Many organizations pay for AI in two separate ways at once: seat-based subscriptions for people using the chat product, and consumption-based charges for applications calling the API. When the combined bill looks expensive, the first job is figuring out which side is driving it, because the two require completely different responses.
The two halves of your AI bill
Seat subscriptions are predictable and headcount-driven. You pay a fixed amount per user per month, and the total moves only when you add or remove people. API usage is consumption-driven. It scales with how much your applications call the model, which can vary enormously month to month. A bill that feels expensive is usually dominated by one of these, and they call for different fixes.
If the seat side is expensive
Seat cost is expensive for one of two reasons: too many seats, or underused seats. Pull the usage data per seat if it is available. Seats assigned to people who log in rarely are pure waste, the equivalent of an unused software license. A quarterly review that reclaims inactive seats is the entire fix, and it requires no engineering at all.
The second seat issue is duplication, paying for overlapping tools that serve the same purpose because different teams adopted different products. Consolidating overlapping subscriptions removes cost without removing capability.
If the API side is expensive
API cost is where the real variability lives, and it is driven by the mechanics of how your applications consume tokens: how many calls per task, how many tokens per call, how often calls retry, and which model each call uses. An expensive API bill is a sign to look at those mechanics, not at your headcount.
The diagnostic is the same one that applies to any high API bill: separate call volume from cost per call, find whether the driver is retries, context growth, or model choice, and attribute the spend to the specific feature responsible.
Why the combined bill hides the answer
When both halves land on one invoice as a single total, neither problem is visible. The seat waste hides inside the API growth, or the reverse. Splitting the bill into its two components is the first move, because a seat problem and a consumption problem have almost nothing in common except the invoice they arrive on.
Attribution across both
The durable fix spans both halves. Map every seat to a named user and cost center, and map every API key to a team and use case. Once both are attributed, the expensive bill resolves into owned line items, and each owner can see and manage their own contribution. An expensive AI bill is almost always an unattributed one, where cost has grown faster than anyone's ability to see where it went.