Co-Existence: The Next Phase of AI

By Ethan Mollick
Portfolio / Penguin Random House, October 2026
240 pages

In an era saturated with AI hype on one side and existential dread on the other, Ethan Mollick offers something rarer and more useful: a clear-eyed, practical field guide to living and working with machines that are sometimes—but not always—smarter than we are. Co-Existence: The Next Phase of AI, the sequel to his 2024 bestseller Co-Intelligence, arrives at exactly the right moment. The chatbot era has given way to agentic systems capable of multi-step, financially valuable work, yet true artificial general intelligence remains (for now) over the horizon. Mollick, a Wharton professor, Academic Director of the Generative AI Labs, and author of the widely read “One Useful Thing” newsletter, focuses squarely on the strange in-between years we actually inhabit.

The book’s central organizing idea is the “jagged frontier.” AI capability is not a smooth, steadily rising curve. The same system that writes elegant code, produces sophisticated strategic analysis, or tutors a student with remarkable patience can invent facts, miss obvious context, or fail spectacularly on tasks that look trivial to humans. This unevenness, Mollick argues, is the defining feature of our current technological moment. Leaders and knowledge workers who treat AI as either an all-powerful oracle or a mere parlor trick will systematically underperform. Those who learn to map the frontier daily, to recognize where the machine is strong and where it is weak, will build a durable advantage.

Mollick moves well beyond the now-familiar territory of prompt engineering. He proposes three practical modes of collaboration that reflect different degrees of integration between human and machine. “Centaur work” involves relatively clean handoffs: the human defines the problem, sets constraints, and evaluates the final output while the AI executes the bulk of the intermediate work. “Cyborg work” is continuous interleaving—human and AI iterating in real time, each refining the other’s contributions. “Conjurer work” delegates greater autonomy, allowing the system to explore solution spaces with lighter human oversight. Each mode has distinct strengths and failure modes. Mollick illustrates them with concrete experiments drawn from his research and his own writing process, including cases in which the AI produced passages he wished he had written himself and others that were flatly uninspired. The tone throughout is warm, lightly humorous, and refreshingly free of both techno-utopianism and doom-mongering.

A recurring and important theme is the risk of cognitive offloading. Large language models trained with reinforcement learning from human feedback are optimized to be helpful, agreeable, and persuasive—sometimes to the point of sycophancy. The danger is not merely that they will hallucinate facts; it is that they will gradually erode our own capacity for judgment, originality, and critical detection of nonsense. Mollick is blunt: the lasting human advantage lies in remaining the final arbiter of quality, context, values, and taste. He offers concrete practices for preserving that edge—structured verification routines, deliberate introduction of friction into workflows, and intentional periods of unaided thinking. These recommendations feel especially urgent for educators and parents concerned about the development of young minds, and for managers responsible for the intellectual capital of their organizations.

The organizational chapter is among the book’s strongest. Mollick contends that simply “adding AI” on top of existing processes rarely produces transformative results. Individual productivity gains are real and measurable, yet organizations designed for a pre-AI world tend to absorb those gains without converting them into collective performance improvement. He sketches a practical leadership–lab–crowd model. Senior leadership must engage deeply with the technology rather than delegating AI strategy to a distant digital or innovation team. Dedicated “labs” must run continuous, rigorous experiments on how the technology can reshape core processes. And the broader workforce—the “crowd”—must be both empowered and expected to surface use cases and surface problems. In a world of increasingly agentic systems, the human role shifts toward goal-setting, oversight of complex multi-step workflows, integration of social and political context, and the ethical judgment that machines still lack. “Things are very weird,” Mollick observes. “They may get weirder.”

Co-Existence is deliberately not a technical treatise. Readers seeking detailed discussions of transformer architectures, scaling laws, or speculative timelines for superintelligence will find the book wanting. Its strength is its disciplined focus on the practical, present-tense questions that most managers, professionals, consultants, and educators actually face right now: How do I productively work alongside something that occasionally outperforms me in my own domain? How do I protect my own thinking from systems designed to finish my sentences? How should institutions—companies, schools, governments—adapt their structures and incentives? The prose is accessible without being simplistic, grounded in peer-reviewed research and real-world experiments yet free of academic denseness or jargon for its own sake.

For Harvard Business Review readers, the book’s greatest contribution is its refusal of false dichotomies. AI will neither replace knowledge workers wholesale in the near term nor remain a peripheral productivity tool. The competitive edge will belong to individuals and organizations that treat coexistence as a craft—mapping the jagged frontier every day, choosing the appropriate mode of collaboration for each task, and deliberately preserving human agency and judgment. At 240 pages, Co-Existence is concise enough for busy executives yet substantive enough to shape strategy discussions at the highest levels. It is the rare technology book that feels simultaneously urgent and measured, optimistic about human potential without being naïve about technological power. Highly recommended for anyone whose work involves thinking, deciding, or leading in the age of intelligent machines.

Mind Map of Co-Existence

 

 

 

 

 

 

 

 

 

 

 

 

Glossary

• Jagged frontier: The uneven, unpredictable boundary of AI capabilities—strong in some domains, weak or unreliable in others that appear similar to humans.
• Centaur work: Collaboration mode with clean handoffs between human and AI.
• Cyborg work: Continuous, interleaved collaboration between human and AI.
• Conjurer work: Mode in which humans delegate substantial autonomy to AI systems.
• Cognitive offloading: The tendency to transfer mental effort to AI, potentially at the cost of one’s own judgment and skill.
• Agentic AI: Systems capable of planning and executing multi-step tasks with limited human intervention.
• Leadership–lab–crowd model: Organizational framework requiring engaged senior leadership, dedicated experimentation units, and broad employee participation.