AI spend tracking sounds like a large project. Set up correctly, the core of it takes about a week and then runs on a half-hour a month. The reason most organizations never do it is not difficulty. It is that no one owns the setup. Here is the setup, in five concrete steps.
Step 1: Inventory every source
List every place your organization pays for AI: model provider APIs, AI-enabled SaaS tools, developer tools, and enterprise contracts with AI add-ons. Record vendor, monthly cost, owning team, and whether it touches customer data. This is the one step that reliably surprises people, because AI spend hides in tools nobody registered as AI spend.
Step 2: Restructure API keys by team
Shared API keys make attribution impossible. Issue one key per team or major use case, with a naming convention that encodes the owner. This is an hour of work and it is the single highest-leverage move in the whole setup, because it converts an undifferentiated total into owned line items automatically.
Step 3: Build the key registry
Create a simple registry, a spreadsheet is fine, that maps every active key to its team, its use case, and its expected monthly cost. This registry is what you will reconcile the vendor billing against each month. The discipline of maintaining it matters more than the sophistication of the tool holding it.
Step 4: Set soft and hard limits
For each key, set a soft limit that fires an alert at around 80 percent of expected spend and a hard limit as a backstop at roughly 120 to 150 percent. The soft limit gives you an early warning with runway to respond. The hard limit protects against runaway costs from bugs or unexpected scale. Reserve hard caps that actually cut off service for experimental keys, not production ones.
Step 5: Run the first monthly reconciliation
At month end, pull the vendor usage data, map it against the key registry, allocate costs to teams, and flag anything that does not match. The first reconciliation takes a couple of hours because you are building the baseline. Every one after that takes about thirty minutes, because you are only looking for what changed.
What you have after a week
With those five steps in place, you can answer the questions that matter: what did each team spend, is any team's spend growing faster than its usage, and is there spend on the bill that maps to no owner. That last question surfaces the forgotten prototypes and abandoned jobs that quietly inflate an AI bill. The setup is not complicated. It just needs someone to own it for one week.