Why You Must Read 'Co-Intelligence' for Working With AI

"Co-Intelligence: Living and Working with AI," by Wharton professor Ethan Mollick, opens from an unusual vantage point for a business book: Mollick studies AI adoption inside real companies and runs field experiments on it, rather than writing about the technology from the outside looking in. The book's central claim is that generative AI functions as what economists call a general-purpose technology, one that eventually touches nearly every task inside a business rather than a single department or function, and that most owners are underestimating how much of their operation it will reach, the same way early spreadsheet software reshaped accounting work well beyond what anyone expected when it first shipped. Mollick's most useful contribution isn't a prediction about the future, though. It's a name for something happening right now: what he calls the "jagged frontier." Some tasks fall easily within AI's reach, and others, despite looking just as difficult on the surface, sit stubbornly outside it, with no obvious way to tell which is which without testing. A field study Mollick co-authored with researchers at Harvard Business School, MIT, and Boston Consulting Group put a real number on that unevenness. Nearly 800 consultants worked on a set of realistic assignments, some with AI access and some without, and the consultants who used AI on tasks inside the frontier and started as the lowest performers on the team improved their output by 43%, closing most of the gap with the strongest performers on the team, a gap that usually takes years of coaching to close by any other method.

We have invented technologies, from axes to helicopters, that boost our physical capabilities; and others, like spreadsheets, that automate complex tasks; but we have never built a generally applicable technology that can boost our intelligence.

— Ethan Mollick, "Co-Intelligence"

The same study found the reverse problem just as clearly. On a task selected specifically because it fell outside AI's current strengths, consultants who used AI were 19 percentage points less likely to land on the correct solution than consultants working without it at all. That's the part of the "jagged frontier" idea that doesn't fit neatly into most companies' AI rollout plans, which tend to assume that if a tool handles one kind of writing or analysis well, it will handle the next similar-looking task about as well too. Mollick's argument is that the frontier can't be predicted from first principles; it has to be mapped through actual use, task by task, inside a specific business, because a tool that drafts a strong marketing email might produce a confidently wrong answer on a pricing calculation that looks, to a non-expert, like the same category of work. The book spends considerable time on why that happens: large language models don't check their own answers against reality, and a model with nothing solid to work from will still generate something that reads as fluent and certain, wrong or not, which is a harder failure to catch than an obviously broken tool would be. Mollick describes two working styles for people who use AI well despite that risk. Some act as what he calls "Centaurs," splitting work cleanly between themselves and the tool along a fixed line; others work as "Cyborgs," staying inside the tool continuously and correcting its output as they go, sentence by sentence rather than handing off a whole task at once. Neither style is inherently better. What both share is that the person doing the work has already tested where the frontier sits for their own tasks, instead of assuming competence in one area transfers automatically to the next.

The mapping exercise Mollick describes lines up with a distinction GH Group has already drawn between where AI actually saves owners time and the tasks small business owners shouldn't hand to AI yet. Both pieces land on the same conclusion Mollick's research does: competence on one task says almost nothing about competence on the next one until someone checks it. Inside the Inside-Out Method™, that checking is part of what belongs to the Structure phase, the point where an owner decides which tasks get delegated, to a person or a tool, and which stay under direct oversight because the cost of a confidently wrong answer is too high to risk on a first try. "Co-Intelligence" isn't a manual for any specific tool, and Mollick is upfront that the underlying models keep changing faster than any how-to guide could track; a reader looking for step-by-step prompts for one particular AI product will find the book more conceptual than that. It works better as a mental model for figuring out where a given tool's frontier actually sits before handing it something that matters, and for staying honest about which parts of that frontier still need a person checking the work before it goes out the door. GH Group doesn't have a confirmed read of this specific title from Gary the way it does for a few other titles on this list; it's included here because the jagged-frontier idea matches a pattern GH Group already sees in owner-led businesses experimenting with AI on their own, not as a personal endorsement beyond that.