# What is AI evaluation?

Source: https://mushbloom.uk/resources/glossary/what-is-evaluation.html

By Mushbloom. Published and updated: 2026-10-12.

## Definition & overview

AI evaluation is the systematic testing of model or agent behaviour against defined requirements. It can examine factual accuracy, retrieval quality, tool use, safety, cost and task completion using representative cases.

## Analogy & mental model

It is a driving test with realistic routes, not one successful trip around the car park.

Keep a repeatable test set covering normal cases, missing evidence and unsafe requests.

## Why it matters

One successful demonstration does not show that a system is dependable. Evaluation exposes recurring errors, missing evidence and unsafe actions before wider deployment. The cases should reflect real user tasks and commercial consequences rather than only easy examples.

## Practical business example

An enquiry assistant is tested on routine requests, spam, missing contact details and demands for unsupported pricing before it is allowed to assist staff.

Illustrative scenario, not a client case study or promised result.

## How it works

1. Define measurable requirements and a representative set of expected outcomes.

2. Run the workflow against normal, ambiguous, failure and adversarial cases.

3. Inspect errors, revise the system and repeat the same tests after changes.

## Core components

- Representative cases

- Explicit success criteria

- Repeatable comparison

Automated judges can be useful but should not replace human review of consequential or disputed cases.

## AI evaluation vs Monitoring

Evaluation tests defined cases before or between releases. Monitoring observes behaviour in use; both are needed because production inputs may differ from the test set.

## Frequently asked questions

What does AI evaluation mean?
AI evaluation is the systematic testing of model or agent behaviour against defined requirements. It can examine factual accuracy, retrieval quality, tool use, safety, cost and task completion using representative cases.

How is it different from Monitoring?
Evaluation tests defined cases before or between releases. Monitoring observes behaviour in use; both are needed because production inputs may differ from the test set.

What is a sensible first step for a small business?
Keep a repeatable test set covering normal cases, missing evidence and unsafe requests. Define measurable requirements and a representative set of expected outcomes.

What limitations should I keep in mind?
Automated judges can be useful but should not replace human review of consequential or disputed cases.

## Sources

OpenAI: Evaluation best practices: https://developers.openai.com/api/docs/guides/evaluation-best-practices

No rankings, AI citations or commercial outcomes are guaranteed.