AI-Native Development

How to Evaluate an AI-Assisted App Prototype Before Release

A release review for founders who received a prototype built with AI tools, with a scenario table, data and dependency checks, failure tests and a clear release decision.

Ragip Diler, founder and director of Gestures Studio
Illustrative drawing of a review table with scenarios checked one by one and one open item selected.

An AI-assisted app prototype is ready for release when its full user journey, saved data, accounts, external services, failure states and usage limits have been checked against written expectations, and every open issue has an owner and a decision. A prototype that only works in the demo is a prototype that has not been reviewed yet.

This guide is for founders and product leads who received an app built with AI coding tools, from a studio, a contractor or their own team. It gives you one review table, the checks that fill it and a way to make the release decision. The review covers product behavior and delivery evidence. A security audit or penetration test is a separate engagement, and an app that handles payments or sensitive personal data needs one before a public release.

AI tools make a convincing demo cheap, which is why the review matters. In Veracode’s 2025 GenAI Code Security Report, models given a choice between a secure and an insecure way to write code chose the insecure one in 45% of 80 coding tasks. Google’s 2024 DORA report estimated a 7.2% drop in delivery stability for every 25% increase in AI adoption. Neither figure describes your prototype. Both describe why it needs evidence before it ships.

Use one review table for every check

Record every check in one table with five columns: the scenario, the expected result, the observed result, the owner and the open item. Writing the expected result before testing turns “it seems fine” into a pass or a fail, and the owner column makes sure nothing waits on nobody.

The rows below come from a fictional class booking app for a climbing gym, the same example used in the mobile app brief. The app was built with AI coding tools and has no AI features of its own.

Use one review table for every check
Scenario Expected result Observed result Owner Open item
New member books and pays for a class Booking appears in the gym’s system and in the app Pass Developer None
Member closes the app mid-payment and reopens it Booking is either complete or clearly not started Booking shows as pending with no way forward Developer Fix before release
Member signs in on a second phone Same bookings appear on both devices Pass Developer None
Member A changes a booking ID in a request Request is refused Member B’s booking details returned Developer Blocker
Gym’s booking API is unavailable Clear message, booking can be retried Spinner never stops Developer Fix before release
Member deletes their account Account and personal data removed from the backend Account hidden, data kept Product lead Fix before release

Run the scenarios on the agreed test devices with real accounts, not only in the builder’s browser preview. A scenario passes only when the observed result matches the expected result exactly.

Check the full user journey and saved state

Check that a person can complete the whole journey on a real phone and that everything they did is still there afterward. AI-generated prototypes often handle the screen in front of them well and lose state between screens, sessions or devices.

Check the full user journey and saved state
Check What to do Passes when
End-to-end journey Complete the core task from first launch to result No step needs a workaround or a reload
Saved state Close and reopen the app, then sign out and back in Everything the person created is still there
Second device Sign in on another phone The same data appears on both
Empty states Use a new account with no data Every screen explains what to do first
Interrupted actions Leave mid-task, lose connection, return The task resumes or clearly restarts
Time and region Change the phone’s time zone and language Dates, times and prices still read correctly

Data that only lives on the device is a common gap. If saved items disappear after a reinstall, the prototype is storing them locally, and the brief needs to say whether that is acceptable.

Review accounts, data boundaries and external dependencies

Confirm that each person can only reach their own data, that no secret key ships inside the app, and that every outside service the app depends on has an owner and a plan for when it fails. These are the checks a demo never shows.

  • Data boundaries. Sign in as two test users and try to open one user’s records from the other account, including by changing IDs in links or requests. Any success is a blocker.
  • Secrets. Ask where API keys and service credentials live. They belong on the server or in a secrets manager, never in the app bundle or the repository history.
  • Roles. If the app has admins or staff, check that a regular account cannot reach their screens or actions.
  • Account deletion. Apps with accounts must offer deletion inside the app, one of the store rules covered in the mobile app brief. Check that deletion removes the record, not only hides it.
  • Dependencies. List every outside service, such as payments, maps, email and AI APIs, with its account owner, its cost and what the app shows when it is down.

