Keeping your cost per AI call honest
Model prices fall every quarter, yet bills keep rising. Four habits that keep AI features profitable.
Large language models are extraordinary, and they are also expensive, slow and occasionally wrong. A surprising number of problems that arrive labelled “AI” are better solved with a rule, a search index or a well-designed form.
1. Is the answer already written down?
If the right answer lives in a policy document or a database, you probably need good search, not generation. Retrieval with a small model on top is cheaper and easier to trust.
2. Can you list the rules?
If an expert can write the decision down as ten rules, write the rules. Use a model only for the messy edges the rules cannot reach.
3. What does a wrong answer cost?
Where a mistake is cheap and a human checks the output, LLMs shine. Where a mistake is expensive and nobody checks, you need guardrails, evals and often a human in the loop — or a different approach entirely.
The most valuable thing we do in a strategy engagement is sometimes saying no.