What changes when your AI bill is in rupees

When a Pakistani team calls a foreign model API, three things happen that have nothing to do with the model. The price is set in dollars. The payment needs a card that can settle in dollars. And the amount that finally leaves the company account depends on an exchange rate nobody on the engineering team was watching. Multiply by a product with real traffic and finance ends up forecasting a line item it cannot actually see.

P/9 is priced in rupees end to end, and the consequences of that are worth spelling out, because they change how you plan, not just how you pay.

Quoted, billed and settled in one currency

Every route in the catalogue has a per-million-token price in PKR, for input and for output, visible before you make a call. The organisation holds a prepaid balance in PKR, and each request is debited from it as it is served. What you were quoted, what you were charged and what your balance shows are the same currency, so a cost estimate made in a planning meeting is still true at the end of the month.

A balance that fails closed

A prepaid ledger does something an invoice cannot: it puts a hard ceiling on spend. When the balance runs out, requests are refused rather than served on credit. That sounds severe until you have been on the other side of a runaway job that discovered a new pricing tier at three in the morning. A refused request is a page; a surprise invoice is a quarter.

The same mechanism works one level down. Each API key can carry its own lifetime credit limit. When a key reaches it, the gateway answers 402 for that key alone and the rest of the organisation's balance is untouched. A staging environment, a contractor's integration or an experiment can be given exactly the budget it deserves.

Chargeback is a filter, not a spreadsheet

Because every request is attributed to a key, a project and a model, the question of who spent what is answered on the dashboard rather than reconstructed later. Open a key and its spend for the window is the first number on the panel, beside the models it reached. Teams that bill AI usage back to product lines can name a key per product and read the number off.

What this asks of you

• Name keys after the caller (a service, a product, an environment), not after the person who created them. Usage is reported per key, and a key called "test" tells finance nothing.
• Give every non-production key a credit limit. It costs nothing and it is the difference between a failed experiment and an expensive one.
• Watch remaining credits, which are always in the sidebar, and top up before the balance runs dry rather than after a caller hits a 402.

None of this is exotic. It is what buying anything in your own currency, from a prepaid balance, has always been like. It just has not been what buying inference was like, until now.