Prototypes now cost an afternoon, so the hard work is judging which one ships

Something shifted this year, and it is not another design tool. The way work moves through a product team is getting rebuilt. Prototypes cost an afternoon now. Agents are becoming a second user of your app. Voice is turning into a real surface, not a gimmick. All of that hits how you plan features and run reviews. Let me catch you up.

The deep cut

• Cheap making moves the hard part to judgment. Once a prototype costs an afternoon, the scarce work is deciding which of the twenty versions ships.
• The design system becomes a contract with the agent. AX means a bot needs to know your objects and verbs, not just your buttons.
• Structure is the product, and you deleted yours. Teams wiring up AI find no legible map and rebuild the information architecture they dropped a decade ago.

When the second diamond stops costing money

For twenty years, teams front-loaded research because building was the part you could not take back. That math flipped. Stanford HAI's 2025 AI Index found the cost of querying a GPT-3.5-level model dropped 280-fold in 18 months. A candidate solution used to cost a sprint. Now it costs a prompt.

So the exploration did not die. It moved to the back of the process, where you build five versions instead of mapping twenty on a wall. Figma's own report backs this up: among teams whose AI projects met or beat expectations, 60% had explored multiple approaches, against 39% of the ones that fell short.

The catch: making is nearly free, so the cost lands on knowing what is good. Call it your judgment budget. A brief with no standard of "good" gets you forty screens by Thursday and no decision by Friday. Write the eval, the criteria you test output against, before anything gets made. That is your new spec.

Your app has a second user now

The old assumption was one human behind the screen, clicking through flows. That is breaking. Users describe what they want and let an agent figure out the steps. The shift from UX to AX, agentic experience, means your app has two audiences. A human sees a "Generate Report" button and knows what it does. An agent needs structured information about what that capability accepts and whether it has permission to run it.

Your design system stops being just components and becomes a contract. The agent needs to know which objects exist, what actions are valid, and which ones require human confirmation. The good news is the same accessibility work you already fund, semantic HTML, ARIA labels, meaningful names, is exactly the structure an agent reads. You are not starting from zero. You are finishing a job you half-did.

The house nobody can find their way through

Here is where it goes wrong fast. Cheap features spun up in an afternoon, each with its own vocabulary and its own idea of what happens when something breaks. Patrick Neeman calls it a Winchester Mystery House, staircases climbing into ceilings, doors opening onto walls. Nine in ten organizations use AI, he notes, but most have not rewired their workflow around it. The features arrived faster than any structure to hold them.

The fix is not glamorous. Define your object model before the feature list: the real nouns and the verbs scoped to each. When a feature offers to "create a follow-up," that has to land on a real object with a defined home. If follow-up is not an object anyone named, the feature just invented a room on no floor.

And kill the global "ask anything" box. It looks generous in review and behaves like a séance once it ships. Scope what a feature can do to where the user is standing. Skip that and the complexity you refused to absorb gets handed to the user.

Voice is a surface with its own rules

Voice is not one thing, and treating it as one is a fast way to build the wrong thing. The eight voice patterns worth knowing run from lifelike avatar calls to plain dictation, and each carries a different tradeoff. On a phone call, nothing stays on screen, so a date or a reference number has to be confirmed in the next turn or texted over. Push-to-talk suits noisy rooms; voice-activity-detection suits hands-free back-and-forth.

The design payoff is real when you match the pattern to the job. Vapi rewrote an agent to sound like a colleague instead of a corporate script, no change to the flow, and outbound conversion went from 14% to 22%. Personality, carried by word choice and pace, moved the number. Pick the pattern first, then sweat the copy.

Three questions for your team

• Where is our judgment budget going? If making is now cheap, do we have a written standard of "good" we test output against, or are we shipping whatever the tool put on top?
• Can an agent understand what our product does without guessing? Have we named our objects and verbs, or is our design system still just buttons?
• Which of our recent features invented a room on no floor? Whose job is it, this cycle, to redraw the map and retire the sealed-off wings?