Co-Intelligence is Mollick’s practical case for treating AI as a collaborator — not a tool, not a threat, but a genuine co-worker. Written from the vantage point of an educator watching his students navigate an entirely new cognitive landscape, it is one of the most grounded and accessible takes on generative AI in print.
The book is most useful for someone new to working alongside an AI. For practitioners already embedded in the tooling, much of it will feel familiar. That is not a criticism — it is a signal of how far the field has moved.
The interesting territory is what the book gestures at without fully theorizing. Mollick reports findings; the structural read of those findings runs further than the book takes them.
The central frame — the jagged frontier
Mollick’s most useful concept is the jagged frontier: the uneven edge of AI capability. AI does not fail the way humans fail. It succeeds brilliantly at tasks we assumed were hard, and stumbles at tasks we assumed were easy.
AI doesn’t fail the way humans fail. It succeeds brilliantly at tasks we assumed were hard, and stumbles at tasks we assumed were easy.
The implication is practical: stop trying to predict what AI will or will not do. Map it by doing. The frontier is real but it is also moving, and the only honest way to know where it is at any moment is to keep working at it.
The BCG study — the empirical anchor
The book’s load-bearing claim sits on Mollick’s own research with Boston Consulting Group. Seven hundred and fifty-eight consultants were sorted into three groups: humans alone, humans with AI, AI alone.
The numbers most often quoted:
- The AI-assisted group completed 12.2% more tasks, 25.1% faster, at 40% higher quality.
- The performance gap between top and bottom performers — previously 22% — collapsed to 4%.
Where previous technical revolutions often targeted more mechanical and repetitive work, AI works, in many ways, as a co-intelligence… [it] can lead to 20–80% improvement in productivity across a wide variety of tasks.
The finding that warrants more attention is the gap collapse. The previous 22% spread between the top and bottom of the distribution was not a productivity spread alone — it was an access spread. Senior consultants carried accumulated context that juniors did not yet have. AI gave the juniors immediate access to that context. The gap did not close because the bottom got faster. It closed because the bottom got access to the same default expertise the top had.
That is a real economic change. It also raises a structural question Mollick does not fully answer: where does the bottleneck go?
What expertise actually is
The book’s closing imperative — have high levels of expertise as a human — is correct, and it is the part that hides the most. Expertise is a word that flattens what is actually two different things: accumulated knowledge, and accumulated judgment.
Knowledge is what just became cheap. Judgment is what didn’t.
The structural reading borrows from Mises. The argument Mises makes in Human Action is that economics is not the study of inner mental states; it is the study of human action — the manifest choices people make given the context available to them. The aggregate of choices, across people and across time, is the substance of an economic system.
If knowledge is now democratized — every junior has the accumulated frontier within reach — then the differentiating capital is no longer information. It is the capacity to choose well. The accumulated weight of someone’s decisions across their domain is what we usually call taste.
Taste is praxeological capital. It accumulates the way Mises says all economic understanding accumulates: through a person directly grappling with material, making choices, observing consequences, adjusting. It does not transfer through reading or listening. AI closes the knowledge gap. It does not close the taste gap.
AI closes the knowledge gap. It does not close the taste gap.
Mollick is aware of part of this. He writes that AI “gravitates toward ‘average’ outputs” and warns against outsourcing taste. He frames expert humans as carriers of “deep subject matter knowledge, broad system leadership, or excellent taste.” The taste concept is present in the book. The structural claim — that taste is the central form of capital in an AI-augmented economy — is not. That is the next move.
The calculator framing
Mollick’s dominant analogy is the calculator. Not electricity. Not the steam engine. The calculator.
It is a deliberate choice and it is the choice an educator naturally makes. Calculators changed how mathematics was taught and applied. The debates about dependency took decades to settle, and arguably never did. The adoption curve was uneven, infrastructural at the human scale, never quite finished. Mollick is asking: what happens when the calculator can write, analyze, argue, and reason?
For comparison: Agrawal, Gans, and Goldfarb’s Power and Prediction uses electricity as the frame — an infrastructure transformation that reorganizes entire industries.
Both framings are useful. They look at different granularities. The calculator frame is about individual augmentation; the electricity frame is about system-level reorganization. The calculator analogy is closer to how the book reads in practice, because Mollick is writing about how people personally engage with the tool. The electricity frame is closer to where the structural argument has to land, because the changes the BCG study points toward are not individual — they redistribute who gets to make which decisions inside a firm.
The misread signal — layoffs as liquidation
The companies currently laying people off in the name of AI efficiency are, on the structural reading, doing the wrong thing. They are liquidating the form of capital that just became most valuable: the human decision-makers whose accumulated taste differentiates one company from another.
The old read of AI-driven layoffs is efficiency gain. The structural read is the opposite. A company laying off knowledge workers in 2026 is signaling, in public, that it has not figured out how to deploy taste-makers in a world where knowledge is no longer scarce. It is making the wrong trade — selling off choosers because it does not know how to use them.
The interesting live counterexample is Dan Shipper at Every. Every has roughly tripled its team in the past year. The team is organized horizontally — individual team members own specific products. Each member is a taste-maker reinvesting cognitive capital into discovery and creative work rather than information processing. The pattern is the opposite of layoff-for-efficiency. It is hire-for-judgment.
The question worth sitting with is not who is right, Shipper or the layoff companies? It is: what structure of a company makes layoffs the rational move, and what structure makes more hiring the rational move? That is a topology question. It is the question this framework exists to ask.
The gentler critique
Mollick is working within an institutional frame that does not fully break from hierarchy. The book still implicitly assumes organizations are trying to optimize within existing structures — how to use AI better inside the current shape, not whether the current shape is the problem.
That is not a flaw. It is an honest position from an educator working with real institutions. But it is the boundary the book stops at, and it is the boundary where the publication’s own argument starts. If taste is praxeological capital, and praxeology is a theory of distributed decision-making, then the hierarchical firm is structurally the wrong vessel for the AI-augmented economy. That claim is too large for Co-Intelligence. It is the next book.
The verdict
For a general reader: essential. The book describes what is happening with AI in working life without overreaching and without prediction. That kind of writing is rare and useful.
For someone already deeply embedded in AI tooling: the BCG study is worth the price of the book. The conceptual frame is directionally right. The implications run deeper than Mollick traces — but the book points clearly enough that the next move is visible.
Assume this is the worst AI you will ever use.
Mollick’s final imperative is the most consequential line in the book. If this is the worst AI we will ever use, then the question of how to organize work around it — including whether to keep hierarchy at all — cannot be deferred.