AI solved the wrong bottleneck

In the summer of 2018, a restaurant in Bangalore installed a kitchen automation system that could plate dishes in half the time of a human line cook. The owner was thrilled. Throughput doubled. Orders left the kitchen faster than ever. But the reviews got worse. Not immediately. Over a few weeks. Diners started describing the food as "fine" and "nothing special," which in restaurant terms is a death sentence delivered politely. The problem was not the plating speed. The chef had started skipping the tasting step. With the machine handling output so quickly, there was pressure to keep the pipeline fed, and the contemplative moment where the chef adjusted the seasoning got compressed, then eliminated. The kitchen was producing food faster. But nobody was thinking about whether the food was good.

I keep returning to that story because it is the most precise analogy I have found for what is happening in product teams right now.

Why: the constraint was never production

Here is the uncomfortable truth that the AI productivity conversation has been sidestepping for eighteen months. The quality constraint in product work has never been the speed of producing outputs. It has always been the quality of thinking that precedes those outputs.

A product brief that takes three hours to write is not slow because writing is slow. It is slow because thinking takes time. The PM has to sit with the problem, identify the real constraint, consider second-order effects, test assumptions against what they know about the customer, and arrive at a position they can defend. The writing is just the last step. But it is the step that AI has accelerated.

AI solved the wrong bottleneck. Production was never the constraint. Thinking was.

I tested this on myself earlier this year and the results were clarifying in a way I did not enjoy. I used an AI tool to generate a product strategy document for a side project. The output arrived in ninety seconds. It was well-structured, clearly written, and entirely reasonable. But when I sat with it for an hour, I realised I disagreed with the core framing, and I only discovered my disagreement because I forced myself to read the document slowly and interrogate each assumption. The AI had produced a plausible strategy. But it was not my strategy. It had not emerged from my thinking. It had replaced my thinking.

The document looked like a decision. But no decision had been made.

How: speed without thought produces volume without value

The pattern I have been watching across product teams in 2025 is what I have started calling the production illusion: the appearance of high output masking a decline in thinking quality. Teams are producing more documents, more briefs, more specs, more decks. The volume is up. But the thinking density per document is down.

At Schneider Electric, years before AI tools existed, I saw a version of this same pattern created by a different cause. The product team had implemented a new documentation template that made it faster to produce specs. The template had pre-filled sections, standard language, and dropdown menus for common choices. Output tripled. But the specs started converging toward a mean. Every document looked professional. None of them contained a genuinely original insight about the customer problem. The template had made it easier to produce specs without making it necessary to think through them. The tool had optimised the wrong step.

That is exactly what AI is doing now, at far greater scale.

I mentor junior PMs, and I have started noticing a specific tell. The ones using AI to generate first drafts produce documents that are smoother, better organised, and more polished than anything I wrote at their career stage. But when I ask them to explain the reasoning behind a particular decision in their brief, there is a pause. Not the pause of someone organising their thoughts. The pause of someone encountering a question about something they have not actually thought through. The document has an answer. But the person who produced it does not.

That is the production illusion at work.

What: separating thinking from production

The practice gaining traction among senior product people I speak with is deceptively simple. Separate the thinking phase from the production phase, and use AI only in the second.

The thinking phase is the messy part. It is the hour you spend staring at a whiteboard, the twenty minutes of writing and deleting the same paragraph, the conversation with a colleague where you realise your original framing was wrong. This phase produces no visible output. It looks unproductive. It is the phase where actual product decisions get made.

The production phase is everything that follows: structuring the document, drafting the prose, formatting the deck, generating the diagrams. This is where AI genuinely helps. It can turn rough thinking into polished output faster than any human. But it cannot do the rough thinking for you. Or rather, it can produce something that looks like rough thinking. The appearance of thought is not thought.

I have been calling this the thinking phase, which sounds almost redundant when you say it out loud. Of course there is a thinking phase. But the point of naming it is that it is the phase most at risk of being skipped, precisely because AI makes the next phase so fast that the gap between "I have not thought this through" and "I have a finished document" has nearly disappeared.

At Nike, I once watched a design lead present a concept that was visually stunning and strategically incoherent. Someone in the room asked how long she had spent on it. "About two days," she said. A colleague leaned over to me and whispered, "I think she spent two days on the pixels and about ten minutes on the problem." That dynamic existed long before AI. But AI has made it possible to compress those two days into two hours while still spending only ten minutes on the problem.

The ratio has got worse, not better.

The honest accounting

I am not arguing against AI tools. I use them daily. They have made specific parts of my work genuinely faster, and I would not go back. But I have also learned where they help and where they deceive, and the distinction is not complicated.

AI helps when you have already done the thinking and need to produce the output. It deceives when you have not done the thinking and it produces the output anyway. The output looks the same in both cases. The product decisions embedded in one have been earned, and the product decisions embedded in the other have been fabricated by a language model that has no stake in whether they are right.

The teams I see struggling most right now are not the ones avoiding AI. They are the ones using it before they have earned their own clarity. They prompt an AI tool with a vague sense of the problem and receive a crisp-sounding document, and the crispness of the output persuades them that the thinking has been done. But it has not. The thinking was skipped. The production just happened to look good.

The product teams doing their best work in 2025 are the ones who have learned to sit with the discomfort of the thinking phase. They stare at blank pages. They write bad first paragraphs. They argue with colleagues about framing. They do the slow, unglamorous work of figuring out what they actually believe before they ask a machine to articulate it beautifully.

Speed is a gift when you know where you are going. When you do not, it just gets you to the wrong place faster.