Why most "AI features" should be a button
Most of the AI bolted into products in 2026 makes them worse, not better. Here is when to ship it and when to walk away.
The pattern we keep seeing
A team has a genuinely useful model: it can summarize, extract, classify, draft. Then the instinct kicks in, and instead of shipping the useful thing, they ship a chat box that can theoretically do the useful thing, plus eleven other things nobody asked for, with a blank text field and a cursor blinking at the user like a dare.
The chat box feels like the more impressive deliverable. It demos well once. It converts terribly, because most users don’t know what to type, don’t want to type anything, and already know exactly what they’re trying to do. They just want the summary, not a conversation about getting the summary.
What we build instead
Scope the interaction to the one thing people actually do. If 80% of the chat traffic turns out to be some version of “summarize this thread” or “draft a reply to this,” that’s not evidence people want a general assistant. That’s evidence people want a button labeled exactly that, placed exactly where the task happens.
Make the input implicit. The best AI features don’t ask the user to describe what they want. They read it off the screen the user is already looking at: the thread they have open, the record they’re viewing, the file they just uploaded. A button that says “Summarize this thread” needs zero typing, because the thread is already the input.
Keep a narrow fallback, not a wide one. When the single-purpose action can’t handle a request (an edge case, something genuinely novel), the fallback is a clearly labeled escalation, a “something more specific? use the full assistant” link, not a wall of chat history the user has to scroll past to find the one thing they need today.
A rough decision framework
We ask three questions before deciding chat vs. button vs. nothing:
- Is there one dominant intent, or many genuinely different ones? One dominant intent (check order status, summarize a document, draft a reply) means a button. Many unpredictable intents across a broad surface (an internal tool used by five different teams for five different jobs) means a scoped chat interface might actually earn its complexity.
- Does the user already have the input in front of them? If the input is already on screen, forcing them to re-type or re-describe it in a chat box is pure friction. Pull it automatically.
- What’s the cost of a wrong answer? High-stakes actions (sending an email on someone’s behalf, modifying a record) need a visible preview and a confirm step regardless of interface. Low-stakes actions (a draft, a suggestion) can run eagerly and just show the result.
What this looked like in practice
On one project, the original spec called for an “AI assistant” sidebar: open-ended chat, available everywhere in the app. We shipped three buttons instead, each tied to a specific screen: “Summarize this ticket,” “Suggest a response,” “Flag as urgent if needed.” Usage went up, not down, because every click did something the user had already decided to do, with zero typing and zero ambiguity about what would happen next. The open-ended chat version, prototyped for comparison, got used for exactly the same three things, just with more friction getting there.
When chat actually is the right call
Chat earns its place when the task genuinely is open-ended and conversational: research, exploration, something where the user doesn’t know the shape of their own question yet. Our own site’s assistant, Verse, is scoped chat for exactly that reason, visitors arrive with genuinely different questions about scope, timeline, and fit. But even there, we constrain it hard: a fixed set of things it will and won’t do, suggested prompts instead of a blank box, and an explicit handoff the moment a question needs a real person. Unconstrained chat is rarely the honest answer. Constrained chat, or no chat at all, usually is.