What is error analysis, and why is it called the most important eval activity?

30-second answerSay your answer out loud first, then reveal.

Error analysis means systematically reading real outputs, especially failures, and categorising what went wrong to find the most common and most damaging failure modes. You read traces, write open notes about each problem (open coding), group the notes into failure categories (axial coding), count them, and prioritise fixes. It tells you which evals to build and what to fix first. Generic metrics chosen without it often measure the wrong things.

The error analysis loop: sample traces, open coding, axial coding, count and rank, then fix top categories and build targeted evals, re-measure, and sample again.

Example output

Failure category	Count (of 100)	Severity	Action
Didn't ask for order ID before answering	18	Medium	Prompt rule + eval
Retrieved outdated policy	12	High	Index freshness fix
Over-long answers on mobile	15	Low	Length constraint
Hallucinated delivery date	6	High	Tool grounding + faithfulness check

Best practices

One person with domain knowledge (or a small group) reads the traces. Don't outsource understanding.
Look at the full trace: input, retrieved context, tool calls, output.
Repeat regularly (weekly early on), because new failure modes appear with new users and features.
Turn each category into a measurable eval (code check or judge) so you can track it.

PreviousWhat is a golden dataset? How do you build one and how big should it be?

Little by little, you're building something great.

What is error analysis, and why is it called the most important eval activity?

Example output

Best practices

Related