Most organizations that track AI spend stop at the team level. That is a good start, and it is not enough to make investment decisions, because budgets are approved by project, not by team. A team of eight might run a dozen projects at once, and knowing the team spent forty thousand dollars last month tells you nothing about which of those twelve initiatives earned it.
Why project-level tracking is harder
Team attribution is straightforward because people belong to teams and rarely move. Projects are messier. One engineer contributes to several projects in a week, a single API key serves multiple initiatives, and a shared service handles requests that originate from different projects entirely. The spend does not arrive labeled with a project name, so the labeling has to happen at the moment the call is made.
The mechanism: tag at the source
Project attribution requires a project identifier attached to each AI call as it happens. The cleanest place to add it is the metadata field most AI APIs support, or a wrapper layer your applications call through. Every request carries a project tag alongside the team tag, and that tag flows into your usage data so spend can be grouped by project without reconstruction after the fact.
If you cannot modify every call, the fallback is separate API keys per project rather than per team. It is coarser, but it produces clean project attribution for the workloads that matter most.
What project tracking reveals
Once spend is grouped by project, three things become visible immediately. Which projects consume the most AI relative to their value, which are quietly scaling in cost as they mature, and which experiments have run their course but never stopped spending. That last category is the most common source of waste, projects that ended in everyone's mind but not in the billing data.
Tie project spend to project outcomes
The point of tracking by project is to compare cost against what the project delivered. A project that spends heavily but ships a revenue-generating feature has different economics than one that spends the same and produces an internal tool nobody adopted. Tracking spend by project is what lets you have that conversation with real numbers instead of impressions.
Project-level attribution is the layer where AI spend stops being an infrastructure cost and starts being an investment decision. Build the tagging in once, and every project's economics become a number you can defend.