Map AI Capabilities to Product Problems
Choose the simplest capability that solves the problem. Generation creates content; classification assigns a label; extraction turns unstructured content into fields; semantic search retrieves conceptually related material; RAG supplies retrieved evidence to a generation step; tool calling lets a model request controlled operations; multimodal models accept or produce content such as images or audio.
Do not reach for an agent when a fixed workflow will work. A deterministic pipeline is easier to test and explain. Use a model for ambiguity or language understanding, and conventional code for arithmetic, permissions, filtering, and operations that require exact guarantees.
FAQ over company docs -> retrieval + grounded answer
Invoice intake -> extraction + schema validation
Support ticket routing -> classification + review threshold
Natural-language reporting -> constrained query plan + read-only tools
Image understanding -> multimodal model + image policy
Refund workflow -> policy engine + approval + audited tool- Use the simplest pattern that meets the requirement.
- Separate language interpretation from exact business logic.
- Define a measurable success criterion for each AI capability.
When is a deterministic workflow generally preferable to an autonomous agent?