Capstone — design your department's monthly token budget

Three ways to see it

1. 1

   The deliverable has two parts. Part one is a Google Sheet titled '[Your dept] AI Cost Model'. Columns: Use case (e.g. monthly policy briefing summarisation), Owner (a named human), Volume per month (count of runs), Avg input tokens, Avg output tokens, Model (Haiku/Sonnet/Opus/4o/4o-mini), Caching applicable (yes/no), Batching applicable (yes/no), Monthly USD, Monthly PKR. Add a totals row at the bottom. Add a sensitivity panel that lets the CFO change one number — exchange rate, model choice, average input length — and see the total move. A live sheet survives every model price change; a static slide does not.

2. 2

   Part two is a one-page memo. Open with the problem in two sentences. List the five use cases with one-line descriptions. Show the total monthly cost from the sheet. Identify the three biggest cost lines and the optimisation lever on each (cache the system prompt, batch the overnight runs, route the easy intent classification to Haiku). State the residual risk: what could make the number wrong? Vendor price hikes, volume growth, accuracy failures forcing a re-run. Close with the procurement ask: a monthly cap, a quarterly review, a named owner. One page. No appendix. Senior officials read what fits on a single page.

3. 3

   Reference scenarios you can adapt. For an FBR audit unit: 5000 returns per month at 2000 input tokens and 400 output, with caching, on Sonnet, runs about USD 70 a month, around PKR 19,500. For an SBP research desk: 200 policy papers a month at 16,000 input and 800 output, batched overnight on Sonnet, about USD 220 a month, around PKR 61,000. For an HBL customer service queue: 30,000 chats a month at 1500 input and 400 output, 70% routed to Haiku with cached system prompt, about USD 85 a month, around PKR 23,800. For an SBP compliance flag review: 10,000 transaction descriptions a month at 300 input and 50 output, all on Haiku batched, about USD 4 a month, around PKR 1,100. For NADRA citizen FAQ in Urdu: 50,000 short questions at 200 input and 100 output mixed Urdu-English, cached system prompt, Sonnet only on escalations, about USD 35 a month, around PKR 9,800. Replace the names with your own.

The why-tree

Try this with Claude

AI-edge prompt: 'Here is my draft memo and cost sheet [paste both]. Critique as a Pakistani CFO would. Find the three numbers you doubt, the one optimisation lever I missed, and the sentence that would make you reject the plan. Then rewrite the memo in 250 words to fix all three.' Treat the AI as a friendly hostile reviewer. That is the cheapest pre-review you will ever get.