How to test an AI-created application: an accessible method

To test an AI-created application, begin with situations whose expected results you can describe. Try the normal journey, input errors and relevant interruptions. Keep observations so you can check that a correction solves the problem without damaging working features.

Write the result before testing

For a form, state what should appear in the summary. For a calculation, prepare an independent answer. For a transmission, define what should exist in the destination service.

“Everything should work” does not help you decide whether a trial succeeds. Prefer: “The four answers appear unchanged in the summary after returning to the form and opening the summary again.”

Prepare contrasting cases

Start with complete, ordinary input. Add an empty required field, long text, accented characters, zero and an out-of-range value wherever these cases are relevant.

Also test action order: going back, double-clicking and trying again. These cases can reveal lost answers or duplicates that an ideal journey hides. Keep the test data so you can repeat the checks after changes.

Let someone discover the tool without directing every click

Give a colleague a business task and observe. Record hesitations and mistaken interpretations. A tool can produce the correct answer while repeatedly encouraging users to enter the wrong information.

Check keyboard navigation and a small screen. W3C emphasises appropriate labels and notifications in forms. A message should help users understand where an error is and how to correct it without re-entering everything.

Describe a defect so it can be reproduced

Record the actions, inputs, observed result and expected result. Add useful context such as screen type. “After Back, the Audience field is empty” enables a targeted correction.

Rank defects by consequence. An incorrect total or falsely announced action takes priority over a decorative misalignment. Ranking depends on use: a seemingly small error may have a significant effect in real work.

Repeat checks after a correction

Repeat the failed scenario and several previously validated journeys. If the project changes frequently, have a competent person automate relevant repetitive checks.

Keep an explicit trial decision: usable within the planned scope, needs correction or needs further examination. Your business tests provide useful evidence, but do not replace technical review of access, data and operation when the project requires it.

Do you have a business task to explore with AI? Discuss your needs with eloRezo.

Content and sources checked on 30 October 2026.