AI spend is going parabolic, and the labels on the bill (OpenAI, Anthropic, Gemini) are about all a CXO gets to work with. The hard part of tying that spend to outcomes is structural.
A major portion of AI spend isn’t COGS. It’s the spend on coding agents producing the software, the spend on building marketing content, the spend on custom sales tooling, the spend on Intercom agents and Sybill analysis. All are structurally shared, all are variable, and none has a bill that maps them to outcomes.
The rest is harder still. Billing data can tell you what a developer spent addressing a single Jira ticket, or what a single support resolution cost. But unit-cost views of outcomes are a different kind of problem, particularly when both Claude Code and Codex delivered the new feature, when ServiceNow and Intercom were used together to fix the customer issue, when Claude Design and Gemini were both used to build the new marketing plan.
The bills we get from Anthropic don’t have the right details for us to attribute spend to business outcomes. This isn’t an Anthropic problem, it’s a visibility problem; regardless, I need to know the ROI of the outcomes we produce, and whether more AI spend would produce better ones.
No single inference provider can give me the data we need to do this. Unit economics are structurally a blended picture, and attributing AI activity across outcomes requires more than activity metrics. The good news: CloudZero’s allocation engine is purpose-built to solve this kind of attribution problem, and, when combined with real-time spend data, connects all shared costs to business outcomes.
Cost-to-serve isn’t the problem. Cost-to-produce is.
For cost-to-serve, CloudZero has decades of practice.
Take three customers using a shared resource like AWS Redshift: Cloud billing data plus tagging (or inferred tags) tells us how much each customer used and how to roll up the cost of features and products. The same approach works for AI infrastructure. Coarse billing makes it harder, but cost-per-customer and cost-per-feature for an AI inference workload is a solved problem.
The hard problem is cost-to-produce. How do I attribute the cost of a subagent running via API key to the marketing campaign asset it was pressure-testing? That context isn’t in a bill (structurally, it can’t be.) Observability telemetry doesn’t carry the business meaning of the work either.
The only way to solve this: Capture the usage data at the moment the agent is working, then run that data through an allocation system robust enough to attribute the spend to the marketing department and the specific campaign. Anthropic billing data will get better over time (CloudZero was the first to support their new Enterprise billing API). But better bills aren’t enough on their own. You need all the spend sources and all the telemetry, threaded through an allocation engine that can connect them to outcomes. That’s just the reality of unit economics. If you want your unit cost, or to understand what it cost to deliver that quarterly OKR, you have to look elsewhere.
So, what do you do?
You bypass the bill and build the spend picture from telemetry. Then you reconcile to the bill when it arrives (and in the rare cases without a bill, telemetry stands on its own). Finance sees accurate total cost, and teams get per-outcome attribution of cost.
Don’t rely on the bill
Using AI yields span-by-span telemetry (what tokens, by whom, on what project), captured the moment work happens. That’s the raw material. But raw telemetry isn’t enough on its own: It tells you the activity, not the business meaning behind it.
What makes this work is the streaming data pipeline underneath: It takes in real-time cost and usage data along with the business context that produced it, and handles the scale of AI telemetry. That data comes from agents we built to capture it where it wasn’t already available, from LLM gateways that already capture it, from SDK wrappers, network capture, and observability-platform integrations we’re adding. These collection sources bring data through the pipeline and into the multi-dimensional allocation engine, where spend is attributed to every relevant dimension and to the outcomes that move the P&L.
That allocation engine is the moat. You can’t do it in Excel. You can’t do it with Tableau, or with an observability platform, or with tools built for cloud cost. We’ve spent the past ten years building this engine: programmatically ingesting cost and usage data, organizing it with fine-tuned allocation logic, passing the most rigorous financial audits. That’s what makes us the platform of choice for the world’s most sophisticated cloud spenders (Coinbase, Nubank, Shutterstock, Duolingo, and many others). It’s what makes us uniquely equipped to do the same for AI costs. All we needed was the data.
With real-time, trustworthy allocation data, I can then get to the crown jewel of financial analysis: unit economics. I know how much I’m spending on every core dimension of my business. I know how much volume my business is doing (for CloudZero, how many dollars of spend I’m managing; for Samsara, how many devices they’re managing; for GrubHub, how many orders they’re processing). I can unite this data and get dimensional unit economics: cost-per-anything (per $ under management, per device, per order, etc.), broken down by customer, by team, by feature, by product, or any other way I might want to analyze my costs and revenue.
The allocation engine
I’m not describing a theoretical future. I’m describing a tangible present. CloudZero’s AI outcome attribution provides all this for AI costs right now. I see it firsthand every day. In February, we adopted Claude Enterprise and we lost all the cost visibility we’d had with them before. Not only were we (and everyone else using Claude Enterprise) limited to their bills for AI cost visibility, but the billing API was nascent and not useful for tracking individual costs. I had no idea who on my team was spending what, let alone why they were spending it.
But using CloudZero’s allocation engine on real-time spend, I have a clear view into who’s spending what, what projects they’re spending it on, and what business outcomes those projects yield.
CloudZero made a bet that our engineers, building agentically with Claude Code and Cursor, would ship faster and at higher quality. I can see whether the bet’s paying off: which engineers ship more features per week, what those features cost to deliver, and where the next investment would pay off. For every AI dollar we spend, I know what it produced. The lump-sum Anthropic bill doesn’t tell me any of that.
The new part is the pipeline that captures real-time telemetry. The hard part is the allocation engine downstream of it. We’ve been the only ones running telemetry-anchored allocation at that scale. Now we’re the only ones running it for AI.AI outcome attribution is one of three major launches we’re announcing today. The other two: AI Hub and our updated UI. Ready to see CloudZero in action? Request a demo here.
