Vik Bajaj, the co-CEO of Jeff Bezos’s new venture Prometheus, said something recently in an interview with the New York Times, recently that deserves more attention than it got. Explaining why building something as complex as a jet engine is hard, he said: “You can’t build something like a jet engine with words alone, not even the words of mathematical equations.”
That is a striking admission to come from inside a company raising tens of billions of dollars to apply AI to engineering. It is, in effect, an admission of AI limits.
Reactions on Prometheus focused on the size of the bet: $12 billion in initial funding, a $29 billion valuation, talk of another $100 billion fund in the works. All this is undoubtedly important, but the more interesting story in pointing out the current limits of AI. There are lessons here for most, if not all CEOs, lessons that go beyond the term “artificial general engineer” itself.
But the real bottleneck is not invention. It’s coordination.
A clue appears in Bajaj’s comments. Designing something as complex as a jet engine, he said, “takes a thousand human minds creatively working together. It is one of the most complex things we do as a species.” Today that process takes the better part of a decade.
Ten years is a long time. Time is spent searching design spaces, running simulations, working out which materials and geometries may solve dizzying arrays of design constraints. But also, on hundreds and hundreds of engineers checking each other’s assumptions, ensure the compatibility of subsystems designed by different teams, and catch conflicting assumptions and even errors that may only surface when one group’s solution collides with another’s; thousands of specialists working from different assumptions, models, and constraints that ultimately produce a coherent artifact. Looked at this way, jet engines are organizational achievements.
AI is exceptionally good at reducing knowledge friction associated with search, simulation, and analysis. AI can also coordinate the work of multiple teams, reducing coordination friction – delays, communication overhead, redundant workflow. AI can reduce several types of friction. The key question is which friction it reduces. Coordination friction is often worth eliminating. Cognitive friction — disagreement, challenge, competing interpretations — the slow, inefficient process by which a large group of specialists converges on a single coherent design without losing the disagreements that catch mistakes early. This is not a bug in how engineering organizations work. Remove these inefficiencies and the process may collapse. It is, in a different form, the same inefficiency that makes science self-correcting: the back-and-forth, the dissent, the redundant checking that looks like waste until it catches the thing that would have failed during actual deployment. Strip it out too aggressively in pursuit of speed, and you get designs that move faster through the pipeline and fail in ways nobody caught in time. Not all optimization is good.
Prometheus is not trying to automate an engineer. It is trying to compress the coordination burden of team-of-teams.
What this means for other C-suites
Few companies are building jet engines. But almost every company has some version of it: decades-long, friction-filled processes, involving a large number of people, in which good ideas become real products. The Bezos-Bajaj framing offers a more useful question than the usual one that CEOs ask about AI, some version of: How fast is this coming, and how worried should I be? This is a question that invites either panic or denial, and both are common right now. A more useful question is: in my own decade-long (or month-long, or quarter-long) cycle from idea to shipped product, which parts of the delay are knowledge/cognitive — search, calculation, drafting, simulation — and which parts are coordination — alignment, sign-off, the slow social process of getting a large group to agree on one design?
That distinction changes what a CEO should expect to buy, and from whom. Tools that compress cognitive and knowledge friction are arriving fast, are increasingly commoditized, and are worth adopting aggressively — the case for hesitation here is weak. Tools that claim to compress coordination friction deserve more skepticism, not because coordination is sacred, but because so far there is no good evidence that an AI system can replace the disagreement and cross-checking that makes teams-of-teams’ outputs reliable. Bajaj’s words point exactly to this: the parts of a large engineering project that are hardest are the parts hardest to codify, to specify in language. A thousand human minds working together is a coordination problem before it is a computational one.
A related trap worth naming; the temptation, once cognitive friction gets cheaper, is to assume coordination friction will shrink too — that faster individual output means faster teams. Perversely, it mean the opposite: when each person can generate ten times as many ideas, drafts, or designs, the volume of things that need to be reconciled grows just as fast, and the coordination bottleneck gets worse.
None of this requires a view on whether “artificial general engineer” is a fair description of what Prometheus is building, or on the wisdom of valuing a 150-person company at $29 billion. What it does is that CEOs ask a sharper question of any AI investment in front of them: which kind of friction is this actually reducing? Bezos’s co-CEO, in a candid moment, has already told us where the hard part is. It is not in the words. It is in getting a thousand minds to agree on what the words should describe.
The most important question for CEOs is not whether AI makes organizations faster. It is whether AI is reducing coordination friction, cognitive friction, or both. The first often increases productivity. The second may be the source of reliability, creativity, and breakthrough innovation. Acceleration and progress are not the same thing.
This article originally appeared in RealClear Markets on June 30, 2026.
