Evaluating AI Systems

AI systems fail in ways you’ve never seen before. A model scores high on benchmarks and still misleads users. An agent completes a task by taking harmful actions and then covering its tracks. A RAG pipeline retrieves the right documents and still returns a fabricated response. Evaluating AI Systems teaches you how to know—and prove—whether your LLM-based applications are ready to ship. In this practical and important book, author Panos Alexopoulos provides a data and decision-driven approach that starts with the questions you need to answer about your system and works backwards to the evidence you need to gather.

Evaluating AI Systems goes beyond how to gather metrics and use eval libraries. In it, you’ll learn to think systematically about evaluating the performance of AI-led applications and agents. You’ll start with the core issues: how to define your evaluation goals and determine which aspects of a system need to be evaluated to achieve them. Next, you’ll learn how to select meaningful metrics and acquire the appropriate evaluation data, along with the trustworthy judgments you need to calculate them.

Because evals without actions are meaningless, you’ll learn how to interpret evaluation results, communicate their meaning and limitations to different stakeholders, and turn evaluation evidence into decisions and improvements. You’ll see how evaluation continues after deployment, allowing you to monitor how your AI system behaves in the real world and identify when its behavior or performance requires attention.

Along the way, you’ll test the book’s systematic approach against some of the most challenging qualities of modern AI systems, such as evaluating robustness when systems encounter unexpected conditions, assessing whether their behavior is fair across relevant groups, and identifying security and privacy risks. Plus, explore the challenges of evaluating increasingly general and adaptable AI systems whose capabilities and behaviors may be difficult to anticipate in advance.

Throughout the book, examples and case studies spanning predictive, generative, and agentic AI demonstrate how the same evaluation principles can be adapted to different systems and technologies. By the end, you’ll have the principles, frameworks, and practical judgment to build evaluations that produce evidence you can act on.