02

Blog · November 20, 2026 · 6 min read

AI-Built Prototype Review: What We Let AI Generate and What a Senior Designer Signs Off

A founder can now describe a product in a chat box and have a clickable, running app by the afternoon. We think that is one of the best things to happen to product design in a decade. We build most of our prototypes the same way, as working software, because a coded prototype answers questions a static mockup can only guess at.

The trouble starts the day that prototype meets real people. A generated app is built for one user, one role, sample data, and a good day. Real users bring the other days. This post explains how we split the work between AI and senior designers when we build prototypes, what our review covers before anything goes in front of a customer, and what we refuse to ship no matter how good the demo looks.

Our rule: AI drafts the surface, a senior designer owns the consequences

The short version of our process is simple. AI handles anything that gets cheaper with more attempts. A senior designer handles anything where one wrong call costs a user money, data, or trust.

In a complex or regulated product, that line matters more than usual. A fintech transfer screen, a health intake form, an insurance quote, or an AI agent that can take actions on someone's behalf all carry consequences a model cannot see from the prompt. The model sees layout. The senior designer sees the person who just sent rent to the wrong account.

What we let AI generate

The first working version of the main flow. We describe the job the user is trying to do, the data they have, and the outcome they need, and we let the model scaffold screens, routes, and components. Ten variations cost almost nothing, so we generate several and throw most of them away.

Layout exploration. Models are fast at trying density, ordering, and grouping. We use that speed to compare three ways of presenting the same information side by side in a browser, where real scrolling and real screen sizes expose problems a static frame hides.

Realistic sample data. Long names, empty fields, foreign currencies, huge numbers, a balance of zero. We ask the model to generate the ugly data on purpose, because generated apps tend to be tested only with tidy data.

Copy drafts for low-stakes surfaces. Empty state hints, tooltips, and first-pass labels. Every word on a high-stakes screen gets rewritten by a person, but a draft helps the conversation start.

Boilerplate states. We prompt for loading, empty, and error versions of every screen as a checklist item. The model produces them quickly. Whether they are right is a separate question, and that question belongs to a person.

What a senior designer reviews before real users see it

Every state, on the running app. We review in the browser, never from screenshots. Loading, empty, partial, expired, offline, permission denied, and the retry after a failure. A generated app usually has a beautiful main screen and a spinner or a blank page for everything else. Users spend a surprising amount of time in "everything else," especially in products that move money or handle claims.

Roles and permissions. Demos run as an admin. Real products have viewers, approvers, members, and people from other companies who must never see each other's data. A senior designer walks through the product as each role and asks what that person should be able to see, change, and undo. Most of the scary bugs we find in generated prototypes live here.

Irreversible actions. Anything that sends money, submits a claim, deletes records, or lets an AI agent act needs a confirmation that shows the real consequence in plain words, plus a clear way back when one exists. Models are good at adding a confirm dialog. They are much weaker at deciding what the dialog must say for a nervous person at 11pm.

Trust and compliance moments. Disclosures, consent, identity checks, fee display, and audit trails. These are decided by someone who knows the product's regulatory context and has read the actual requirements. A model will produce something that looks like a disclosure. Looking like one is the risk.

Component drift. Each prompt tends to invent its own button, table, and modal. After a week you have four shades of gray and three corner radii with no names. We consolidate into a small set of tokens and components before the prototype grows, because drift compounds and cleanup gets more expensive every day.

Accessibility basics. Keyboard focus order, labels on every input, contrast, target sizes, and screen reader behavior on the main flow. Generated code often looks accessible and fails the moment you put the mouse away.

The story the screens tell. A senior designer asks one question across the whole flow: does a first-time user understand what just happened and what happens next? No automated check answers that, and it is usually the difference between a prototype people tolerate and one they trust.

Where we draw the line

Some decisions never go to a model in our studio, however good the output looks:

  • What the product should do. Scope, priorities, and what to leave out come from research and judgment about the business.
  • Wording on high-stakes screens. Fees, errors in payment flows, health information, denials, and anything with legal weight is written and reviewed by a person.
  • Whether a finding is real. If a usability session suggests a change, a senior person decides whether it reflects a pattern or one tired participant.
  • Whether it is ready. "Ready for real users" is a sign-off with a name attached. A model has no name to attach.

What we reject on sight in an AI-built prototype

These come up often enough that we keep a standing list:

  • A main flow with no designed error state, only a generic "Something went wrong."
  • A spinner with no text on any action that moves money or submits data.
  • An admin view used as the only view because that is how the demo was built.
  • Confirm dialogs that ask "Are you sure?" without saying what will happen.
  • Real customer data loaded into a prototype that has no access controls.
  • Copy that promises speed, security, or accuracy the product has not earned yet.
  • Five slightly different buttons on the same screen.
  • An AI feature that acts without showing what it is about to do or offering a way to stop it.

Why we work this way

We could slow down and draw everything by hand in a design tool first. We choose working software because it surfaces the hard questions earlier, in the medium users actually touch. AI makes that affordable for a small senior team, and it lets us spend our hours where they matter most: on the states, roles, and moments of risk that decide whether people come back.

We could also let AI carry the work all the way to launch. We do not, because the cost of a confident, wrong screen in a regulated product lands on a real person, and on the company's reputation with them. Speed from the model, judgment from senior people, and a named human sign-off at the end. That combination is the whole method.

A quick self-check for your own AI-built prototype

If you built your prototype with AI and want to know whether it is ready for real users, try this in an hour:

  1. Log in as every role your product has and try to do something that role should not be allowed to do.
  2. Turn off your network halfway through the main flow and see what the user sees.
  3. Load the ugliest realistic data you can imagine and check every screen.
  4. Read every confirm dialog out loud and ask whether a stranger would know what happens next.
  5. Put the mouse away and finish the main task with only the keyboard.

If more than one of those turns up a surprise, the prototype is doing its job as a prototype. It is telling you what still needs a designer's judgment before launch.

Work with a senior team that builds this way

We are a boutique studio that builds working prototypes with AI and keeps senior people accountable for every decision that touches a user's money, data, or trust. If you have an AI-built prototype and want a clear read on what it needs before real users see it, we can help.

Book a discovery call or email ash@uxsignalstudio.com.

Let’s talk about your next step

Have a question or a project in mind?

Tell us where you are and what you want to figure out. We’ll review your message and reply within two business days.

Contact UX Signal Studio

We’ll use your details to respond to your inquiry. This form does not subscribe you to a mailing list. Privacy policy.

Prefer to talk? Book a discovery call

Get the next essay in your inbox

One practical piece on UX and product design, every week or two. No noise.