AI is easy to add and surprisingly hard to make useful. A model can summarize, classify, generate, route, and plan. The product question is not what the model can do. It is where a model can remove a real point of friction without creating a new one.
Start with the job
Describe the workflow without mentioning AI. Who is trying to accomplish what? What information do they need? Where do they hesitate? Which step is repetitive, slow, or dependent on specialist knowledge?
Only then decide where a model belongs. Sometimes it is a drafting layer. Sometimes it is a classifier that routes work. Sometimes it should not be there at all.
Match the pattern to the uncertainty
Anthropic’s practical taxonomy is useful here. Prompt chaining works when a job can be decomposed into known steps. Routing works when different inputs need different specialists. Parallelization helps when independent perspectives can be gathered at once. Agents are appropriate when the path cannot be known in advance and the system can safely inspect its environment.
- Known path: use a workflow with explicit steps and checks.
- Many categories: route the request to a focused path.
- Independent work: parallelize and combine the results.
- Open-ended task: use an agent with clear tools, limits, and stopping conditions.
Make the boundary obvious
Useful AI products make context visible. They tell the user what information was used, what action is about to happen, and where confirmation is required. The product should make correction easy instead of making the user negotiate with a black box.
Evaluate the whole workflow
A good output is not enough. Measure time saved, correction rate, completion rate, escalation quality, and the cost of wrong answers. Evaluate the workflow with representative examples before you optimize the prompt. A more eloquent response that does not improve the job is not progress.