The Problem With "AI"
"AI" describes too much. When a team says they want to "use AI," they might mean a classifier, a retrieval system, a generative model, or an agent that plans across steps.
These are not the same. They fail differently. They require different infrastructure.
A Practical Taxonomy
Predictive — classifiers, regression, anomaly detection. High-reliability, deterministic, evaluable.
Generative — LLMs, diffusion models. High-capability, probabilistic, hard to eval.
Agentic — planning systems, multi-step reasoners. High-complexity, emergent failures, require execution contracts.
How to Choose
Start with the failure mode. If you can't tolerate inconsistency, predictive. If the output is the value (text, image, code), generative. If the problem requires multi-step reasoning with real-world consequences, agentic — and plan for the overhead.