When a Chat Box Beats a Button (and When It Doesn't)

Three big consumer apps shipped AI chat into their main flows in the same week. Spotify, Waze, and Superhuman all decided you should be able to talk to the app instead of tapping through it. That is a pattern worth reading closely, because your team is about to face the same question: does this flow deserve a chat box, or does it just need a better screen?

Let me catch you up on what shipped, what it tells you about where chat helps, and where it quietly adds work.

Talk when the menu can't hold the request

Spotify's new assistant lets Premium users type or speak to shape what plays. You can ask for artists you have never heard, then add Taylor Swift, then narrow to her newer stuff, all in one back-and-forth. No menu could hold that many turns without becoming a maze.

That is the real tell. Chat earns its place when the request has too many combinations to fit on a screen. Mood, tempo, era, and artist mixed together is a hard thing to design buttons for. A sentence handles it fine.

The same logic shows up in Waze. Its new Gemini search lets you ask for "a coffee shop that's open right now" or "parking close to Grand Mall". Open now, near a place, cheapest gas. Those filters stacked together are annoying to tap and easy to say.

The best move might be less chat, not more

Here is the part that should stick with you. In the same Waze release that added a talking assistant, the team also shipped a "less chatty" mode. It cuts the voice prompts so your podcast doesn't get interrupted, keeping only turns and hazards.

So one team, one launch, went both directions at once. More conversation where you're searching, less where you're just driving. That is a mature read of when talking helps and when it gets in the way. Not every moment wants a voice in it.

Waze also added personalized navigation that learns your driving habits with no chat at all. The lesson: AI in the flow does not mean a chat box everywhere. Sometimes it means the screen just gets smarter and stays quiet.

Chat is fine until it acts on your behalf

Asking is low risk. If Spotify plays the wrong song, you skip it. Nobody gets hurt. That is why music and search are safe places to start with a chat layer.

Sending is a different story. Superhuman's new auto-draft reads your important emails and writes replies in your tone, plus two variations. When it works, you send with almost no edits. When it doesn't, it agreed to a meeting at a post-midnight time, and by default it leaned toward saying yes to a pitch.

That gap matters for your roadmap. A wrong song costs nothing. A wrong email costs your credibility. Co-founder Rahul Vohra says 40% of drafts got sent within a day, and 60% of those went out with no editing. Good numbers, but they only hold because the human still hits send. Keep that checkpoint in any flow where the AI acts, not just answers.

The data you already own is the real draw

The flashy demo is telling Spotify to make a playlist more upbeat. The stickier feature is quieter: you can ask when you first played a song, or what genres you've leaned on lately. Spotify already had that data. Wrapped proved people love it. Chat just makes it available on demand instead of once a year.

That is the part worth copying. A chat layer over data your users can't easily reach is more valuable than one that only reshuffles the same buttons. The question is not "can they talk to it" but "does talking reach something the screen never surfaced."

The deep cut

All three companies used a mix of models under the hood, picking whatever fits the task. Spotify and Superhuman both said so plainly. That means the chat box is not the product. It is a thin front on a stack you can swap and tune.

So don't spend your review arguing about whether to add a chat feature. Spend it on two things: which task has too many combinations for a screen, and what happens when the AI is wrong. If the request is complex and a wrong answer is cheap, chat is a strong bet. If a wrong answer is expensive, keep a human hand on the send button. Granola's rebrand is the tell here too: it took the notes so people could stay in the conversation. The AI did the boring part and left the human the part they wanted. That is the bar.

Three questions for your team

1. Name one flow where the request has too many combinations to fit on a screen. That is your best candidate for chat. What is stopping us from testing it there first?
2. For any flow where the AI acts instead of answers, where is the human checkpoint, and what does a wrong action cost the user?
3. What data do we already own that users can't easily reach today? Could a chat layer surface it, the way Spotify opened up listening history on demand?