If the product has its own AI feature

Test the AI feature against the risks the OWASP Top 10 for LLM Applications 2025 names first: prompt injection, sensitive information disclosure and unbounded consumption.

Review accounts, data boundaries and external dependencies
Risk Test Passes when
Prompt injection Put instructions inside content the AI reads, such as “ignore your rules” in a message The feature keeps to its task
Sensitive information Ask the feature for other users’ data or its own instructions It refuses, and no private data appears
Unbounded consumption Send many requests quickly and very long inputs Limits apply and costs stay capped
Unsuitable input Upload the wrong file type, an empty input or an unclear photo A clear message explains what to try
Wrong output Review a set of example inputs and outputs Bad outputs are handled before the person sees them

Inspect failure recovery and usage limits

Make the app fail on purpose and check that a person can recover. A release-ready app handles a slow network, a declined payment and a busy service with a clear message and a next step, and it caps the usage that costs money.

Inspect failure recovery and usage limits
Failure How to test it A ready app
No connection Turn on airplane mode mid-task Explains the problem and keeps the person’s input
Slow response Throttle the network Shows progress and stays usable
Declined payment Use a test card that declines Keeps the selection and explains what to do
Double submit Tap the main button twice Creates one record, not two
Service limit Exceed an API’s rate limit in testing Waits and retries, or explains the delay
Crash Force an error in a test build Crash reporting records it with enough detail to fix

Ask for the usage limits in writing: requests per user, file sizes, and the monthly spend at which paid services stop or alert someone. Running costs that nobody capped are a release issue, not an accounting detail.

Agree the release decision and unresolved issues

Close the review with one of four decisions, written down with the open items and their owners. A clear decision is more useful than a long list of findings, because it tells everyone what happens next.

Agree the release decision and unresolved issues
Decision When it fits What happens next
Release Every scenario passes and no open item affects people’s data or money Ship and monitor
Release with known limitations Open items are minor and documented for users or the team Ship, then fix in the next version
Fix first Any item touches data access, payments, deletion or crashes Fix, then rerun the affected scenarios
Rebuild a part One area keeps failing because of how it was built Scope the rebuild before more features

Classify each open item as a blocker, a fix before release or a later improvement, and keep the classification with the item. Blockers and data issues are never “release with known limitations.”

If the review shows the work is earlier than expected, compare a prototype vs MVP vs production app and choose the stage the evidence supports. For the questions to ask before work starts, see the AI-native development process guide.

Frequently asked questions

Is AI-generated code safe to release?

AI-generated code is safe to release when it has passed the same review as any other code: a person checks it, tests cover the main journey and failures, and data access and secrets are verified. In Veracode’s 2025 research, models chose an insecure option in 45% of tasks, so unreviewed AI output is not release evidence.

Who should review an AI-assisted prototype?

A developer who did not build the prototype should review the code and the checks in this guide, and the product owner should confirm the journey and the release decision. Apps that handle payments or sensitive data also need a dedicated security review.

How long does a prototype review take?

The review takes as long as the scenarios in the table. Start with the core journey, data boundaries and failure tests, since those decide whether the app can be released at all.

What is the difference between a prototype review and a security audit?

A prototype review checks that the product works, keeps people’s data apart and recovers from failure. A security audit or penetration test looks for ways an attacker could break in, and specialists run it as a separate engagement.

Can we release with known issues?

Yes, when the issues are minor, documented and owned, and none of them affects people’s data, payments or account deletion. Anything that touches those areas is fixed first.

Discuss the next stage of your prototype

Bring the prototype and your review table, even if it is half filled. Explore how Gestures approaches AI-native prototyping and product building, and we can work out what the evidence supports next.

Working through a decision like this?

Bring your current product or website and we can discuss what needs attention.

Start your project