The short version
- Define the system boundary before evaluating risk or performance.
- Measure the full workflow, not only the underlying model.
- Treat monitoring and human escalation as system components.
01
Model versus system
A model maps inputs to outputs. A system combines that model with data pipelines, orchestration, user interfaces, policies, logging, and people who act on or review the result.
02
Core system layers
Most deployed AI systems contain an input layer, context or retrieval layer, model layer, control layer, output interface, and observability layer. Ownership should be explicit at each boundary.
03
How to evaluate an AI system
Evaluate task quality, latency, cost, reliability, safety, accessibility, and user outcomes. Include failure recovery and human review rather than testing only ideal prompts.
Common questions
Questions worth asking
Does every AI system use generative AI?
No. AI systems can use rules, predictive models, optimization, computer vision, speech models, generative models, or combinations of these methods.