AI strategy for small companies: the payback period test

Companies ask me what to build with LLMs. My answer is a test, not a roadmap: every AI use case must clear both of these before you spend engineering time.

1. A payback period you can write down. "It will make us innovative" is not a payback period. "Reduce manual quote review from 40 minutes to 8" is. If you can't price the outcome, you can't defend the project when the model bill arrives.
2. A failure mode you can survive. LLMs are wrong at a rate. If the system can be wrong in a way a customer or auditor notices — pricing, health, compliance — you need guardrails, review queues and a rollback plan before the demo. If the cost of wrong is small, ship fast.

What I actually do in engagements: rank the use-case portfolio by payback and implementation risk, kill the demos dressed as strategy, and put the first model in front of real customers with a measured target. Two or three projects a year, owned by someone with a number on their scorecard — not an "AI lab."

The boring truth from running frontier models in production: the model is the cheap part. The expensive part is the workflow around it — evals, data access, review loops. Budget for that or don't budget at all.