Finance asks where the AI spend went, and the provider console reports a total.
A total is the beginning of what finance needs. They also want to know which team spent it and whether the spending worked. An engineering leader at a global industrial manufacturer described the shortfall without much patience for it. The provider's native reporting, in their words, gives "raw data" and not much else. The reporting value they wanted sat well beyond what the provider offers natively.
Raw data is an accurate description, since the provider console is built to explain the invoice and reporting past that point is left to the customer.
The reasonable version of the objection
We hear the pushback constantly and it's usually correct. A reseller partner said it to our faces directly, pointing out that the major providers already ship native AI spend tracking and that the reason to add another tool wasn't obvious from where they sat.
That partner is right to press on it, and the answer depends entirely on where a team already is. A product leader at a major cloud cost management vendor put the dividing line in concrete terms. At roughly 5 million a year in cloud spend on a single provider, they said, "use native tools, it's good enough." Once a second provider arrives, the environment takes on the same complexity that FinOps teams already know.
A single provider at moderate spend is the case where the native console is genuinely enough and there's no reason to buy anything. The rest of this post covers the five things AI cost reporting has to do, which is how you find out whether that still describes your environment.
Some teams believe a gateway covers the rest. A cloud and AI cost engineering lead at a large enterprise security vendor told us the observability problem had largely been handled inside the enterprise by AI gateways. That's partially true and worth taking seriously. A peer at another security vendor working on the same problem saw where that stops, noting that their open-source LLM gateway gave them cost controls for the model providers while missing all of the ancillary spend around them.
Five things AI cost reporting has to do
A console that delivers all five for a given environment is doing the job.
Every provider visible in one place. Multi-provider visibility is the first thing to break, and the break is easy to miss because each console is individually fine. Once a second model provider, a hyperscaler's hosted model service, and the coding tools all bill separately, reconciliation happens by hand every month. The spend threshold quoted above is a useful marker.
Per-user visibility. Provider consoles report at the account or key level. If two hundred developers share an organization, the console shows one number. Acting on that number, whether that means finding the heaviest user or spotting someone routing simple work to an expensive model, needs a layer the console doesn't have.
Attribution below the service level, down to individual tools and workflows. Provider consoles hit their granularity ceiling at this level. The founder and managing partner of an AI consultancy went looking for exactly this and hit the limit. Their cloud console could report what transcription cost and what summarization cost, "But not at the LLM level." Service-level attribution, which is what cloud billing was built to do, stops one level above where AI cost decisions actually get made.
Cost tied to an outcome. No provider console knows whether the money produced anything, because it has no view of the business result. Boards and CFOs ask for exactly that, and the console is structurally unable to supply it.
Chargeback the department will accept. Allocating spend to a department only counts as chargeback when the department accepts the number. Without that confidence, the allocation gets disputed every time it's presented.
Signals you've outgrown it
A second provider gets added. This is the cleanest trigger and the easiest one to date precisely, though nothing degrades on the day it happens. The reconciliation work starts accumulating from that point.
Alerts arrive after the decision. An SVP of innovation at a technology solutions provider described the timing problem. "FinOps is like historical. By the time you get that alert, it's too late." Provider consoles are billing systems, which makes them retrospective by design. A response that consistently begins after the spend is committed arrives too late to change the number.
A request comes in that takes a spreadsheet to answer. The specific request varies, but answering it required exporting from two places and joining them by hand, and the same request will come back next quarter.
What the next layer up changes
A customer gave us the clearest explanation of why the consoles stay shallow. A VP of Engineering at a large cybersecurity company worked through the incentive out loud. "The AI companies are not going to solve that problem. Because it goes against their business purpose." Solving it would fall to someone outside the providers, they said, and "Maybe it's you guys."
We wouldn't put it more strongly than that customer did. A provider has no reason to build the tool that recommends using less of their product, and a FinOps lead at a major cloud provider asked us directly what we add beyond their own native dashboards.
The next layer up changes the unit of analysis. Provider consoles are organized around the bill, so the finest thing they can describe is a charge. The layer above is organized around the request, so the same spend can be sliced by user, team, workflow, tool call, and outcome, because those attributes travel with the request instead of the invoice. All five of these follow from that difference.
Where we fit
For teams missing two or more of the five, the fastest way to see what's missing is to look at their own workloads instead of a demo environment.
Connecting Providers is the multi-provider view in one place. AI Assistants covers per-user detail across the major AI coding assistants. Attribute Spend to Teams & Departments is the chargeback question. AI Outcomes is how cost gets tied to a business result rather than to a charge. Debug Logs & Traces is the per-request granularity underneath all of it.
Teams getting all five today should stay on the console. We'd rather say so now than three months into an evaluation.
