Let me start with something that gets forgotten in almost every conversation about corporate strategy, including conversations about AI.

Companies are not higher beings. They are not oracles with privileged access to the future, or rational actors operating on superior information, or entities with the kind of institutional wisdom that transcends the limitations of the individuals who compose them. They are collections of flawed, anxious, competitive people, people very much like you and me, trying to make decisions under uncertainty while simultaneously managing their own self-doubt, their own organisational politics, and their own very human fear of being left behind.

This matters for understanding what happened with AI-first companies, because the decisions those companies made look very different when you view them through the lens of human psychology rather than corporate strategy. They were not the result of sophisticated analysis that simply turned out to be wrong. They were the result of fear, competitive pressure, and the very understandable desire to be on the right side of what looked, in 2022 and 2023, like it might be the most significant technological shift in a generation.

Understanding that clearly is the first step toward navigating what comes next.

Let me invite you to a free webinar where Aakash Gupta and I will show you hands-on how to become a truly modern AI PM. If you use the right AI, you can achieve the best product results in record time.

We'll show you how to operate like the best of the best on Tuesday, August 11th at 9 am PT.

Sign up for free here:

Now, back to the post:

When ChatGPT arrived, the comparison to the iPhone was not unreasonable. The iPhone genuinely changed how humans relate to technology, created entirely new categories of products and businesses, and made things possible that had previously been inconceivable to most people. The question of whether generative AI would do the same was a legitimate one, and the people who concluded that it would were not being naive.

What happened next, however, reveals how the comparison was misapplied.

The iPhone changed everything, but it changed everything slowly and through the products built on top of it, not through the iPhone itself being marketed as the product. Nobody bought an iPhone because it used a particular chip architecture or because the operating system was written in a specific language. They bought it because of what it allowed them to do: navigate, communicate, capture moments, access information, be entertained. The technology was the infrastructure. The experience was the product.

Generative AI got inverted. The technology became the product. Companies did not ask "what problem does this solve for our users and how can AI help us solve it better?" They asked "how do we make this an AI product?" and then worked backwards from there. The API became the marketing label. The model became the feature. Being AI-first became a positioning statement rather than a description of how the technology was actually serving users.

That was, in retrospect, a straightforwardly bad strategic bet, and the companies that made it are now living with the consequences.

There is something specific that needs to be said about GPT-3.5, which was the model that triggered the broadest wave of AI-first enthusiasm, because understanding what it actually was versus what it appeared to be explains a great deal of what went wrong subsequently.

GPT-3.5 was an extraordinarily impressive piece of technology. Its ability to generate fluent, contextually coherent text across an enormous range of topics and formats was genuinely unprecedented, and it created, for many people encountering it for the first time, a powerful impression of interacting with something that understood them.

It did not understand them. It was producing statistically likely continuations of text patterns learned from an enormous corpus of human writing. The impression of comprehension was an artefact of how good it was at that task, not evidence of anything resembling genuine understanding. The gap between the impression and the reality was, and remains, significant.

Companies that built strategies on the assumption that the impression was the reality discovered this gap the hard way. Chatbots that seemed to handle customer queries fluently in testing hallucinated confidently in production. AI-generated content that read well on first pass turned out to be unreliable in ways that were difficult to predict or control. Features that were demoed impressively in controlled conditions behaved unpredictably at scale with real users asking real questions.

The technology was genuinely impressive. It was not as generally capable as it appeared, and the gap between what it seemed to be able to do and what it could reliably be deployed to do in production environments was much larger than the hype cycle suggested.

It would be tempting, and some people have been tempted, to categorise the AI boom alongside 3D television, VR, cryptocurrency, and NFTs: technologies that generated enormous excitement, attracted enormous investment, and ultimately delivered much less than they promised.

That comparison is not quite right, and it is worth being precise about why.

