Why AI spend is unattributable by default

The invoice arrives, the number is bigger than last month, and nobody can say which team caused it. This is the default state of enterprise AI spend, and it is not an accident of sloppy teams. It is structural.

How it happens

• One shared API key per provider, created the week AI adoption started.
• Keys copied between teams because requesting new ones was slower.
• No metadata on requests: the provider sees one anonymous stream.
• Multiple providers, each with its own console and its own invoice.
• Self-hosted models metered by nobody at all.

Why spreadsheets fail

The administrative fix, asking teams to log their usage, decays within weeks. Engineers route around process that slows them down, and every new application quietly inherits an old key. Attribution built on goodwill loses to attribution built into infrastructure, every time.

The structural fix

• All model traffic through a gateway, cloud and self-hosted alike.
• Identity attached to every request: team, application, user.
• Budgets and limits enforced at the gateway, not at invoice review.
• Telemetry exported into the observability platform you already run.

What good looks like

Every request attributable in one place. Finance gets chargeback it can defend. Teams get budgets they control. And the conversation shifts from who spent this to whether it was worth it, which is the conversation an AI program actually needs.

Goverent implements cost attribution and budgeting on the gateway you already run. [email protected]