain. (From The Map You Design and the Map You Trace, 29 July 2026.)
RAM 1998 was built only after that operating reality had been established. Self-learning algorithms ranked suppliers using continuously updated evidence. Buyers retained authority to change the weighting between price and delivery. Firm Fences defined what the system could do, when it had to stop or redirect an action and which human remained accountable.
Within three months, next-day delivery rose from 51% to 97.3%. It remained there for seven years.
The technology did not remove the human from the loop. It gave people a better loop to operate.
McKinsey’s enterprise second brain
McKinsey has now described an AI operating system that acts as an enterprise “second brain.”
Its AI OS model maintains enterprise context, reasons across functions, recommends or executes decisions and learns from outcomes. McKinsey also calls for explicit decisions about what agents may execute, what requires human judgment and what should escalate to senior leadership.
This is an important recognition: enterprise AI requires an operating architecture. A collection of models, agents and use cases will not govern itself.
But a second brain can still learn the wrong organization.
McKinsey describes companies redesigning workflows and then building digital twins, ontologies and semantic layers to codify processes, policies, relationships and institutional knowledge.
What happens when the institutional knowledge is incomplete?
What happens when the official workflow excludes the spreadsheet that actually runs the operation, the supplier contacted outside the system, the manager who manually reconciles conflicting records or the technician who holds orders until four o’clock to achieve their performance incentives?
I have made this point about digital twins before. In January, I wrote that a twin can only model what someone thought to include. What gets included, and what gets left out, often depends on the role of the person deciding. In July, I argued that without validation, a flawed understanding becomes a flawless, high-resolution model of the wrong organization. An AI operating system built on that twin inherits the same blind spot, and acts on it. The difference with AI is that the impact is immediate. In past technology eras, a flawed model took months to show up in results, and people had time to notice. An AI operating system acts on it at once.
The resulting AI OS may become process-model right and real-world wrong.
It can reason rapidly across functions, enforce its rules consistently and learn from every recorded outcome while remaining blind to the condition determining the result.
Shadow systems are evidence, not noise
McKinsey’s article comes closest to the issue in one of its questions for CEOs: what incentives and controls are needed to ensure that important decisions and actions happen inside the AI OS, “and not in organic shadow systems”?
McKinsey is right that incentives drive shadow systems. In 1998, that was the whole story. Technicians were measured on call volume, and stopping to submit a part order interrupted a service call, so they held their orders until late afternoon. A policy already told them to order immediately after each call. The measure overrode it. The control existed; the incentive won.
That is the risk in answering shadow systems with new incentives and controls. Until you trace what the existing incentives actually reward, you cannot know whether the new one removes the shadow system or simply creates the next one.
A shadow system is not noise to be pulled back inside the architecture. It is evidence of how the operation actually runs.
A second brain first needs operating reality
This is why Hansen ARA™ RAM 2025™ running on SLAP OS™ begins with Phase 0™.
How the two operating systems fit together. On the left, the operation runs as parallel strands. Phase 0™ traces backward from the outcome and brings to light the strands the documented map leaves out, incentives, shadow systems and workarounds, within a Firm Fence of named human authority; real outcomes loop back into it through the Infinity Loopback. That validated operating reality feeds McKinsey’s AI OS on the right, shown with its layers as McKinsey describes them. Illustrative.
The sequence is:
Start with the outcome being produced now.
Trace backward to establish the real operation.
Identify shadow functions, incentives, workarounds and external agents.
Establish the Firm Fence around that operating reality.
Use independent AI models to challenge the understanding.
Preserve disagreement through separate Main and Shadow panels.
Assign decision authority and escalation to a named Human Orchestrator.
Feed outcomes back through the Infinity Loopback without rewriting what the system originally understood.
SLAP OS™ does not assign the human the residual work AI cannot complete. It provides the operating protocol through which human and AI agents reason, disagree, act, stop and learn together.
And because the first brain is the organization’s own validated understanding of how it operates, not a platform, it is not tied to McKinsey’s architecture or anyone else’s. The Hansen AI OS is built to feed whichever AI operating systems an organization adopts, today’s and those still to come. McKinsey states the same principle as its first rule for building an AI OS: “Own the context; rent or replace the technology.” The context worth owning is the one that has been validated against how the operation actually runs.
The Human Orchestrator is not there to approve every answer. The role is to establish what the agents are acting on, decide what authority they have, recognize when evidence no longer supports the frame and remain accountable for the outcome.
That responsibility does not shrink as AI improves.
It becomes more consequential because the system can act faster, across more functions and with greater reach.
The human role that does disappear
One human role should diminish: the person waiting at the end of the process to approve or reject an answer they did not help frame and may no longer be qualified to assess.
That is not meaningful control. It is an accountability transfer.
The organization has allowed the system to define the question, assemble the evidence, perform the reasoning and propose the action. It then asks a person to absorb responsibility by clicking approve.
Choudary is right to reject that future.
