Somewhere in most companies adopting AI, there is a finance person who will ask a simple question this year: what are we spending on this? The honest answer at a lot of places is a shrug. Usage shows up weeks later on an OpenAI or Anthropic invoice, split by nothing useful, and nobody can say which team, feature, or customer owns it.
That shrug is becoming a product category.
The category arrived
Count the comparison pages. CloudZero published a roundup of AI cost management tools. Finout published two: one on AI cost observability, one on FinOps for AI. Amnic published on both AI cost tracking and FinOps for AI. Braintrust, Maxim, Tokonomics and Comet have each published their own list of LLM cost tracking tools. When seven vendors write the same listicle in a year, the category is real regardless of what any single product does.
The products behind the lists fall into three groups. Gateways and proxies: LiteLLM tracks spend per key, user and team across providers, Helicone sits in front of your API calls and logs per-request cost. Observability platforms: Langfuse attaches token counts and cost to every LLM call as a trace, and Datadog sells LLM Observability with cost estimation built into existing APM dashboards. FinOps platforms: Vantage ingests Anthropic and OpenAI billing directly alongside cloud spend, and Mavvrik and Kion both market token-level tracking with chargeback to teams and cost centres.
Per the FinOps Foundation’s own surveys, essentially all FinOps teams are now managing AI spend in some form. The foundation publishes a phased adoption path for it. Finance departments have noticed.
What the tools measure
Look at where each product sits and a pattern shows up. Every one of them meters at the API call: tokens in, tokens out, model, price. That is the part that was always easy. The usage object comes back with every response, and multiplying by a rate table is a weekend of work.
The part none of them answer is the one leadership asks. Which pull request did this spend produce? Which customer’s feature request consumed the most? What did the last shipped increment of work cost, all in? A gateway knows the request came from key sk-proj-.... It has no idea that the request was one of forty an agent made while implementing a refund flow.
This is the same structural problem infrastructure FinOps solved with resource tags: an API call has no resource to tag. It is a transaction. Attribution has to be reconstructed from context around the call, and context around the call lives in agent session logs, git history, and ticket systems, not in the billing stream.
Why leadership will care
Two forces push this from “engineer’s curiosity” to “recurring agenda item”.
The first is that AI is moving from experiment to line item inside every technology function. When a few engineers prototype on a personal key, the bill is noise. When agents write a meaningful share of production code and support workflows run on models, the spend is a budget line that grows every quarter, and finance will want it allocated the way cloud spend is allocated: by team, by product, by customer.
The second is evals. Anyone running agents seriously is already building evaluation harnesses to make behaviour consistent. Cost is just another dimension of the same consistency question. An agent that passes evals but costs four pounds per ticket against a two-pound support saving has failed an economic eval, and no quality benchmark will catch that. “Cost per unit of work” is the metric that connects the eval report to the invoice.
The gap I keep running into
I spend my days in coding agents, and the interesting thing is that the raw data for good attribution already exists on disk. Both coding agents I use write every model response to a local session log with token counts and a computed cost, including for subagent conversations. Nothing about that is secret. What is missing is the join from those logs to the work: the branch, the pull request, the ticket.
That join is what I am building and writing about now, under the name Agent Spend. First step was verifying the premise, that per-message and per-subagent cost can be reconstructed exactly from session logs. It can; I will show the numbers in the next post.
The trackers arriving now will answer “how much did we spend”. The open question, and the more useful one, is “what did we get for it”.