3D television was largely a solution looking for a problem: a marginal enhancement to an existing experience that most users found more trouble than it was worth. Cryptocurrency addressed real problems in specific contexts but was overwhelmed by speculative excess that had nothing to do with the underlying technology. NFTs were almost entirely speculative from the beginning, with genuine use cases that were vastly outnumbered by the noise. VR is a game console that tried to fool us into believing it’s an alternative to the analogue world.

Transformer-based AI is different in kind. It has already changed meaningful things: the way software is written, the way certain categories of content are produced, the efficiency of specific knowledge work tasks, the accessibility of capabilities that previously required specialist expertise. These changes are real, documented, and in many cases significant. The companies and products that have integrated AI in ways that solve genuine user problems have genuinely improved. That is not nothing. It is actually quite a lot.

The failure was not in believing that the technology mattered. It was in believing that it mattered more immediately, more universally, and more transformatively than it did, and in structuring business strategy around that overestimate rather than around the more modest and accurate picture.

The era of pushing AI into products to see what happens, of shipping features because they have the AI label rather than because they solve a problem, is drawing to a close. The next wave of AI models from the major labs has, so far, produced incremental improvements rather than the step-change capability that would justify a new round of maximalist enthusiasm.

What that means is that the window for clear-eyed, analytically grounded product thinking is wider than it has been for the last two years.

The practical implication for PMs is simpler than the analysis above might suggest.

Remain a Product Manager.

That sounds almost dismissive, but it is genuinely the most useful thing I can say. The core discipline of product management, understanding user problems deeply, prioritising ruthlessly based on value and effort and risk, building things that create real and measurable improvement in people's lives or work, and being honest about what the evidence says regardless of what the narrative says, does not change because a new technology arrives. It applies to AI features exactly as it applies to everything else.

The PM who evaluated AI features using the same rigorous criteria they applied to everything else in the last two years made better decisions than the PM who suspended those criteria because the technology felt different. The companies that asked "does this AI feature solve a real problem for our users better than the alternative would?" before building it are in better shape than the companies that asked "how do we make this AI?"

This is the moment where the most analytically rigorous, least hype-susceptible product managers can genuinely differentiate themselves. Not by being sceptical of AI as a technology, which would be its own form of error, but by applying the same standards to AI features that good PMs have always applied to every feature: does it create real value? Is the evidence for that value solid? Is the risk of it failing or behaving unpredictably in production understood and accounted for? Are we building this because users need it, or because the market expects us to have it?

The answers to those questions, applied honestly, will produce better products than the AI-first positioning did. And in an environment where many companies are now quietly retiring features they shipped too hastily, the PM who avoided that trap has earned something valuable: credibility with their team and their leadership that will last longer than any product cycle.

Market corrections, whether in financial markets or in technology adoption, tend to feel like failures in the moment and look like course corrections in retrospect.

The AI-first enthusiasm was not a mistake in the sense of being irrational given the information available at the time. The technology was genuinely impressive. The potential was genuinely significant. The competitive pressure to move quickly was real. Most of the people who made aggressive AI bets were not foolish; they were operating under genuine uncertainty and leaning toward action, which is usually the right instinct in fast-moving environments.

What the correction reveals is where the genuine value was all along: not in the technology itself as a product, but in the specific, well-considered applications of that technology to problems that users actually have and that the technology is actually equipped to address reliably.

That is the space where the best product thinking will operate for the foreseeable future. Not AI-first. Not AI-sceptical. But AI-where-it-genuinely-helps, assessed with the same rigour that has always separated good product decisions from bad ones.

The companies and PMs who get that right, in this moment of post-hype clarity, have an unusual opportunity. The noise has reduced. The users who were briefly impressed by AI novelty are now asking whether it actually helps them. The bar for what counts as a genuine AI product improvement has been raised by the experiences of the last two years.

Meeting that bar requires no special relationship with the technology. It requires the same discipline that good product management has always required.

A question worth sitting with: of the AI features your product has shipped in the last two years, which ones would you build again knowing what you know now, and which ones were driven more by pressure than by genuine user need?

The honest answer to that question is a useful guide to where the discipline needs to be applied more carefully going forward.