What is AI observability?

When a normal program breaks, you read the logs. When an AI app gives a wrong answer, the logs often say nothing useful, because the "decision" happened inside a model. AI observability fills that gap by recording what the system did at every step, so you can reconstruct it afterward.

What gets recorded

A typical setup captures a trace for each request: a timeline of every step the app took. That usually includes:

• Each call to a model, with its timing and how many tokens it used.
• Each tool call an agent made and what came back.
• Any documents fetched by retrieval-augmented generation.
• Errors, retries and the total cost of the request.

Whether to store the actual prompts and answers is a separate choice. They can hold private data, so many teams make it opt-in.

Why it matters

Think of a flight recorder. You hope never to need it, but when something goes wrong, it is the only way to know what happened and in what order.

For an AI agent that takes many steps on its own, this matters more. One bad step early on can send everything after it in the wrong direction, and without a trace you cannot tell which step it was.

Observability also helps with the quieter problems:

• Cost: you can see which feature or user burns the most tokens.
• Quality: you can spot hallucinations in real traffic, not only in tests.
• Drift: you can notice when answers get worse after a model update or a prompt change.
• Accountability: you can show what an agent did and why, which security and compliance teams often ask for.

How it differs from evals

Evals test a model on a fixed set of examples you chose in advance. Observability watches the real traffic after launch. You want both: evals tell you whether a change is safe to ship, and observability tells you how it behaves once people use it.

Standards and tools

Many teams use OpenTelemetry, an open standard for collecting traces and metrics, which has conventions for describing generative AI calls. Using a shared format means you can switch the tool that displays the data without rewriting your instrumentation. Check the project's documentation for the current status of those conventions, since they have been evolving.

A simple starting point

If you build with AI, begin by logging one record per model call: which model, how long it took, how many tokens, and whether it succeeded. Add tool calls next. That alone answers most "what just happened?" questions.