Part 1: Cheap subsidised AI has made our products lazy

Open any tech product in 2026 and you'll see the same feature in the same place. Summarise this Slack channel. Summarise this document. Summarise this meeting. Summarise this thread. Sometimes the summary is genuinely useful, when the use case warrants it. But most of the time the button is there because somebody could ship it, not because the use case asked for it.

I want to talk about why every product looks like this. Because it isn't a coincidence, and it isn't because users asked for it.

The reason every product looks like this is that AI is cheap right now. Not actually cheap. Just cheap to us. The labs running these models are burning through investor money at a rate that makes Uber's early years look modest, and the IPOs of OpenAI and Anthropic are circling. When those land, the burden of paying for all this shifts. Some of it goes to public markets. Most of it ends up with the people using the product, which is to say, all of us.

Cheap AI changes how products get built. When inference is effectively free, there's no friction on adding another AI feature. No exec asks whether the eighth summary button is worth it. No product team has to defend the cost. We've become blind to the questions that usually discipline product thinking, what does this cost to run and what return are we getting for it. The need for speed has won out over the patience to ask.

What this produces, in practice, is bolt-on. Every roadmap has an AI line item. Every product page has a sparkle icon. Every workflow has a button that calls a model whether or not the model is the right thing to call. Teams ship features quickly because the leadership chant is the same in every company: we need AI in this. The thinking happens, when it happens at all, after the feature is out.

And the laziness goes deeper than the feature itself. It runs all the way back into how the feature got built. Tokens get burned long before anything ships, on workflows nobody questioned, on prompts that did more than they needed to, on agents that ran twice because nobody bothered to set up the first one properly. AI is excellent at improving efficiency on bounded, low-risk tasks. We are using it for everything, including the parts where the speed gain is illusory and the cost is real.

The feature itself rarely starts in the right place. It doesn't start from a user, or a use case, or a real piece of context. It starts from the tool. We're not building things people need. We're finding places to put a model. This applies just as much to AI-native products as to legacy ones bolting on a chatbot. Same instinct, different starting position. The product team that should be asking what's actually worth building here is asking instead what can this model do that we can ship by Friday.

This isn't sustainable, and almost nobody is talking about it. The IPOs are coming. The subsidy isn't going to last forever. The price of inference will move closer to its real cost. And when that happens, every product team that spent the last two years adding AI to everything is going to be asked a much simpler question than they're used to. Is this feature worth what it now costs to run?

That's a healthier question than it sounds. Some features will earn their price tag easily. Others will get rescoped, narrowed, or quietly turned off, and that's fine. The harder bit is the conversation with users, because the bar AI has set is "look what this can do" and the bar after pricing arrives might be a little lower. Building back from that takes careful strategy, considered thinking, and a much better understanding of where AI actually belongs in the use case.

None of this is a complaint about AI. The tools are useful. They have changed what I can produce in a day, the kinds of problems I can take on, the speed at which I can move from idea to test. They have made me a more productive designer. And they have opened up fields I wouldn't have gone near two years ago. I am building and shipping my own software now. The rate of improvement isn't slowing, and I am genuinely intrigued about what's coming.

The laziness isn't the tool's fault. Lazy features are a direct response to how easily we've been able to ship, and to how comfortable that ease has made us. None of this is a problem we should be waiting on. The reckoning will force the question, but the question was always there. The work ahead, when the price of inference catches up, isn't to use less AI. It's to think harder about which uses earn their place. What that thinking looks like in practice is where this conversation goes next.