Insights
Why Generative AI Deepens Innovation Bottlenecks
Most organisations are buying speed at the stage that was never slow.
I have been sitting with a framework from Harvard Business School this summer — a working paper by researchers proposing what they call a human bottleneck perspective on innovation (HBS WP 26-094, 2026). Their premise is quietly radical. The constraints that hold back the innovation process are rarely technical. They are cognitive and social. They live in how people generate ideas, how they judge novelty, and how ideas travel through social systems. And generative AI does not act uniformly on those constraints. At each stage it relieves some and deepens others.
That last part is the sentence I keep returning to. We talk about AI as though it were a tide that lifts everything. It is closer to a lever. Where you place it determines whether something moves or breaks.
Consider the stages
Consider the stages. Idea generation was genuinely constrained — a small group of people, a finite number of hours, the limits of one’s own reading. AI relieves that constraint almost completely. Organisations can now produce more plausible ideas in an afternoon than they previously produced in a quarter.
But screening was also constrained, and AI does not relieve it in the same way. Judging novelty requires knowing what has been tried, what failed, what the organisation can actually absorb. That judgement rests on tacit knowledge held by a small number of people. So when ideation output multiplies and screening capacity stays fixed, the bottleneck does not disappear. It relocates — and it lands on the people who were already the constraint.
I suspect this is why so many AI programmes feel oddly disappointing from the inside. Output rises. Decision speed does not. Leaders read this as an adoption problem and buy more tools for the stage that was already fast.
A second literature sharpens the point
A second literature sharpens the point. Work published this year in Frontiers in Psychology distinguishes two ways people hand thinking to a machine. In dependent offloading, the person delegates the judgement itself and accepts the result. In autonomous offloading, the person delegates the production and keeps the judgement — using the output as material to interrogate. The behaviour looks identical from the outside. The same tool, the same prompt, the same document arriving on time. What differs is where the human mind stayed in the loop.
Organisations measure the artefact. So they cannot tell these two apart. And over time, if only the artefact is measured, the cheaper of the two habits wins.
Which bottleneck are we actually relieving?
What follows is not a technology strategy. It is a design question, and it can be asked in an hour:
Which bottleneck are we actually relieving? Name the stage. If it is generation, expect volume, not better decisions.
Which bottleneck did we just deepen? If generation multiplied, who now has to judge all of it — and did we give them anything?
Where is judgement still held by too few people? That is where contribution is scarce, and where the organisation is most fragile.
Are we rewarding autonomous or dependent offloading? Ask people not what they produced, but what they rejected, and why. Rejection reasoning is the most legible signal of thinking we have, and almost nobody collects it.
What our particular system was actually short of
There is something clarifying in the bottleneck framing. It refuses the fantasy that a tool improves an organisation evenly. It puts the question back where it belongs: not what the technology can do, but what our particular system was actually short of.
Most organisations, I think, were never short of ideas. They were short of the courage and the capacity to judge them. No model relieves that. It only makes the shortage more visible — and, perhaps, more expensive to keep ignoring.