Why your AI bill never matches your dashboard

Every team running LLMs in production has two numbers for the same month. One comes from telemetry: count the tokens, multiply by a price sheet, sum. The other comes from the provider: the invoice, the thing that actually gets paid. Ask both what April cost and you get two answers. The interesting question is why.

None of it is mysterious. Prices change mid month, and the sheet your script downloaded in March does not know about the change. Cached reads and batch jobs bill at different rates than the naive tokens-times-price math assumes. Some requests never pass through your tracing at all: a teammate's script, a retried job, a key shared with a vendor tool. Those show up on the invoice and nowhere else. Then there is rounding, currency, and the quiet fact that your dashboard groups by local time while the invoice cuts at UTC month end. Each cause is small. Together they guarantee the two numbers drift.

Engineering trusts the dashboard because it can attribute: this agent, this team, this workflow. Finance trusts the invoice because it is what leaves the bank account. Both are right, and at month close they meet. The invoice always wins that meeting, because it is the number that gets paid. But all the explanations live on the other side. When the two do not tie out, someone spends an afternoon in a spreadsheet trying to make them agree, and next month the same afternoon happens again.

Reconciliation is an old accounting habit applied to a new bill. Put measured and billed at the same grain: date, provider, model. Subtract. The delta is the whole product. A delta near zero tells you your telemetry can be trusted, which is worth more than it sounds. A persistent delta is information: spend nobody attributed, a stale price, a trace that never landed. Either way, every dollar on the invoice traces back to dated events instead of an estimate.

You do not need a product to see your own gap. Pick one month and one provider. Export usage from your telemetry, download the invoice CSV, group both by day and model, subtract. Most of the work is pricing history and discounts, and that is also the part that hurts to maintain by hand. That maintenance is what Kenda automates. If you would rather see your own delta than read about one, book a call and we will reconcile one invoice with you.