The answer is not to remove the human from the loop. It is to stop treating the human as the last control attached to a loop designed somewhere else.
What humans can now do
AI allows people to recognize and process more operating strands than any individual could follow alone. It can test competing interpretations, monitor changing conditions, surface disagreements and carry evidence from one decision into the next.
That creates capabilities previously beyond practical reach.
But technology cannot determine by itself whether the enterprise’s understanding reflects the real operation. It cannot assume accountability for the outcome. It cannot decide which contradiction matters when the record and the world no longer agree.
Those remain human responsibilities.
The future is not humans performing the shrinking residue of work machines cannot do.
It is humans establishing the operating reality, setting the purpose and authority of the system, and using multiple forms of intelligence to achieve outcomes neither could reliably produce alone.
Human-as-checker is a diminishing role. Human-as-orchestrator is not.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
How the Hansen-McKinsey AI Operating Systems Address Choudary’s “Human-in-the-Loop Fallacy.”
Posted on November 4, 2026
0
By Jon W. Hansen, FCIPS | Procurement Insights | Hansen Models™
In a LinkedIn post this week, Sangeet Paul Choudary challenged one of the AI industry’s most reassuring promises: there will still be a human in the loop.
His argument, developed in the Harvard Business Review article “The Human-in-the-Loop Fallacy”, is that organizations are designing an increasingly marginal role for people. AI performs the work. Humans review its output, manage exceptions and retain whatever tasks the technology cannot yet perform.
As AI improves, that frontier contracts.
Choudary is right about the problem. A person employed as the machine’s permanent proofreader does not represent a compelling future of work. Reviewers may also lose the knowledge and attention needed to recognize errors as they become increasingly removed from performing the work themselves.
But the human in the loop is not the fallacy.
Putting the human at the end of the loop is.
The human should not receive a finished decision
Choudary offers four examples of organizations moving beyond task automation.
IKEA reportedly used a chatbot to handle millions of customer enquiries while retraining employees for remote interior-design work. Virtual production on The Mandalorian moved visual-effects decisions from post-production onto the set. Garfield Law used AI to make previously uneconomic small-debt cases commercially viable. John Deere reorganized around agricultural production systems rather than individual equipment lines.
These examples demonstrate that AI can change capabilities, markets, organizational structures and where decisions are made.
They do not demonstrate that the human has left the loop.
At IKEA, people moved into more complex customer work. On The Mandalorian, specialists influenced decisions earlier in production. Garfield Law still requires people to define acceptable legal outcomes, risk and accountability. John Deere’s leaders changed organizational authority and decision rights.
The humans did not disappear. Their position changed.
They moved from reviewing completed work to helping determine what work should be done, how decisions should be formed and which outcomes matter.
That is not the elimination of the human in the loop. It is the elevation of the human from reviewer to orchestrator.
Starting with the constraints AI removes is still starting with AI
Choudary contrasts two transformation approaches.
One begins with today’s work, automates tasks and assigns the remainder to people. The other begins with the constraints AI removes and asks what new decisions, capabilities, markets and structures become possible.
The second is unquestionably more ambitious. It still begins one step too late.
Before asking what constraints AI removes, the organization should establish what is producing its current outcomes.
That does not mean documenting the official process, mapping the workflow or asking people how work is supposed to happen. It means starting with the result being produced today and tracing backward through the people, incentives, decisions, informal handoffs, external agents and workarounds producing it.
Only then can the organization answer two questions:
Without that grounding, organizations can automate the wrong work or build an entirely new capability on a false understanding of how the enterprise operates.
Every task succeeded. The operation failed.
In 1998, a Department of National Defence maintenance operation was achieving 51% next-day parts delivery against a 90% contractual requirement.
Every function appeared to be doing its job.
Technicians met their call-volume targets. Buyers processed the orders they received. Suppliers shipped. Couriers transported the parts. Finance followed its controls.
The operation still failed.
We began with the 51% outcome and traced backward. That revealed that technicians were holding parts orders until late afternoon. The practice helped them meet their own performance target, but it narrowed procurement’s supplier and transportation options, collided with courier cutoffs and customs clearance, and made a part costing $100 in the morning cost as much as $900 later in the day.
That condition did not appear in the documented process. No one had asked when the orders arrived. A task-level AI assessment would not have found it because each task, viewed independently, was being performed correctly.
The failure existed in the relationships between the tasks.
The Map You Design vs. The Map You Trace: the same system, mapped from two starting points. The designed map runs cleanly to its destination. The traced map starts from the failure, leaves the mapped process, and surfaces the obstacles the plan could not contain. (From The Map You Design and the Map You Trace, 29 July 2026.)
RAM 1998 was built only after that operating reality had been established. Self-learning algorithms ranked suppliers using continuously updated evidence. Buyers retained authority to change the weighting between price and delivery. Firm Fences defined what the system could do, when it had to stop or redirect an action and which human remained accountable.
Within three months, next-day delivery rose from 51% to 97.3%. It remained there for seven years.
