Most conversations about AI and product work begin with capability and end with a feature list. A model can draft the spec, generate the prototype, summarise the interviews, write the test. All of that is true, and all of it is the least interesting part.
The useful question is narrower: which cost actually moved?
The cost that collapsed
For most of the last two decades, the expensive part of product work was production. Writing the document, building the clickable prototype, turning research notes into themes, producing the first version of anything. Those tasks took days, and the days shaped the process. We wrote lengthy specs because writing was the only way to align a team. We built high-fidelity prototypes late because they were expensive. We batched research synthesis because each round cost real effort.
That cost has fallen sharply. A first draft is now minutes. A prototype that looks real is an afternoon. Synthesis of twenty interview transcripts is a prompt, a review and an edit.
When production gets cheap, the bottleneck does not disappear. It moves.
Three shifts worth naming
From artifact to argument. When everyone can produce a polished document, the document stops being evidence of thinking. What becomes scarce is the argument inside it: why this problem, why this approach, what would prove it wrong. Teams that were previously differentiated by the quality of their artifacts will find that advantage evaporating. Teams differentiated by the quality of their reasoning will find it amplified.
From production to editing. Generation is abundant; selection is not. The skill that compounds is the ability to look at ten plausible options and explain why eight of them are wrong for this context. That is taste, but not in the vague aesthetic sense. It is the ability to hold a model of the user, the constraint and the trade-off in your head at the same time, and to notice when an option quietly violates one of them.
From certainty to verification. AI output is fluent. Fluency is easily mistaken for correctness, especially when it arrives in the shape of a chart, a citation or a confident paragraph. The new overhead is verification: checking the claim, the number, the source, the assumption. Teams that do not budget for verification will not get faster. They will get faster at being wrong.
A prototype used to be a claim about feasibility. Now it is a claim about taste. Neither is a claim about demand.
What did not change
The problem is still the problem. A model can generate a feature for a problem you have not understood, and it will do so beautifully. The output will be coherent, on-brand and irrelevant. Understanding whose problem you are solving, at what moment, and what it costs them today remains stubbornly human work, not because machines cannot process language, but because the evidence usually lives in behaviour, context and things people cannot articulate.
Trade-offs still need an owner. Every product decision gives something up. AI can enumerate the options and describe the consequences, but it cannot decide which consequence your team is willing to live with. That is a question of values, risk appetite and strategy, and it belongs to a person.
Trust is still earned in use. A generated interface can look finished. It cannot make the product reliable, coherent over time, or honest about what it does with someone’s data. Those properties come from craft and accountability, not from the speed of the first draft.
The prototype trap
The most seductive use of AI in product work is the instant prototype. It is genuinely useful, and it creates a specific failure: it answers the feasibility question so quickly that teams skip the desirability question.
A clickable prototype can make a weak idea feel inevitable. The room responds to the polish, not the premise. Before the prototype, the team was arguing about whether anyone needed this. After it, the team is arguing about button placement. The first argument was the important one.
The discipline is to separate generation from evaluation. Build the artifact, then put it down. Ask what problem it solves and what evidence supports that problem. Show it to people who have the problem, not people who admire the craft. Keep the prototype cheap enough that you are willing to throw it away.
A practical posture
Three habits have helped me get more from these tools without letting them flatten my judgement.
First, use them for counterarguments, not confirmation. Asking a model to argue against a decision is more valuable than asking it to justify one. It will produce objections you have not considered, and you can decide which ones matter.
Second, keep the decision log outside the tool. A short record of what was decided, why, and what would reverse it. When the context changes or a new model produces a compelling alternative, the log is what stops the team from relitigating from memory.
Third, separate the work of making from the work of choosing. Give generation its own time and space, then switch modes deliberately. The quality of the second activity determines the value of the first.
The uncomfortable part
AI raises the floor. The average document is better, the average prototype is more convincing, the average analysis is more complete. That is good for the industry and uncomfortable for individuals, because it makes the average harder to distinguish.
The differentiator is not access to generation. Everyone has that. The differentiator is the discipline of knowing what to throw away: the ability to look at a large volume of plausible work and say, with reasons, this one, not those.
That is a product skill. It was a product skill before any of this, and it is the part that still takes years to learn.