Why Your Cloud FinOps Tool Can't See Your AI Spend

Most FinOps teams spent the last five years building a solid practice around cloud spend. Tagging policies, cost allocation, showback reports, commitment coverage. Then generative AI arrived, and a fast-growing share of the technology budget started flowing through channels that practice was never built to see.

If your monthly close now includes an "AI" line that nobody can fully explain, you are not alone. The problem is structural, and it starts with where the data lives.

1. The invoice line hides the consumer. When a team calls a model through AWS Bedrock or Azure AI Foundry, your cloud cost tool sees a service charge. It does not see which application, team, or person generated the tokens. Tagging helps with infrastructure, but a single shared API key used by five teams produces one undifferentiated cost stream.

2. Direct vendor spend sits outside the cloud bill entirely. Anthropic, OpenAI, Google AI Studio, Cursor, GitHub Copilot, Microsoft 365 Copilot. Each bills separately, on its own cycle, with its own console and its own definition of usage. None of that reaches your cloud FinOps platform.

3. Seats and tokens are priced on different models. A coding assistant seat is a fixed license. API usage is metered per input and output token, with different rates per model, discounts for cached tokens and batch jobs, and costs that rise with context length. Putting both in one spend basis takes deliberate work.

Why this matters for the FinOps lifecycle

The FinOps Framework moves through three phases: Inform, Optimize, Operate. AI spend breaks the first one. Without allocation down to the team and the user, you cannot run showback. Without showback, there is no accountability. Without accountability, optimization becomes a central team chasing engineers for answers after the invoice lands.

AI spend also behaves differently from compute. Anyone with an API key or a seat can commit spend, with no provisioning step for finance to approve. Model choice, retries, and context length can make the same feature cost many times more. And an agent stuck in a loop does not stop on its own.

What AI spend visibility actually requires

A working FinOps for AI practice needs a few capabilities that cloud tooling does not provide:

One spend basis across vendors. Metered API usage, seat licenses, and fixed-price contracts counted together.
Attribution to people and teams. Every token and every seat mapped to a department, team, user, or agent, synced from your identity directory.
Model-level granularity. Cost, tokens, and requests broken down by vendor and model, so you can see which workloads drive the trend.
No change to the request path. Visibility should come from vendor usage and billing data, without routing production AI traffic through a proxy.

How KeepAir closes the gap

KeepAir is a free AI spend governance platform built for exactly this problem. You connect each vendor with one API key, with no code changes. KeepAir backfills historical usage, imports your org structure from Microsoft Entra or your HR system, and maps every token and seat to the person and team behind it.

From there, FinOps gets showback by department, team, user, and agent, with usage, seats, and contracts in one view. KeepAir covers 15+ vendors, including Anthropic, OpenAI, Google, Microsoft 365 Copilot, GitHub Copilot, Cursor, AWS Bedrock, and Azure AI Foundry. It reads usage and billing metadata only and does not collect prompt or response content.

The takeaway

Your cloud FinOps practice is not broken. It was simply designed for a cost model where provisioning came before spend. AI reverses that order. Bringing AI into the practice you already run starts with visibility at the level where spend is actually committed: the team and the person.

See where your AI budget goes.