GLOSSARY
Observability
Observability describes how well a system can be understood through measurements, events and traces. For an AI application, useful questions include which step produced an output, which version was involved, and where time or cost accumulated.
Investigate an incorrect proposal precisely
If a draft contains the wrong delivery date, saying that the AI made a mistake offers little guidance for improvement. The team needs to distinguish an outdated source from a misunderstood statement or a value changed during a handover.
Appropriate measurement points can shorten diagnosis and clarify responsibility. They can also reveal repeated model calls or slow interfaces that drive operating effort. Orchestration therefore needs traceable handovers without automatically copying every confidential input into logs. Operational staff need enough evidence to act, while access should remain connected to their role.
Record what answers a defined question
Before implementation, specify which failures must be diagnosable and which data is necessary for that purpose. Permissions and retention limits also apply to diagnostic records. A practical test deliberately introduces an error: can its cause be identified and corrected without unnecessary disclosure of customer information? The amount of collected data alone does not answer that question.