Guides

Built an app with AI? What to check before using it at work

Test workflows, data, permissions and deployment to see what the prototype still needs before rollout.

Updated on 6 min
An application prototype connected to an operational system through checks on access, data and recovery.

Before using an AI-built app at work, check that it does the job with realistic users, data and errors. A successful demo tells you little about permissions, data accuracy or whether you can recover after a failure.

AI can help you build a first version. The next step is to understand what you have and what still needs testing. The same principle applies to software built with other tools. Start from what the app needs to do and check how it behaves.

Define one use case with clear boundaries

“The team will use the management app” is too broad to test. Pick a workflow, such as receiving a request, assigning it, working through it and recording the outcome. Identify the allowed users, the data involved and who approves the result.

For that workflow, write down both the expected behavior and the actions that must be prevented. A first version with observable behavior gives you something you can test.

Test the whole workflow, including interruptions

Reopen the record after saving the form. Edit it from another account, check the report that uses it and verify any message sent to an external service. For multi-step actions, record what happens when one step fails.

Suggested test cases for a first release
CaseWhat to checkEvidence to keep
Complete dataThe workflow reaches the expected outcomeSteps, user and final record.
Required field missingThe app explains what is missing and leaves data consistentMessage and data state.
Repeated submissionThe duplicate operation is prevented or clearly identifiedRecords, outgoing messages and IDs created.
User with a different roleActions and information respect permissionsAccounts used and the outcome of each attempt.
External service unavailableThe failure is visible and there is a recovery pathAlert, attempt and next step.
Reopening after an interruptionIt is clear what was saved and what still needs doingState before and after resuming.

Use these cases as a starting point. Define checks specific to payments, documents, sensitive data and actions that are difficult to undo before involving real users.

Check sign-in and permissions separately

Sign-in establishes who the user is. Authorization determines which resources and actions they can access. OWASP distinguishes the two and recommends checking permissions on every request. Verify the action itself as well as what the interface displays. OWASP Authorization Cheat Sheet.

Set up accounts with different roles and verify allowed and forbidden operations. If the app separates customers, check that an account cannot read or change another company's data. Include APIs, exports and files when they are in scope.

Check what happens to the data

Where does the data live? How are customers, tasks and documents related? Does a change appear in every relevant view? Do people understand what happens when a record is deleted? Answer these questions before bringing in operational data.

For recovery, consider code, database records and attachments separately. PostgreSQL, for example, documents different tools and steps for dumping and restoring a database. The plan for your app should explain what is backed up and how the restore will be tested. PostgreSQL, SQL Dump.

Use demo data in tests where possible. If you need real data, agree on the environment, authorization and handling with its owner.

Review dependencies and credentials

Dependencies are the external components the application uses. Identify them, how they get updated and which services require credentials. In npm projects, npm audit reports known vulnerabilities in the dependencies it checks. Use separate checks for business logic and permissions. npm Docs, npm audit.

Evaluate alerts in the context of the project. Before applying an update, check compatibility and the workflows it may affect. For credentials, check where they are stored, who may use them and how to replace them while keeping services running.

Prepare the release and a way back

Before the app goes into use, identify the version to release, who will release it and how to check the result. Use a test environment to run through the agreed steps before applying them in production.

Rolling back can involve code and data. If a release changes the data or database schema, publishing the previous code may be insufficient. Identify what you can restore, which data might be lost and who decides when to suspend use.

Decide whether to start, fix issues or investigate further

Make the decision from the evidence
What you foundPossible next step
Agreed cases work and the limits are acceptableConsider a limited first rollout with a named owner and a way to handle issues.
Problems can be reproduced and scopedEstimate fixes and checks before extending use.
Permissions, data or dependencies remain unclearInvestigate those areas before handing the workflow to the team.
Code or authorized access is missingEstablish what materials are available before committing to work.

An app can already be useful with a small, clearly defined scope. The decision depends on the consequences of errors and the controls available. For security certification or broader assurance, use a review appropriate to the risks.

A relevant LuCz project

Matteo took over a partially working management app originally built with AI and brought it into operational use at a company. The case described by LuCz reports more than 20 users and work on development, integrations and releases. See LuCz stabilization and software maintenance. Each application needs its own assessment.

For the longer-term challenge of understanding and maintaining the code, see technical debt in the AI era. The immediate question here is which tests will inform the next rollout decision.

Frequently asked questions

Do I need to know the code to run these checks?

You can describe workflows, roles and expected results without programming. Technical checks on permissions, data and deployment need knowledge of the stack and agreed access.

Do bugs in an AI-built app mean it needs to be rebuilt?

Reproduce the problems and assess the code, data, dependencies and goals. Ask the proposal to identify what can be kept, what needs fixing and what needs replacing.

Have a first version and want to know what to test before the team uses it?

If a check brings up something you cannot assess, start with one workflow. Explain who needs to use it, what should happen and where you got stuck. In the initial conversation, Matteo reviews the situation and whether LuCz is a fit before you agree to any work.

On the booking page, enter your contact details and a short description, then choose a day and time. Any technical assessment, access or subsequent work is agreed separately, with the scope, responsibilities and cost made explicit.

Sources

Sources were consulted on November 2, 2026. Adapt these test cases to the system and the consequences of putting it into use.

Tags

#AI#vibe coding#software maintenance#MVP#code review

Written by

Matteo Lucrezio

Matteo Lucrezio

Startupper | Lead Software Engineer

Book 30 minutes with Matteo

Leave your contact details, then choose a day and time for a free 30-minute initial conversation. Any technical assessment and subsequent work are agreed separately.

How can we help?

Enter your contact details to continue to the calendar.

We use it to follow up on your request and manage the booking. This does not subscribe you to promotional messages.

We save your details immediately to handle your request. They also stay in this browser for 24 hours, so you do not have to type them again if you reload the page.

No suitable time or prefer to write? Email me at matteo.lucrezio@lucz.dev.