Products ship with the same AI panel; people hunt for the off switch, and spend has outrun returns. Build features with AI, not AI features.

Open any product you pay for and look at the top right corner. There is a sparkle. Behind it is a panel that slides in from the right, says “Ask anything”, and offers three suggestions you did not ask for. Close it and it comes back after the next update, sometimes with a new name.

I should be the last person to complain. I write code with AI every working day. When I built 131 variations of one button to study hover feedback, a model wrote most of them. My design files are generated from code that AI wrote. I am about as far from a sceptic as a designer gets, and I still spend part of every week looking for the switch that turns a sparkle off.

That is the gap this piece is about. The tools got genuinely good. The products built on top of them mostly did not, because most of them are the same product.

The same product, forty times

Here is the feature. A chat panel, docked right. A text box with a placeholder in the second person. Three suggested prompts written by someone who has never used the app for real work. A sparkle icon, because the sparkle is now the universal sign for “we added it”. It summarises the thing you were already reading. It drafts the thing you were about to write. It answers questions about the page you are on, with a disclaimer that it may be wrong.

I have seen that panel in a notes app, a spreadsheet, a design tool, a mail client, a team chat, a video call, a browser, a photo editor, a slide deck, a calendar, a music player and a bank. Not similar panels. The same panel. If you took the logos off, nobody on the team that built one could pick theirs out.

There is a simple reason it converged. The model vendors sell the same capability to everyone, through the same API, with the same demo. A company that buys the demo and ships it has not designed anything. It has forwarded it. The feature is named after the technology, “AI assistant”, “Copilot”, “Intelligence”, because there is no user goal to name it after. Nobody ever shipped a “database feature”. They shipped search.

None of this is a new observation. In 2023, before most of these panels existed, Amelia Wattenberger wrote Why Chatbots Are Not the Future and showed why a blank text box is the worst way to expose a capable model: it tells you nothing about what it can do. The same year Maggie Appleton published her Language Model Sketchbook, subtitled “or why I hate chatbots”, with three working alternatives, one of them a model that lives in the right-click menu of the document you are writing. The industry read both pieces and shipped the text box anyway.

Forced on, not opted in

The part people actually resent is not that the panel exists. It is that it was turned on for them, and that off is hard to find.

Microsoft is the clearest case because it is the biggest. Copilot went into Notepad, the Snipping Tool, Photos and Widgets, with the button in the toolbar and no plain way to remove it. In March 2026 the company said it would cut back on “unnecessary Copilot entry points”, starting with those four apps, and the head of Windows wrote that Microsoft would be more intentional about where Copilot goes. That is a company admitting, in public, that it had added AI to a text editor without a reason.

The reaction is not confined to grumpy enthusiasts. A Hacker News thread titled “I’m drowning in AI features I never asked for and I hate it” passed three hundred points, and the top complaints were from people who use AI daily by choice: they were objecting to the version they could not decline. Outside the industry it is wider still. In Pew’s June 2025 survey, half of American adults said AI’s growing role in daily life left them more concerned than excited, against one in ten who felt the other way.

Meta went further than Microsoft. In spring 2025 a blue and purple ring appeared in the WhatsApp chat list, then in the search bar, and there was no way to remove it. Meta called the feature optional. You could choose not to tap it. That is the definition of optional a company reaches for when the feature cannot survive a real one.

Every designer knows what a feature that people keep turning off is called. It is a bug with a marketing budget.

What the loop costs

The toggle that resets is a small thing, and that is why it does so much damage. A setting is a promise. The product said “this is yours to decide”, you decided, and an update decided back. The click to fix it costs nothing. What costs is what you learn from it: this product will override me.

Once learned, that rule does not stay on the one feature. Every update becomes something to audit. You stop reading release notes for what is new and start reading them for what has been done to you. People delay updates, pin old versions, and search for “disable” before they search for “how to”. That is the cognitive friction: not the dismissals themselves, but the vigilance they teach. Attention that should go to the work goes to guarding the interface.

Then it moves from the product to the brand. In January 2025 Apple paused its notification summaries for news apps after the feature rewrote a BBC alert into a false claim that a murder suspect had shot himself. The summary was on by default. The mistake was the model’s, but the name on the notification was Apple’s, and the BBC complained to Apple, not to the model. A feature forced on carries the brand on every error it makes.

