When Ideas Become Cheap

If AI can give you fifty plausible directions before lunch, getting ideas is no longer the hard part. The hard part is deciding which two deserve tomorrow.

Producing alternatives used to carry enough cost that many possibilities never became concrete. Generative AI changes that. Ideas that might once have remained hypothetical can now become drafts, designs, prototypes, or competing approaches quickly enough to actually compare.

Research has already demonstrated gains in some kinds of work. In a 2023 experiment published in Science, professionals given access to an AI writing tool completed assigned tasks in less time and produced work that received higher quality ratings.¹

That is a real improvement. But once one part of the work gets faster, the rest of the process has to keep up with it.

Making one part of a process dramatically cheaper does not make the whole process cheap. It changes where the constraint shows up. If you can generate 100 ideas in the time it used to take to produce 10, and each needs similar attention, you've effectively increased your evaluation load tenfold. Unless your ability to evaluate those ideas grows with it, the bottleneck hasn't disappeared. It has moved.

This can create the illusion of progress. You are producing more, seeing more, and moving faster, but the number of good decisions you can make may not have changed at all.

Finding the next bottleneck

The useful question changes with the bottleneck. Instead of asking how to produce more, you have to ask what is now preventing the work from getting better.

Imagine a small cleaning company preparing to advertise a new service. AI helps the team produce several clear versions of the message. One emphasizes a lower price; another emphasizes saving the customer time. Both describe real features of the service. The team cannot decide which benefit should lead.

It could ask for another round of wording. But it already has messages it could use. The unresolved issue is not whether the sentences sound good. It is whether the customers it wants to reach care more about saving money or saving time.

The next useful task is to investigate that difference. The team could review what customers have actually asked about, then compare the two messages in a test with similar prospective buyers. Rather than asking colleagues which version they like, it would look at which message leads people to inquire about the service.

The result might still be inconclusive. But the team would be gathering evidence about the decision it needs to make, rather than hoping another rewrite will settle it. AI could help organize the responses. It cannot turn an invented customer reaction into an observed one.

The team also does not need to review every version equally. A message that promises something the business cannot deliver can be rejected without a lengthy debate. Slightly different wording need not become a separate strategic decision. Attention can go to the differences that might actually change the choice.

For your own project, start with what you are trying to finish. What would let you take the next step with a reasonable basis for doing so? In this example, the team has enough usable language but not enough knowledge of its customers. Naming that missing piece changes what it should spend its time on.

When the bottleneck moves again

Suppose the same business learns which message works and begins attracting more inquiries. If it then struggles to schedule enough cleaners, better advertising may no longer be its most useful improvement. The priority has shifted from attracting customers to having enough staff to serve them.

The marketing improvement worked. But the larger goal was not a busier inbox. It was more work the company could deliver well. A business that keeps measuring success only by inquiries can miss the fact that it now needs a different kind of improvement.

AI need not be limited to giving you more options. A study published in the Quarterly Journal of Economics in 2025 found that AI assistance increased the number of customer issues support agents resolved per hour. That is a gain in completed work, not merely an increase in generated text.²

Nor is there a reason to assume that every part will improve together. The task is to notice what has changed in your particular work, rather than decide in advance that human judgment will always be scarce or that another tool will always solve the problem.

This is where the moving-bottleneck idea becomes useful beyond the first burst of productivity. The goal can stay the same while the most valuable improvement changes. You may need to stop optimizing the part you have just become exceptionally good at.

The next time AI makes part of your work dramatically easier, ask yourself:

What becomes the hard part now?