Cyborgs, Centaurs, and Self-Automators are already in your workforce. The question is which org structure you are building around them.
AI agents are not replacing your workforce one for one. They are splitting it into two tiers: people who judge and direct AI output, and people who operate AI-augmented workflows. New research now measures the split. In a field study of 244 Boston Consulting Group consultants, Randazzo, Lakhani, and colleagues found three modes of working with AI, and only two of them build skill. The bifurcation is real, it is happening now, and its shape is a design choice. This piece lays out the Two-Track Architecture for designing it deliberately, and the Verification Premium for compensating it correctly.
Everyone is asking whether AI will replace people. On the ground, that is not what is happening. AI agents are not replacing workers one for one. They are restructuring the work itself, and they are doing it faster than most HR functions are prepared to absorb.
The restructure has a shape. As agents take on more of the doing, the human role splits in two. Some people become judges of agent output. They verify, approve, escalate, and intervene. Others become operators of agent-augmented workflows. They direct agents toward tasks, package the outputs, and move them into business processes. A third pattern hides inside the second: workers who hand the task to the agent and stop developing both their judgment and their craft.
The question for HR is no longer whether the split will happen. It is what kind of split your company ends up with, and who lands on which side.
The bifurcation is not coming. It is here. The question is whether you design it or inherit it.
Until recently, the bifurcation argument was inference. We drew it from decades of automation research and early deployment patterns. As of late 2025, it is measured.
Steven Randazzo, Karim Lakhani, Katherine Kellogg, Fabrizio Dell’Acqua, Ethan Mollick, and colleagues ran a field study of 244 Boston Consulting Group consultants moving through a real AI-assisted decision workflow. They found three distinct modes of working with AI, and they named them.
Cyborgs fuse with the AI across the whole workflow. They assign it personas, iterate continuously, and use it to shape both the problem and the solution. About 60% of the consultants worked this way. Crucially, Cyborgs were newskilling. They came out of the work with a new area of AI expertise they did not have before.
Centaurs keep a clear line between themselves and the AI. They own the problem definition and the direction, and they hand specific, bounded steps to the AI. Centaurs were upskilling. They got sharper at the domain expertise they already had.
Self-Automators offload the task more fully. They pass work to the AI and integrate the output back into the deliverable with limited further engagement. And here is the finding that should stop every HR leader: Self-Automators increased neither their domain expertise nor their AI expertise. Both skill trajectories were flat.
Self-Automators do not increase domain expertise. They do not develop AI expertise. Both skill trajectories are flat.
That is the bifurcation, measured. Across the same nominally identical human-in-the-loop workflow, three behaviors emerge and only two of them build capability. If your company is deploying AI agents, you already have all three modes in your workforce, in proportions you have almost certainly not measured.
The same pattern shows up in the psychology. Guy Champniss, writing in Harvard Business Review, surveyed more than 1,200 employees across ten sectors in the US and UK and identified what he calls competency debt: the sense that the more you use AI, the less competent you become. But his data complicates the simple story. He found that psychological debt was almost twice as high among people who used AI rarely, a score of 60, as among people who used it many times a day, a score of 36, and higher among those who used it only for simple tasks, a score of 46, than among those who used it for complex and strategic work, who scored 35. The damage does not come from using AI. It comes from using it shallowly. Champniss arrives from psychology where Randazzo arrives from skill measurement, and they converge on the same conclusion: the mode of use, not the fact of use, determines whether your people grow or atrophy.
The organizational-level data says the same thing from above. In a 2026 HBR Analytic Services survey of 385 decision-makers, 59% had AI in production, but only 35% could point to measurable ROI. High deployment, low value capture. That gap is the bifurcation playing out at company scale: organizations are deploying AI without redesigning the work around it, so they capture the productivity of the operator tier and miss the value that only a deliberate judge tier produces.
Here is where HR has a decision to make.
The realist read says the bifurcation happens automatically at every company that deploys agents, so HR’s job is to react to it. The more useful read, and the one I will argue for, is that the shape of the split is something you build. The comp differential between the tiers, the mobility between them, the skills you require of each, the escalation logic that connects them, every one of those is a design decision. You can make it deliberately, or you can let it happen to you.
