THE KEY ANSWER
The quality of an assistant depends on the entire system: sources, retrieval, instructions, the model, and the interface. Define the correct result, verify it on representative tasks, and show the user the sources and limitations of the response.
What does a “good answer” mean in your process?
For a consultant, a good answer may mean correctly indicating the procedure. For a salesperson, current offer terms are what matter. For an analyst, the completeness of the comparison will be important. The general criterion of “sounds reasonable” does not allow for comparing versions of a solution. Document the required facts, prohibited promises, and the procedure for handling missing information.
Prepare examples of a correct, partially correct, and unacceptable response. Ask two people familiar with the process for an independent assessment. If they disagree, clarify the rules before correcting the model. Technology cannot decide for an organization which of two conflicting customer service instructions is binding.
Context and references: Anthropic: Demystifying evals for AI agents
Where does the answer come from, and can it be verified?
An assistant using documents should allow navigation to a specific fragment, not just the file name. Quoting is a starting point for verification, not proof of the truth of the entire statement. Check whether the cited fragment actually supports the provided information and whether the document is still in effect.
Plan for refreshing, removing, and versioning sources. If a procedure is withdrawn, its old copy should not reappear through the index, conversation history, or cache. You need a content owner who confirms currency. Without this, even a carefully designed assistant will efficiently retrieve outdated information.
No answer is part of good service
The system should recognize situations where information is missing to complete a task. Instead of a generic message, it can indicate a missing order number, ambiguous terms, or the need to speak with an expert. Such an ending still helps the user by stating what to do next.
Do not base automatic decisions solely on the confidence declaration generated by the model. It is better to examine verifiable conditions: the presence of a source, field completeness, identifier consistency, and the result of business validation. Some matters should always be escalated to a human due to their consequences, even if the answer sounds very convincing.
What must work after the first presentation?
Any change to the model, instructions, or sources can alter the result. Maintain a set of control cases and record the configuration version when reporting a bug. In addition to response quality, monitor wait time, task cost, and whether the user had to start over. A test result without work context does not show full utility.
Ensure basic interface features: the ability to correct data, interrupt a task, and report a problem. Separate the draft from the approved result. In case of a failure, preserve the user's work content and show an alternative path. The assistant should shorten the path to closing a matter even when automation cannot complete it.
WHERE TO START
Bring this into your project.
- Describe the correct outcome and prohibited behaviors.
- Check the currency and access to sources.
- Design for no-answer scenarios and case handover.
- Test changes to the entire system, not just the prompt.
Choose one thing your process is missing today. It's a useful topic for your first conversation with the team.
QUESTIONS AND ANSWERS
Frequently asked questions.
Does citation eliminate hallucinations?
No. The model may provide a source that does not confirm the entire answer. Verification should include the consistency of the specific claim with the document fragment and the currency of the material.
Is it enough to use a larger model?
A larger model will not automatically fix outdated sources, incorrect permissions, or conflicting rules. First determine where in the process the error occurs, then compare possible fixes.
Sources and context
- Anthropic: Demystifying evals for AI agents ↗
The 2026 material describes combining automatic evaluation with human evaluation. The examples of enterprise assistant handling in the article are original scenarios.
Prepared by the ALGOV team. Current as of September 8, 2026. Examples describe possible scenarios, not results from client projects. How we create our guides.