The last stage is the strangest. In August 2024 the head of Procreate recorded a video to say the app would never add generative AI, and the artists who use it cheered. A promise not to ship a roadmap item went viral. That is what the loop buys, in the end: a market in which the absence of your feature is a selling point for your competitor.

Trust is cheap to lose here for a structural reason. One forced toggle does not teach “this feature is annoying”. It teaches “this company will override me”, and a rule about the company applies to everything it ships next, including the feature that would have been good.

The bill

So why does it keep shipping? Because of what it cost to build the thing underneath.

UBS now estimates the big cloud companies will spend about $4.1 trillion on capital expenditure between 2026 and 2028, after roughly $492 billion in 2025. In 2026 Amazon, Alphabet and Microsoft together are expected to spend 102 percent of their cloud revenue on capex. Every dollar that comes in from cloud goes back out as chips and buildings, and then some.

That money has to come back from somewhere, and the somewhere is us: every product team that buys tokens and every customer who pays for the seat. Which is why the panel is on by default. A feature that is off is a feature nobody is billed for.

Now put that next to what the buyers are getting. MIT’s NANDA group looked at 300 public deployments and interviewed 150 leaders, and found that 95 percent of companies saw no measurable return from their generative AI pilots. The 5 percent that did had integrated the tool into one specific workflow and learned from it. Gartner’s own number is kinder and still bleak: one in five AI initiatives delivers any measurable return, and one in fifty delivers something transformational.

Is this panic?

Not yet, and the word is wrong anyway. Panic is fast. What I see on product teams is slower and more embarrassing: the roadmap has an item called “AI” with no user problem under it, and for the first time someone with a budget is asking what it is for.

The newest version of the item is called an agentic strategy. Everyone has one. Almost nobody who has one can show you a map of how the work in their company actually gets done: who starts a task, what they open, who they wait on, where it stalls. An agent is a worker. You cannot hire a worker for a job you have not described, and the description is the part nobody wrote.

Those are the two halves of the same mistake. The spend assumed demand would appear once the capability existed. The product teams assumed a feature would appear once they had the API. Both skipped the step where you find a person with a goal and ask what would help. That step is called design, and it is the step that got cut because the board wanted the sparkle on the slide by the next quarter.

Jakob Nielsen called this technology the first new UI paradigm in sixty years: you state the outcome you want instead of the steps to get there. Take him at his word and the docked panel looks even stranger. A new paradigm arrived, and the industry gave it the interface of a 1990s support chat.

You can tell which features skipped it. They are the ones named after the technology. They are the ones that are on by default and hard to turn off, because a feature nobody asked for has to be forced to get used. And they are the ones whose success metric is “engagement with the panel” rather than anything the person came to do.

Features with AI, not AI features

I have one rule for my own work, and it has held up: build features with AI, not AI features.

The 131 buttons were a feature built with AI. A person wanted hover feedback that answered a real question. A model wrote the variations in an afternoon so I could spend my time judging them. The person who eventually uses the button sees no sparkle, no panel, no disclaimer. They see a button that behaves well. AI is in the process, not the product.

The reservation card in my last piece is the same. AI writes the states, the long French name, the failed photo, the slow network, so the design is judged in the material it ships in. Nobody on the receiving end has to know a model was involved, and nobody has to opt out of anything.

The sparkle panel is the opposite shape. AI is in the product, and the process that should have come first, finding the goal, never ran.

The ones you do not notice

It is fair to ask what the other kind looks like, so here are four I use, and none of them has a sparkle on it.

Gmail’s Smart Compose, from 2018, is the oldest and still the cleanest. You type, grey text appears ahead of your cursor, you press Tab or you keep typing. There is nothing to open and nothing to close. It is named after what it does, it lives in the box you were already typing in, and it does nothing until you are mid-sentence. Now look at the same inbox in 2026. A “Summarise this email” button sits on top of a three-line message. Same company, same model family, two different shapes, and only one of them was designed from a goal.

Photoshop’s Remove tool is a brush. You paint over the thing you want gone and it is gone. It sits beside the healing brush it replaces, in the toolbar where that job has always been. The model is in the tool and the name is not.

