An operator at a circular station surrounded by floating red data screens, representing the practice of measuring AI systems.
Evaluation is the control room around an AI system: many instruments, many views, one question - is this fit for use?

AI evaluation is the whole practice of judging whether an AI system is fit for use. It is not one test or one tool. It is the umbrella over everything you do to answer three questions: Is the output correct? Is it safe? Does it hold up in the real conditions you plan to deploy it in? Evaluation combines standardized tests, adversarial probing, human review, automated scoring, and ongoing monitoring after launch. No single method covers all three questions, so a serious evaluation stacks several methods together.

A concrete analogy

Think of certifying a new car before it goes on sale. A crash test is one instrument. A fuel-economy test is another. A test driver on real roads is a third. The regulator does not pick one and call the job done. They run the full battery, then keep watching for recalls once the car is on the road. AI evaluation works the same way. A benchmark is the crash test: a fixed, repeatable measurement. Red-teaming is the stress driver trying to break it on purpose. Human review is the expert judgement no automated score captures. Live monitoring is the recall watch after launch.

Evaluation is the parent concept

Beginners often say “benchmark” when they mean “evaluation”. They are not the same. Evaluation is the practice. A benchmark is one tool inside it. The hierarchy looks like this:

The practice
AI evaluation judging whether an AI system is fit for use
Methods
Benchmarks Red-teaming Human review LLM-as-a-judge Live monitoring
Building blocks
Test cases Rubrics and metrics Datasets the individual inputs and scoring rules each method runs on

Each method answers a different slice of the question. A benchmark gives you a fair, repeatable score against a fixed rubric. Red-teaming probes for failures a benchmark never lists, from jailbreaks to unsafe advice. Human review catches the quality that no automated metric captures, like tone, nuance, and factual grounding. LLM-as-a-judge uses one model to score another model’s output at scale. It is fast and cheap, but it can be fooled, so teams use it to supplement human judgement, not replace it, and have experts check a sample to confirm the automated scorer agrees with them. Live monitoring watches the system in production, where real users send inputs no test set anticipated.

Four things you can evaluate

The single most useful distinction in this field is what you are actually measuring. The same word “evaluation” covers four different targets, and confusing them is a common, expensive mistake.

Level 1 Model The raw model on its own. Knowledge, reasoning, and language ability, measured with fixed benchmarks.
→
Level 2 System The model plus its prompts, retrieval, and guardrails, judged end to end as one product.
→
Level 3 Agent A system that plans, calls tools, and acts over many turns. You judge the trajectory, not one answer.
→
Level 4 Workflow Several agents or steps chained into a business process. You judge the whole pipeline and its outcome.
  • Model evaluation : tests the raw model in isolation, usually with standardized benchmarks for knowledge and reasoning.
  • System evaluation : tests the model together with its prompts, retrieval, and guardrails, the thing your users actually touch.
  • Agent evaluation : tests a system that acts over many steps, so you score the plan, the tool calls, and the path, not one reply.
  • Workflow evaluation : tests a chain of steps or agents wired into a real process, judging the end result of the pipeline.

A model can score well on a benchmark and still fail as a system, because your prompt, your data, and your guardrails all change the outcome. This is why a strong model evaluation never guarantees a strong product. You have to evaluate at every level you deploy at.

If your product retrieves documents before answering, you also need RAG evaluation , which measures whether the retrieved context was relevant and whether the answer stayed grounded in it. For a step-by-step walkthrough of running these tests in practice, see the guide on how AI models are evaluated . Use this page as a hub: pick the level you are deploying at, then follow the link to the method that measures it.

Further reading