20 January 2026
What finance in Kuala Lumpur asks after an AI feature ships
The questions are predictable. Preparing the answers before the first full invoice arrives is cheaper than reconstructing a launch month from memory.
Once a feature is live, finance usually asks four things: what we expected to spend, what we actually spent, what is a one-off launch cost, and what will repeat. Teams that shipped without writing those four sentences spend the next month reconstructing them from Slack.
Launch months mix evaluation runs, a burst of staff testing, and real users. If those three are not separated on the invoice — and they often are not — someone has to reconstruct them from dates. We ask the product owner for the go-live date and the date internal testing stopped. Those two dates do more work than a cost model built before anyone had traffic.
Recurring spend is the number finance actually needs for the next quarter. It is not the launch invoice divided by thirty. It is the last two quiet weeks, annualised with a note about seasonality if the product has any. A food-delivery chatbot in Ramadan is not a quiet-week story; say so.
One-off costs include the evaluation set that was run twenty times, the embedding rebuild, and the extra region that was tried and dropped. If those stay in the “run rate” column, every later month looks like a saving even when nothing changed.
If you are about to ship, write the four sentences before go-live and keep the invoices for the first ninety days in one folder. That folder is the pack we ask for in a compute cost review. It is also the pack your finance partner will thank you for when the first large bill lands.