Practical guides
AI-built app handoff checklist: seven checks before client delivery
A practical checklist for agencies delivering business software: prove the workflow, test real permissions, handle failure, reopen the export, and agree who maintains the app.
Before you hand an AI-built business app to a client, ask for evidence of one completed job. A polished dashboard may explain the idea. A successful handoff shows that the right person can finish the work, find the result later, and recover when something goes wrong.
This checklist is for agencies, consultants, and founders reviewing a client portal, booking tool, quoting app, or internal workflow. It is a proposed acceptance method you can adapt to your project. It is not a security certification or a claim that a particular AI builder has passed these checks.
Start with a small acceptance contract
Write one sentence that identifies the user, action, and durable result. For a hypothetical booking tool: “An authorized receptionist creates a reservation, the customer can see that reservation after signing in again, and a second customer cannot see it.”
Then name the test accounts, device, environment, and release being reviewed. Use synthetic data. Agree which failures block delivery and which limitations the client can knowingly accept.
Keep a record for each check: expected result, observed result, evidence, remaining work, and the person responsible. Record a failure as a failure. “The agent said it worked” is not an observed result.
1. Complete the whole business workflow
Start as a new ordinary user rather than an administrator with a pre-filled account. Create the important record, edit it, reload the page, sign out, and return in a fresh session. Check the list view and the detail view. If another authorized role needs the record, check that person's view too.
For the booking example, a useful test includes creating a reservation, moving its time, and checking the updated time after a fresh sign-in. A successful confirmation banner alone leaves persistence untested.
Keep: the release identity, synthetic record reference, expected state, and observed state after returning. Screenshots can support the record, but should not expose personal information or credentials.
Delivery question: can the client complete the job without a developer repairing data or changing permissions midway?
2. Test a second person's permissions
Authentication answers who signed in. Authorization decides what that person may do. An application can have a working login screen and still expose another customer's records.
OWASP recommends least privilege, denying access by default, and checking permissions on every request. Apply that guidance to your actual roles and records. Create two ordinary test users and, where relevant, a separate staff user. In an environment you own or are authorized to test, check that the second customer cannot read or edit the first customer's private record, including through its direct URL.
Check both the allowed operation and the denied operation. A hidden button is not evidence that the server rejects the action. Ask a qualified developer to inspect the request-level boundary when the product handles sensitive information.
Keep: the role-to-action expectations and observed allow/deny results. Never put login credentials in the public handoff.
3. Try a failure the customer is likely to encounter
Choose failures connected to the business job. A booking app might receive a duplicate submission or a request for an unavailable slot. A quoting app might receive an invalid quantity. A client portal might lose its connection while saving a change.
Before testing, state what should happen: a clear explanation, no duplicate record, preserved input where appropriate, and a safe way to retry. Then check the stored result after recovery. A visible error message and correct underlying data are separate observations.
Do not generate real charges, send messages to real customers, or disrupt a live service to exercise a failure. Use an authorized test environment and suitable test integrations.
Delivery question: can an ordinary user understand the next step, and can support identify what happened?
4. Make one second change and rerun the earlier checks
A client rarely stops after the first version. Request a small, concrete improvement: a new required field, an additional filter, or a revised approval step. Keep the original acceptance record and repeat the relevant checks after the change.
This is a maintenance rehearsal. It shows whether the current project, its data, and its verification evidence can support another iteration. It does not establish the cost or reliability of every future change.
Keep: what changed, which checks were repeated, their results, and any regression still open. Agree who authorizes future work and how it will be priced before the client needs an urgent fix.
5. Complete the task on the actual device
Ask the person accepting the application to use the device and browser they expect to use at work. Test form labels, keyboard behavior, scrolling, navigation, validation errors, and the final submission.
A desktop layout at a narrow width can reveal overflow. It does not prove a phone's keyboard, touch controls, camera, or file picker works. Record which devices were actually tested and which remain untested.
Delivery question: can the intended user finish the full task on that device, including correcting a mistake?
6. Assign recurring costs and operational responsibilities
List the services the application needs, who owns each account, who pays, and what usage can change the bill. Check current provider terms and the project's actual consumption before making a cost claim. Separate observed charges from estimates.
Agree who handles incidents, renewals, data recovery, customer requests, and future changes. A useful handoff has a named support route and a clear boundary around what the agency will maintain.
If the application needs a custom domain, treat DNS and email as part of the launch review. The domain owner, existing records, HTTPS checks, and rollback plan belong in the brief before anyone changes production traffic.
7. Reopen the exported project in a clean environment
If source portability is part of the sale, verify the actual delivered package. Follow its setup instructions from a clean environment. Check that the expected source files, dependency information, configuration instructions, and known limitations are present.
Record the commands or steps needed to run it and the result you observed. Identify external services that still require client-owned accounts or configuration. Keep private keys and passwords out of the export.
An export button or a ZIP filename does not establish that the package can be run or maintained. If this check has not been completed, label portability as unverified rather than promising a working handoff.
Copy this client acceptance record
Use these fields for each business workflow:
- Workflow and release: the job being accepted and the exact version tested.
- Users and environment: roles, synthetic records, devices, and test location.
- Expected result: what successful completion looks like.
- Observed result: what actually happened, including a fresh-session check.
- Evidence: a private test record or appropriate redacted proof.
- Open work: limitations, severity, owner, and agreed next step.
- Operations: account ownership, recurring costs, support, and rollback responsibilities.
- Acceptance: who reviewed the result and when.
Put the checklist in the project brief
When evaluating Admas or another builder, bring the acceptance contract into the project before implementation. Ask to inspect the files and results, then verify the workflow yourself. Use the AI app builder buyer's guide to frame the wider evaluation.
The handoff is ready when the agreed checks have evidence and the client understands the remaining responsibilities. That is a more useful promise than calling every attractive demo production-ready.
Method and evidence
An original editorial acceptance checklist, source-checked on November 10, 2026 against OWASP authorization guidance. The booking example and pass/fail criteria are proposed examples, not measured customer results. This guide does not certify Admas or any generated application. Admas owns this publication and is one of the products a reader may evaluate.
Put the idea to work
Bring your next useful problem.
Describe a workflow, inspect the result, and keep improving it in Admas.
Start a project