Aviation made this exact decision forty years ago. When automation arrived in commercial cockpits, every airline faced the same fork: pilots could become judges of the automation or operators of it. The airlines that thrived designed the role on purpose. They specified what pilots had to verify, what they had to monitor, what they had to be able to do by hand when the automation failed, and how training had to change to keep those skills alive. The airlines that drifted let the role hollow out, and the consequences showed up in accident investigations for decades. The FAA’s own Flight Deck Automation review documented the pattern in detail.
Lisanne Bainbridge anticipated all of it in 1983, in a paper she titled “Ironies of Automation,” a phrase the field has used ever since. Her insight was that the more automation you add, the more skilled the supervising human has to be, because the situations the human is left to handle are by definition the ones the automation could not. An undesigned bifurcation produces the opposite. It lets the judge’s skills decay precisely when you will need them most.
The shape of the bifurcation is a design decision. The airlines that won designed it. The ones that drifted paid for it in accident investigations.
The same choice is playing out in boardrooms right now. Jan-Emmanuel De Neve of Oxford, with Jeffrey Hancock of Stanford and Kate Niederhoffer of BetterUp, surveyed 1,294 desk workers across the US, Canada, and the UK and watched the fork in real time. On the automation side, Jack Dorsey cut nearly half of Block’s workforce, writing that intelligence tools had changed what it means to run a company, and Goldman projects an 11% decline in headcount across its investment-banking clients over three years. On the augmentation side, Aon committed to AI literacy for its 60,000 employees, and Microsoft under Nadella paired its cloud and AI transformation with a company-wide reskilling investment. Automation is the path of contraction. Augmentation is the path of expansion. The bifurcation you design depends on which one you choose.
The mechanism that makes the bifurcation real inside a workflow is what Nate Jones has been describing as the Judge Layer. It is the structural layer that sits between an AI agent’s output and the action that output triggers, where verification, approval, escalation, and judgment live. Without it, agent errors flow downstream uncorrected. With it, agents become safe to deploy at scale.
The judge layer is not necessarily human. It can be an AI judge, and often should be, with one firm rule: the model that produces the work cannot be the model that judges it. Separation is the point. But even an automated judge layer needs humans above it, owning the cases that carry real consequence and setting the standard the AI judge is held to.
Nate’s writing on this is the deepest operational treatment I have seen, and anyone building the verification function inside their company should read it in full here: You gave your AI agent real tools. Here’s the 4-part control layer it’s missing + the Judge Layer implementation guide.
This piece takes the Judge Layer as the operational given and runs the question that follows from it. If the Judge Layer is the mechanism, what is the human staffing model that makes it work? Who sits in that layer, who sets its standards, how are they developed, and how are they paid? That is the bifurcation question, and it is squarely HR’s to answer.
The Randazzo modes describe how individuals behave. The HR design question runs one level up: what organizational structure should contain those modes, and what compensation structure aligns the modes you actually want to grow? Two frameworks.
The Two-Track Architecture is the deliberate design of the judge tier and the operator tier as distinct tracks, with defined movement between them. It is not an org chart you draw once. It is six design decisions you make on purpose.
Role definition: what each tier actually does, written in language a recruiter could post. Cyborgs and Centaurs both belong in the judge track, in different sub-modes. Self-Automators are the warning sign. A role that produces Self-Automator behavior is a role that should be redesigned or moved onto a development track, not left to atrophy.
Decision rights: what each tier can approve, escalate, override, or hold. This maps to the two questions the Randazzo team found at the heart of every human-AI collaboration: who decides what needs doing, and who decides how it gets done.
Get Annette Snyder, The Workforce Paradox’s stories in your inbox
Join Medium for free to get updates from this writer.
Comp differential: the gap between the tiers, which the next framework addresses directly.
Mobility path: how an operator becomes a judge, and how you avoid building judges who never operated and therefore cannot verify.
Skill currency: the practice and re-certification judges need to keep the manual capability the role occasionally demands. This is the aviation lesson applied to knowledge work.
Escalation logic: when a judge escalates to a human with deeper context, and who that person is.
