a16z wrote something last week that I've been turning over in my head: 2026 is "the death of the prompt box for mainstream users."
My first instinct was to skim past it. Another "AI is evolving" take. But then I actually read it.
Their point: the next wave of AI products has zero visible prompting. The app observes user behavior, reads context, and acts. The user just reviews the output — approve or reject.
That's not a UX tweak. That's a different product philosophy entirely.
A few months ago we were debating a feature for a clinical management product. The idea: AI reviews a patient's history before their appointment and surfaces a short list of things worth discussing in the visit.
Good idea. Then someone asked how it would work, and everyone defaulted to the same answer — "the doctor types a prompt, AI responds."
Why? Because that's still the mental model. Box, input, output.
But why should the doctor type anything? The system already knows the appointment. Knows the patient history. Knows the time. It could just act. Show up with the summary. The doctor dismisses it or uses it.
That's a meaningfully better product. And it has no prompt box.
Look at the fastest-growing AI tools right now — Cursor, Lovable, Bolt. What do they share?
The user describes an intent. The tool takes steps toward it. The user doesn't prompt each step; they steer the output. They accept, reject, redirect.
The difference sounds subtle. It's not. In one model, the user manages every step. In the other, the user manages the outcome.
That shift completely changes what a PM is actually designing.
Old model, the PM asks: What should the user type? How should the response display? What happens if the prompt is bad?
New model, the questions are different:
When should the system act without being asked? What context is it reading? At what threshold does it proceed autonomously versus pause for confirmation?
And the hardest one: Can the user understand why the system made that decision? Does it matter if they can't — as long as they can correct it?
These aren't AI feature questions. They're questions about the fundamental behavioral contract between a product and its user.
Here's what I think most PMs are missing about this shift.
Removing the prompt box doesn't mean removing user control. It actually means expanding it — but changing how it works. Instead of approving every action, the user sets the rules once. "Remind me 24 hours before appointments." "Draft invoices automatically but always ask before sending." "Never do this without my explicit confirmation."
Those preferences are prompts. They just don't look like prompts. And they take thirty seconds to configure instead of thirty seconds per use.
So what we're really designing is: an interface where the user tells the system how to behave, once — and the system remembers.
The harder design question is knowing when not to do this.
Proactive, autonomous AI isn't right for every product or every moment. A text input where a clinician writes a note mid-appointment — that should stay a text input. But a scheduling system that already knows every appointment, every patient, every gap — why is it still waiting for someone to ask?
Seeing that distinction clearly is, I think, where product work is moving.
The prompt box isn't dying everywhere. We're just finally getting better at knowing where it never belonged.