Your AI Bill Is Split Across Bedrock, OpenAI, and Azure. Nobody Can Explain It.

Your AI spend is unexplainable because it is split across providers that do not share a billing model: Bedrock bills per token by model, OpenAI bills per token with separate batch and cached rates, Azure OpenAI bills per token or per provisioned throughput unit. Different teams adopted different providers, and no native tool joins them into one view. You fix it by tagging every AI call with team and use-case metadata, landing all usage data in one place, and reporting cost per feature rather than cost per provider.

This is now the mainstream FinOps problem, not an edge case. The FinOps Foundation State of FinOps 2026 report found 98% of FinOps teams manage AI spend, up from 31% two years ago, and the single most requested capability across the entire survey is granular monitoring of AI spend: tokens, LLM requests, and GPU utilization. The demand exists precisely because commercial tooling has not delivered it.

Every LLM call in your codebase should carry two pieces of metadata: which team owns it and which feature it serves. On Bedrock that is cost allocation tags and inference profiles. On OpenAI and Azure it is per-key or per-project separation plus your own request logging. The mechanism matters less than the rule: no untagged AI call ships to production.

Then land everything in one ledger. It does not need to be sophisticated, a daily job pulling usage from each provider API into a warehouse table beats any dashboard you cannot query. When I built this discipline into an AI rollout at a 180-person organisation, the combined visibility surfaced about $124,000 a year in identifiable AI-related savings, most of it invisible while the spend sat in three separate consoles.

Prompt caching is the highest-leverage switch most teams have not flipped: repeated input, system prompts, personas, document preambles, is billed at a fraction of the standard input rate when cached. If your prompts share a large static prefix, you are paying full price for the same tokens thousands of times a day.

Batch endpoints are the second lever: most providers process asynchronous batches at roughly half the interactive rate. Anything that does not need a real-time answer, classification jobs, embeddings, nightly summarisation, belongs on the batch path.

The State of FinOps 2026 report also found 78% of FinOps teams now report to the CTO or CIO, up 18 points from 2023. AI cost is being treated as an engineering discipline, and that is correct: the person who can fix an expensive prompt is an engineer, not an accountant.

Apply the same operating cadence you would to AWS: budgets per team and per feature, anomaly alerts on daily token spend, and a monthly review where each AI feature defends its unit economics. An AI feature whose cost per user grows faster than its value per user is a product decision waiting to be made.

Do not start with a tool evaluation. Start with the ledger: pull last month usage from every provider console you know about, ask each team lead which AI features they run, and reconcile the two lists. The gap between them is your shadow AI spend, and finding it costs nothing but an afternoon.

Then instrument the top three features by spend, flip on prompt caching where the prefix is static, and move the obvious batch workloads off the interactive path. In my experience the first pass at this, on an account that has never had it, pays for the effort within the month.