The Verification Premium is the principle that judges should be paid, promoted, and developed at a premium to the operators they oversee. The logic is asymmetric risk. The marginal cost of a slow execution is recoverable. The marginal cost of a bad judgment in a high-automation environment can be catastrophic, and invisible until it lands.
The Randazzo findings make this concrete. Cyborgs and Centaurs are building capability. Self-Automators are not. If your comp structure pays all three the same, you are paying the same for the mode that compounds value and the mode that quietly erodes it. The Verification Premium fixes the incentive. It pays for the trajectory you want your workforce on.
Aviation, again, got here first. A captain earns materially more than a first officer, and not for stick-and-rudder skill. The premium is for judgment under failure. It is a compensation structure that aligns pay with the moment that matters most.
Seven moves, by seat at the table.
This is not a future-of-work question. It is a this-quarter org-design question. Every company deploying AI agents is creating two tiers, and we now have the measurement: three modes in 244 consultants under the same nominal workflow, with two building skill and one building none. The companies that design around what the research shows will pull away from the companies that inherit the pattern by accident, the same way the airlines that designed the human role around automation pulled away from the ones that drifted. This is the first turn of the flywheel.
The second article in this series argued that AI cognitive decline is no longer theoretical. This piece argues that its structural counterpart, the bifurcation of work into judges and operators with measurable skill trajectories, is the design conversation HR cannot afford to skip.
Next up: the five forms of inertia that keep companies from having this conversation at all.
For CHROs: the Two-Track Architecture is the workforce planning conversation you need this quarter.
For CHRO/Total Rewards or Compensation: the Verification Premium is the comp structure your competitors are already redesigning.
For CFOs: 71% of CIOs say they will cut AI budgets that cannot show value within two years. Funding the judge tier is how you show it.
For CTOs and CIOs: every agent deployment should be audited for whether the Judge Layer exists, keeps the producing and judging models separate, and works.
For CLOs and GCs: Self-Automator behavior in regulated workflows is the compliance failure mode hiding inside the bifurcation.
For CEOs: you are either designing the bifurcation or inheriting it. The research now gives you the diagnostic instrument.
The second article in this series argued that AI cognitive decline is no longer theoretical. This piece argues that its structural counterpart, the bifurcation of work into judges and operators with measurable skill trajectories, is the design conversation HR cannot afford to skip.
Next up: the five forms of inertia that keep companies from having this conversation at all.
The Workforce Paradox is a series on how AI is reshaping the architecture of work, leading to a white paper later this year. Subscribe for the full series. Follow on for daily soundbites on what is actually happening in HR and AI right now.
Notes
Randazzo, S., Lifshitz, H., Kellogg, K. C., Dell’Acqua, F., Mollick, E., Candelon, F., and Lakhani, K. R. (2025). Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise. HBS Working Paper 26–036.
Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775–779.
Federal Aviation Administration. (2016). Operational Use of Flight Path Management Systems: Final Report of the Performance-based Operations Aviation Rulemaking Committee Flight Deck Automation Working Group.
HBR Analytic Services. (2026). What Drives AI Value: Why Modernization and Workflow Integration Matter. Sponsored by Appian. Survey of 385 decision-makers, fielded March to April 2026.
Champniss, G. (2026, May 1). The Psychological Costs of Adopting AI. Harvard Business Review. Survey of more than 1,200 full-time employees across ten sectors in the US and UK, conducted for Meltwater Consulting.
De Neve, J. E., Hancock, J. T., and Niederhoffer, K. (2026, April 15). Why Companies That Choose AI Augmentation Over Automation May Win in the Long Run. Harvard Business Review. Survey of 1,294 full-time desk workers across the US, Canada, and the UK, January 2026.
CIO budget figure (71%, two-year value threshold) reported by Davenport, T. H., and Srinivasan, L. (2026, March 17). 7 Factors That Drive Returns on AI Investments, According to a New Survey. Harvard Business Review.
Nate Jones, the Judge Layer. https://open.substack.com/pub/natesnewsletter/p/agent-judge-layer-production-control. Used with attribution for the Judge Layer concept.
Originally published at https://theworkforceparadox.substack.com.