The technology did not remove the human from the loop. It gave people a better loop to operate.
McKinsey’s enterprise second brain
McKinsey has now described an AI operating system that acts as an enterprise “second brain.”
Its AI OS model maintains enterprise context, reasons across functions, recommends or executes decisions and learns from outcomes. McKinsey also calls for explicit decisions about what agents may execute, what requires human judgment and what should escalate to senior leadership.
This is an important recognition: enterprise AI requires an operating architecture. A collection of models, agents and use cases will not govern itself.
But a second brain can still learn the wrong organization.
McKinsey describes companies redesigning workflows and then building digital twins, ontologies and semantic layers to codify processes, policies, relationships and institutional knowledge.
What happens when the institutional knowledge is incomplete?
What happens when the official workflow excludes the spreadsheet that actually runs the operation, the supplier contacted outside the system, the manager who manually reconciles conflicting records or the technician who holds orders until four o’clock to achieve their performance incentives?
I have made this point about digital twins before. In January, I wrote that a twin can only model what someone thought to include. What gets included, and what gets left out, often depends on the role of the person deciding. In July, I argued that without validation, a flawed understanding becomes a flawless, high-resolution model of the wrong organization. An AI operating system built on that twin inherits the same blind spot, and acts on it. The difference with AI is that the impact is immediate. In past technology eras, a flawed model took months to show up in results, and people had time to notice. An AI operating system acts on it at once.
The resulting AI OS may become process-model right and real-world wrong.
It can reason rapidly across functions, enforce its rules consistently and learn from every recorded outcome while remaining blind to the condition determining the result.
Shadow systems are evidence, not noise
McKinsey’s article comes closest to the issue in one of its questions for CEOs: what incentives and controls are needed to ensure that important decisions and actions happen inside the AI OS, “and not in organic shadow systems”?
McKinsey is right that incentives drive shadow systems. In 1998, that was the whole story. Technicians were measured on call volume, and stopping to submit a part order interrupted a service call, so they held their orders until late afternoon. A policy already told them to order immediately after each call. The measure overrode it. The control existed; the incentive won.
That is the risk in answering shadow systems with new incentives and controls. Until you trace what the existing incentives actually reward, you cannot know whether the new one removes the shadow system or simply creates the next one.
A shadow system is not noise to be pulled back inside the architecture. It is evidence of how the operation actually runs.
A second brain first needs operating reality
This is why Hansen ARA™ RAM 2025™ running on SLAP OS™ begins with Phase 0™.
How the two operating systems fit together. On the left, the operation runs as parallel strands. Phase 0™ traces backward from the outcome and brings to light the strands the documented map leaves out, incentives, shadow systems and workarounds, within a Firm Fence of named human authority; real outcomes loop back into it through the Infinity Loopback. That validated operating reality feeds McKinsey’s AI OS on the right, shown with its layers as McKinsey describes them. Illustrative.
The sequence is:
SLAP OS™ does not assign the human the residual work AI cannot complete. It provides the operating protocol through which human and AI agents reason, disagree, act, stop and learn together.
And because the first brain is the organization’s own validated understanding of how it operates, not a platform, it is not tied to McKinsey’s architecture or anyone else’s. The Hansen AI OS is built to feed whichever AI operating systems an organization adopts, today’s and those still to come. McKinsey states the same principle as its first rule for building an AI OS: “Own the context; rent or replace the technology.” The context worth owning is the one that has been validated against how the operation actually runs.
The Human Orchestrator is not there to approve every answer. The role is to establish what the agents are acting on, decide what authority they have, recognize when evidence no longer supports the frame and remain accountable for the outcome.
That responsibility does not shrink as AI improves.
It becomes more consequential because the system can act faster, across more functions and with greater reach.
The human role that does disappear
One human role should diminish: the person waiting at the end of the process to approve or reject an answer they did not help frame and may no longer be qualified to assess.
That is not meaningful control. It is an accountability transfer.
The organization has allowed the system to define the question, assemble the evidence, perform the reasoning and propose the action. It then asks a person to absorb responsibility by clicking approve.
Choudary is right to reject that future.
The answer is not to remove the human from the loop. It is to stop treating the human as the last control attached to a loop designed somewhere else.
What humans can now do
AI allows people to recognize and process more operating strands than any individual could follow alone. It can test competing interpretations, monitor changing conditions, surface disagreements and carry evidence from one decision into the next.
That creates capabilities previously beyond practical reach.
But technology cannot determine by itself whether the enterprise’s understanding reflects the real operation. It cannot assume accountability for the outcome. It cannot decide which contradiction matters when the record and the world no longer agree.
Those remain human responsibilities.
The future is not humans performing the shrinking residue of work machines cannot do.
It is humans establishing the operating reality, setting the purpose and authority of the system, and using multiple forms of intelligence to achieve outcomes neither could reliably produce alone.
Human-as-checker is a diminishing role. Human-as-orchestrator is not.
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related