The customer's CFO asks what the AI system will cost per month at full rollout. How do you estimate it?

Strong answer

Build it bottom-up: users x requests per user per day x average input and output tokens per request x model price, plus embedding and re-indexing costs, vector database and hosting, observability, and human review time. Use pilot telemetry for token counts rather than guesses. Give a range with stated assumptions and show the levers: model routing to cheaper models for simple requests, prompt caching, shorter context. Compare against the value metric (hours saved x loaded cost) so the CFO sees cost per outcome, not just a bill. Commit to reporting actuals monthly.

Weak-answer indicators

• Quotes only the model API price
• No ranges or assumptions
• No link to value