Somewhere in the last two years, adding AI to a product turned into adding a chat window to a product. A little bubble in the bottom right, a sparkle icon in the toolbar, a panel that slides out and asks how it can help. It became the default move, and most of the time nobody asked for it.
We went through this ourselves. When we rebuilt tidi earlier this year, the obvious version had a chat panel on the side. Every canvas tool has one now. We built the product without it, and the AI reaches the canvas through MCP instead, which means an agent can create and move and read cards without a human typing into a box inside our app. The interesting part is what that does to the interface. There’s nothing to explain, no empty state telling you what to ask, no second way to do everything you can already do.
The bubble is usually a tell. It means the AI got added at the end, so there was nowhere for it to go except on top.
Sprinkled versus structural
There’s a real difference between a product that uses AI and a product with an AI feature.
Sprinkled AI sits next to the work. Summarize this. Draft that. Ask a question about the thing you are looking at. It demos well and it’s the first thing to get ignored, because using it costs the user more effort than doing the task themselves.
Structural AI sits underneath the work and the user never names it. It’s the matching that puts the right two people in front of each other. It’s the parsing that turns an uploaded document into something the app understands. It’s the reconciliation that makes a number correct. Nobody opens the app to use it, and nobody would keep using the app without it.
The second kind is harder to build and much harder to market, which is probably why so many products shipped the first kind.
A text box is a hard interface
The thing that gets lost in all of this is that chat is a genuinely difficult interface, and we started treating it as the easy option because it’s the fastest one to build.
A normal interface shows you what it can do. Buttons, menus, a canvas with things on it. You learn the capability surface by looking at it. A text box shows you nothing, so the user has to guess what the product understands, and most people guess low. They try one obvious thing, get a mediocre answer, and never open it again. That’s the real reason those panels go unused, and it isn’t a prompt problem.
Chat also moves work onto the user. A button is one click. The same action in chat is a sentence you have to compose, then read back, then verify. For anything you do more than once, that’s a worse deal, and people notice quickly.
Then there’s the part that makes it hard to trust. Ask the same thing twice and you can get two different answers, with no way to see what it did or nudge it slightly. Direct manipulation has been winning for forty years for a good reason. You can see the object, you can move it, and you can undo it.
None of that makes chat wrong. It’s the right interface when the space of things someone might ask is genuinely unbounded, like writing or open-ended exploration, where no menu could ever hold the options. It’s the wrong interface for work that’s structured and repeated, which is most of what products are for.
The part that surprised us
The most interesting product feedback we’ve seen this year was not about a feature. It was people going out of their way to say they were glad there was no AI in the thing they were using.
That’s worth sitting with. Not neutral, not “didn’t notice.” Relieved. A year of every tool sprouting a sparkle icon has trained a real group of people to read AI in the interface as a signal that the product is going to guess at their work, get it wrong in a confident tone, and make them check it.
So “no AI” is now a positioning available to you. Not forever, and not in every category, but right now it lands, and almost nobody is using it because it sounds like admitting you’re behind.
What we’d actually build
If AI makes the product do something it genuinely couldn’t do before, build it, and bury it. Put it in the pipeline, not in a panel. Let it show up as the product being unusually good at something rather than as a place to type.
If it doesn’t, skip it. The absence is not a gap you need to apologize for, and shipping a chat window to cover for that is how you end up maintaining a surface nobody uses.
The one place we’d spend the effort instead is making the product reachable from the outside, which is the thing we wrote about in May. An MCP server or a clean API means agents can use what you built, in whatever tool your user already lives in. That’s AI in the product without a single pixel of AI in the interface, and it’s the version people keep.