The ability to see what an AI use is doing and how well, through logs, traces, measures, and sampled review.
Definition
AI observability is the ability to understand, from the outside, what an AI use is doing and how well it is doing it. It combines logs of prompts, outputs, tool calls, and actions; the model and version in use; measures of cost, speed, and failures; and, most important in operations, sampled human review of output quality. Conventional monitoring tells you that a system is up. Observability tells you what it did and whether it was right.
History
The term comes from control theory and was adopted in software operations for systems that are hard to inspect directly. It has been applied to machine learning and generative AI as tools have become more complex and less predictable.
When to Use
Plan observability before launch for any AI use that matters, and more for agents that act, because after an incident you can only investigate what you recorded.
Step-by-Step
- Decide what to record: inputs, outputs, tool calls, actions, version, and who approved.
- Protect the logs: they may hold sensitive data, so control access and set retention.
- Measure quality by sampling checked output and chart it.
- Track cost, latency, and failures.
- Set alerts for limits, rates, and signals.
- Review on a schedule.
Examples
- An agent log that records each tool call and the approval, used to reconstruct an incident in minutes.
- A weekly p-chart of the sampled error rate.
Common Pitfalls
- Logging everything with no retention rule.
- Only technical measures, no quality sampling.
- Logs nobody reads.
- Logs that expose sensitive data to wide access.