Know where to look when AI breaks

A one-page map of the questions to ask at each layer of your AI stack.

AI systems don't fail the way traditional software does. Outputs are probabilistic, dependencies shift, and the symptom you see is usually a few layers away from what caused it. Latency spikes in the application, but the real problem is a retrieval failure. Cost doubles overnight because someone swapped a model and nobody flagged it. 

This cheat sheet covers seven layers, from business impact down through agents, model operations, and cloud infrastructure. At each one, the questions worth answering first. 

What you'll get answers to: 

• Why an agent took the action it did, and whether your guardrails hel
• Whether a bad response came from the model, the prompt, or the data it retrieved
• What you're trading away when you tune for quality, latency, or cost
• Whether your infrastructure is keeping pace with inference demand