Over 200 consultants at Boston Consulting Group got the same strategic task. Same AI tool. Same time pressure. Same training.
They produced three completely different kinds of outcome — and it had nothing to do with the tool.
Some consultants got the highest accuracy of anyone in the study. Some grew new skills. Some turned in fast, polished work that quietly hollowed out their own expertise. The difference wasn't what AI could do. It was how each person decided to work with it.
This is the layer no prompt guide covers. Not what the model knows. Not whether the system survives production. What you become while using it.
The One-Liner
Centaur delegates tasks. Cyborg shares the workflow. Self-Automator delegates the whole job.
Core question: who decides what gets done, and who decides how it gets done?
Hold that frame. It separates all three categories cleanly.
The Study
"Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise" — Harvard Business School Working Paper 26-036, December 2025.
Authors: Steven Randazzo, Hila Lifshitz-Assaf, Katherine C. Kellogg, Fabrizio Dell'Acqua, Ethan Mollick, Francois Candelon, and Karim R. Lakhani — a team from Harvard, Wharton, MIT Sloan, and Warwick.
They didn't survey people about their AI habits. They studied the conversational fabric itself — nearly 5,000 interactions across the full back-and-forth transcripts of 244 BCG consultants doing real knowledge work with GenAI. Who prompted what. Who checked what. Who accepted output without reading it.
That transcript-level view is what makes the finding sharp. People's self-image doesn't match their transcripts. The categories come from behavior, not identity.
The Two Axes
Every mode in the study sits on two questions:
- Who selects what needs to be done? — who frames the problem, picks the steps, decides what matters.
- Who identifies how it gets done? — who executes each step, produces the draft, runs the analysis.
Keep control of both and you're a Centaur. Share the second and you're a Cyborg. Give up both and you're a Self-Automator.
| Mode | Paper label | Decides WHAT | Decides HOW | Share |
|---|---|---|---|---|
| Centaur | Directed Co-Creation | Human | Human — AI assists | 14% |
| Cyborg | Fused Co-Creation | Human | Shared — AI drives execution | 60% |
| Self-Automator | Abdicated Co-Creation | AI decides both | AI | 27% |
Now each one in detail.
Centaur: Strategic Selector
The Centaur stays firmly in the driver's seat. Humans decide both what needs doing and how to do it. AI gets discrete, well-scoped subtasks — never the wheel.
Three practices dominated Centaur conversations in the study:
- Mapping the problem domain — asking AI for general background information
- Gathering methods information — requesting a specific technique, formula, or procedure
- Refining human-generated content — polishing writing the human produced
One consultant asked GenAI how to calculate a market's growth rate over a multi-year span, got the formula back, and then ran the calculation in Excel personally.
That last step is the whole behavior in miniature. The Centaur borrows knowledge, not labor.
The payoff: Centaurs achieved the highest accuracy of all three groups on business recommendations. They also strengthened their existing domain expertise — treating AI as an intelligent search engine and writing coach keeps your own skills in the loop.
The cost: they developed no new AI-related expertise, and stayed cautious about changing established workflows. Some expressed ethical unease about taking credit for AI-assisted work. The Centaur masters the old game better but rarely learns a new one.
Cyborg: Fused Collaborator
The most common mode — 60% of participants. Here the human still decides what they're trying to accomplish, but GenAI drives how the work gets executed: drafting, analyzing, generating options, reworking outputs.
Not "prompt once, accept once." Cyborg collaboration is an iterative dialogue — the researchers describe it as unfolding through probing AI outputs, extending ideas, and validating results in a seamless rhythm of joint problem-solving.
Cyborg interaction practices:
- Modularizing — breaking the task into steps and looping AI through each
- Validating — checking outputs against their own judgment
- Re-running with new data — feeding additional context back in to improve results
- Pushing back — challenging outputs that conflict with their own view
The payoff: Cyborgs developed entirely new AI-related expertise — the researchers call it "newskilling" — while keeping their domain knowledge. New skill stack, old skill stack, both intact.
The cost: a persistent trap. GenAI confidently reinforces the wrong path, and fluency in the dialogue can feel like correctness. Cyborg mastery has to include knowing when not to trust the conversation.
Self-Automator: Dangerous Delegate
27% of participants. The Self-Automator consolidates the entire problem-solving workflow into one or two interactions: dump all the data into a single prompt, accept the output, move on.
The researchers describe this work as fast and polished but lacking depth — output that resembles something completed for the consultant rather than with them.
The number that should make you wince: 44% of Self-Automators accepted AI output with zero modification. Most of the rest made only superficial edits.
