When not to use an LLM

Half of what we scope does not need a model. Knowing which half is the job.

There is a test we run before any agent gets built. If the questions are fixed, the answers are structured, and nothing about the next question depends on the last answer, you do not want an agent. You want a form.

We learned this scoping an interview product. The instinct was to let an agent run everything, including the parts that were a fixed list of scale questions. On paper it sounds richer. In practice an agent walking someone through fixed items is worse on every axis: slower for the respondent, noisier as data, and it costs tokens to do what a radio button does.

Where the model earns its keep

A conversation earns its cost exactly where a form fails: when the interesting answer is the one you did not know to ask for. A follow-up that depends on what the person just said. A request for the document they mentioned in passing. Digging under a vague answer instead of accepting it. That is real agent work, and nothing else does it.

The boring rule

So the rule we ship by: forms where the questions are fixed, agents only where open-ended probing earns its cost, and ordinary code for everything in between. The model is the expensive specialist on the team. You do not send the specialist to do data entry.

This is also the cheapest credibility move an AI studio can make: telling a client where AI does not belong in their product. The pitch that includes a no is the one that gets believed.