AI 功能示範可行，推出前應測試甚麼？

An AI feature can produce useful results in a demo and still encounter problems when connected to real data, users, and business workflows.

For developers, moving from a prototype to a production application involves more than checking whether a model generates a relevant response. It also means evaluating output quality, verifying data access, handling failures, and deciding how the feature will be monitored.

Here are five practical checks to consider before releasing an AI-powered feature.

1. Define How You Will Evaluate the Output

A response that sounds convincing is not necessarily correct or useful.

Before testing, define what a successful result looks like for your application. For a document-based assistant, this might mean answering a question using relevant source material, avoiding unsupported claims, and acknowledging when the available information is insufficient.

Create a small evaluation dataset containing representative questions, expected evidence, and edge cases. Reuse these test cases when changing prompts, models, or retrieval settings to identify improvements and regressions.

The evaluation criteria should reflect the application's purpose. A tool that summarizes documents may need different checks from an assistant that helps users update business records.

2. Separate Retrieval Problems From Generation Problems

If your application uses retrieval-augmented generation (RAG), an incorrect answer can originate at different stages.

The relevant information might not be indexed, the retrieval system might return unrelated passages, or the model might misinterpret the context it receives.

Consider a hypothetical internal knowledge assistant that gives an incorrect answer about a company's leave policy. Before changing the prompt, a developer could check whether the retrieval step returned the correct policy document. If it did not, the investigation would focus on document indexing or retrieval. If the correct passage was retrieved, the next step would be to examine how the model used that information.

This approach helps narrow down the source of the problem instead of treating every incorrect answer as a prompt issue.

3. Verify Data Access and Permissions

When an AI feature uses private documents or business records, its access rules should match the application's requirements.

Check that authorization is enforced before restricted information enters the model's context. Also consider whether cached responses, retrieved records, or diagnostic logs could expose information to unauthorized users.

Useful checks include:

• Confirming that retrieval respects user and record permissions.
• Testing whether one user's results can appear in another user's session.
• Reviewing how sensitive information is handled in logs.
• Checking that actions affecting business records require appropriate authorization.

The specific controls depend on the application's architecture and risk. The important point is to test access permissions as part of the feature rather than assuming the AI component automatically handles them.

4. Test Failure Scenarios

AI-powered applications depend on components that can fail independently, including model APIs, databases, retrieval services, and external integrations.

Identify the failure scenarios relevant to your application. These may include API timeouts, rate limits, empty search results, or responses that do not match an expected output format.

Depending on the use case, safeguards might include:

• Validating structured responses before using them.
• Applying bounded retries where retrying is appropriate.
• Showing clear error messages when a request cannot be completed.
• Providing a fallback workflow when the AI feature is unavailable.

For features that modify business records or trigger external actions, also consider duplicate requests, interrupted operations, and authorization checks.

Not every application requires the same safeguards. Choose them based on what the feature does and the consequences of an incorrect result.

5. Monitor the Feature After Release

Pre-release testing can identify many issues, but it cannot cover every interaction that may occur in production.

Monitor indicators relevant to the application, such as response latency, failed requests, output-validation errors, and user feedback. For retrieval-based features, it may also be useful to investigate recurring unanswered questions or cases where relevant documents are not found.

Logging can help developers investigate problems, but diagnostic data should be handled with appropriate access controls, redaction, and retention limits.

Use what you learn to prioritize improvements. Depending on the problem, the next change might involve the prompt, retrieval configuration, source documents, integration logic, or application interface.

A Practical Pre-Release Checklist

Before making an AI feature available to a wider audience, consider whether:

• [ ] Representative test cases have been evaluated.
• [ ] Retrieval quality has been checked, if applicable.
• [ ] Data access and permissions have been tested.
• [ ] Relevant failure scenarios have been considered.
• [ ] Monitoring and responsibility for ongoing issues are defined.
• [ ] A suitable way to disable or roll back the feature is available.

This checklist is a starting point, not a guarantee of production readiness. The appropriate release criteria depend on the feature's purpose, complexity, and risk.

Connecting Engineering Checks to Business Goals

Technical testing is only one part of an AI implementation. Teams also need to understand the problem they are solving, the intended outcome, and how they will determine whether the solution is useful.

For a broader perspective on selecting AI use cases, assessing readiness, planning pilots, and measuring outcomes, see this AI implementation roadmap for businesses.

The two perspectives complement each other: business goals help define what the team should evaluate, while engineering tests provide evidence about how well the proposed solution performs against those requirements.

Final Thought

A successful demo is a useful starting point, but it is not sufficient evidence that an AI feature is ready for production.

Evaluating outputs, checking permissions, testing failure scenarios, and monitoring results can help developers make more informed release decisions. The right level of testing depends on the application and the risks involved.

For developers working with AI applications, which area has required the most attention in your projects: output evaluation, retrieval quality, permissions, or integration with existing systems?

Disclosure: This article was prepared with AI assistance and should be reviewed by the author for technical accuracy before publication.