The payoff is real. Fast. Polished. For routine tasks, tight deadlines, or problems sitting well inside GenAI's capability, full delegation is efficient. The productivity gains aren't imaginary.
The cost lands later. When work requires judgment — framing the real problem, deciding what matters, evaluating competing narratives — abdication hollows out the exact capabilities that make a professional valuable. Self-Automators developed neither new domain expertise nor new AI-related skills. The output looks done. The skill growth is zero.
Why Abdication Feels Like Productivity
The Self-Automator's work passes review more often than it should. It's fluent, formatted, and finished-looking. Nothing in the output screams "no one was home."
That's the trap. Skill atrophy is invisible in the artifact. It shows up months later — in the meeting where you can't defend the analysis, in the judgment call you haven't built muscle for, in the interview where your expertise turns out to be one prompt deep.
The underlying pattern: when you give up control over what you do, you rarely keep control over how you do it. Decide the "what" and you keep leverage over the "how." Abdicate the "what" and the "how" tends to follow it out the door.
Match the Mode to the Task
The paper's practical takeaway isn't "be a Centaur." Each mode has a correct use. The failure is using one mode for everything.
When error cost is high — financial forecasting, operational diagnostics, anything where a wrong answer compounds → run Centaur. Execute the core analysis yourself. Use AI for targeted support. Maximum accuracy, and your expertise keeps deepening.
When the task rewards creative extension — ideation, rhetoric, stakeholder persuasion, drafting with flair → run Cyborg. The iterative loop extends ideas further than either side alone, and you pick up AI fluency while doing it.
When the task is routine — bounded, low-stakes, inside GenAI's competence, under deadline → Self-Automator is fine. Full delegation is efficiency here, not surrender.
Abdication is only a trap when judgment is the job.
The Human Layer
If you've read the earlier posts in this series, you know the stack:
┌───────────────────────────────────────────────┐
│ Layer 0: OPERATOR LEVEL — Work Mode │
│ How you divide judgment between you and AI │
│ Centaur / Cyborg / Self-Automator │
│ │
│ ┌─────────────────────────────────────────┐ │
│ │ Layer 3: Harness Engineering │ │
│ │ Whether the system survives production │ │
│ └─────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────┐ │
│ │ Layer 2: Context Engineering │ │
│ │ What the model knows at each step │ │
│ └─────────────────────────────────────────┘ │
│ ┌─────────────────────────────────────────┐ │
│ │ Layer 1: Prompt Engineering │ │
│ │ How you ask │ │
│ └─────────────────────────────────────────┘ │
└───────────────────────────────────────────────┘
Layers 1–3 decide what the system does. Layer 0 decides what you do — and what you're left with afterward.
Same model, same context pipeline, same harness. A Cyborg and a Self-Automator sitting in front of it produce different work and, more importantly, become different professionals over time. The tooling stack is shared. The skill outcome is chosen.
This also reframes the "AI will replace you" panic. In the study, nobody's job got replaced. What diverged was expertise. The consultants who kept judgment in the loop compounded. The ones who abdicated got a productivity bump and a flat skill curve.
Which One Are You?
Your transcript would tell the truth. Short of that, honest signals:
You're running Centaur if you ask AI for the formula and compute the answer yourself. Your prompts request information, methods, and polish — never the conclusion.
You're running Cyborg if your sessions are long dialogues. You push back on bad output, re-run with better inputs, and can't always tell where your idea ends and the model's begins.
You're running Self-Automator if your prompts open with "here's all the data" and close with a request for the final deliverable. If you accept output more often than you edit it. If a task finished in one round trip.
Most people run all three, in the same day. That's fine. The question is whether you're choosing the mode — or drifting into whichever one is fastest.
The Bottom Line
Prompt engineering is what you ask. Context engineering is what the model knows. Harness engineering is whether the system survives production. Work mode is whether you survive it.
The study's most hopeful finding is the last one: the most valuable professional isn't a fixed type. It's an adapter — someone who deliberately switches modes per task. Centaur when the stakes demand accuracy. Cyborg when the work rewards co-creation. Self-Automator when the task is genuinely routine.
The tools will keep improving. Your judgment is the only component that can't be automated away — because the moment you automate it yourself, it's already gone.
References
- Randazzo, S., Lifshitz-Assaf, H., Kellogg, K. C., Dell'Acqua, F., Mollick, E., Candelon, F., Lakhani, K. R. — "Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise," Harvard Business School Working Paper No. 26-036, December 8, 2025. SSRN 4921696
- HBS AI Institute — The Three Ways Professionals Work with AI – Which One Are You?