Figma’s rename layers command takes a chore nobody wanted and turns it into one keystroke. The result is a list of sensible names in the panel you were already looking at. No chat, no suggestions, no “Ask anything”.

Descript’s remove filler words is the same shape for audio. One command, the ums are gone, and you are back in the transcript.

Emily Campbell’s Shape of AI catalogues the interface patterns that have emerged around models, and it is worth an afternoon. What strikes me reading it is that the patterns that work, the ones she files under inputs and governors, are mostly invoked from inside the work, and the sparkle panel is not a pattern at all. It is a place to put the patterns you have not designed yet.

What they share is easy to list. Each is named after a verb. Each lives where the work already happens. Each is invoked, so there is nothing to turn off, because it does nothing until you ask. And each solves a problem a person could have described before the model existed. The model made the solution possible. It did not pick the problem.

Four questions before you ship one

If an AI feature is on your roadmap, these are the questions I would want answered before a line of it gets built. They are ordinary product questions. That is the point.

• Who asked? Not “would users like this”, but which person, on which task, got stuck in a way this unsticks. If the answer is “the board”, stop.
• What is it named after? If the name is the technology, there is no goal in it yet. Rename it after what it does for the person. If you can’t, you don’t have a feature.
• Would you ship it off by default? If the feature only earns its place when forced on, it has no place. The people who want it will find the switch; the ones who don’t should never have to.
• Would it survive at cost price? The tokens are subsidised today by that $4.1 trillion. Price the feature as if they weren’t. If it dies, it was never worth more than its marketing.

Most sparkle panels fail all four. The good AI features I have seen, and there are some, pass all four and have no sparkle on them, because they were designed the way everything else was: from a person with a goal, backwards.

In the room

The questions are easy to agree with and hard to use. I lead design at a fintech and report to the CEO, and I have no budget authority either. Here is what has worked for me when the roadmap arrives with “AI” on it and no problem under it.

Do not argue with the strategy. “Should we have an agentic strategy” is a yes or no question with a board behind the yes, and you lose it in the room. Change the question to “which job”. Nobody with a strategy can object to being asked where it applies.

Volunteer for the map. Offer two weeks to document how three real tasks get done today: who starts them, what they open, who they wait on, where they stall. Leadership rarely refuses, because it is cheap and it looks like progress on the strategy. The map is where an agent either finds a job or does not, and you are the one holding it. You have moved the decision from a slide to evidence without saying no to anyone.

Bring the four questions as a form, not an objection. A one-page brief that every AI item fills in before it gets built. Nobody can argue against filling in a form. Items that cannot fill it in die quietly, and nobody has to be the person who killed them.

Rename it in the ticket. “AI assistant” becomes “Draft the reply” or “Find the duplicate”. When there is no verb to rename it after, that shows in the ticket, and the product manager notices before you have to point.

Price it. Ask what the tokens cost per seat at list price and put that beside the price of the seat. You do not need budget authority to ask a question whose answer has a currency sign on it.

Propose the opt-in version as the experiment. Ship it off by default and measure who turns it on and keeps it on. That is a real signal, and it is the version you wanted anyway. If leadership insists on default-on, ask for an off switch that is one tap away and survives updates. That is a fallback you can usually win, and it protects the thing the loop above destroys.

And when you lose, lose well. Ship it, instrument it with a task metric rather than panel opens, and bring the number back in a quarter. Being right in a quarter beats being right in the meeting, and it is the only kind of right that changes the next roadmap.

The sparkle is a confession

Nobody chose the icon on purpose, either. Fast Company traced how the sparkle took over AI and found no first use, only a drift from the emoji people used when they wrote about models to the glyph on the button, because it read as magic and nobody had a better idea. A sparkle icon tells you two things. The company has a model, and it did not know what to do with it. It is a label that says “AI goes here” in a product where nobody found the place.

The investment will sort itself out or it won’t. I am not an economist, and the people who are don’t agree. But the product side is not a mystery, and it is not a bubble. It is a generation of teams that were handed a capability and told to ship it, and skipped the part of the job that would have told them where.

Find the person. Find the goal. Build the feature with the AI, then take the sparkle off. Nobody will miss it.

Nobody asked for the sparkle button was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.