The invoice tells you that a call cost four dollars and twelve cents. It cannot tell you whose budget it came out of, whether the prompt carried a patient record, or whether the policy said that model was allowed to see it. Raidu meters the call and decides the call at the same instant, and signs the result.
Cloud cost tooling assumes capacity you provision in advance. AI spend is consumption a model, or an agent, decides at runtime. Tokens have no reserved-instance ceiling, a retry storm looks like ordinary traffic, and a planning loop spends until something stops it.
Every product in this category is one of two things, and the difference is not a feature list. It is where the software sits relative to the call.
Reconciles usage after the provider has issued it, allocates it to an owner, and ranks what looks wasteful. Broad estate coverage, including seats and infrastructure that never touch a gateway. The ceiling is structural: billing metadata carries no prompt, no policy and no decision, so the record can price a call but cannot describe it.
The cost is computed at the moment of the call, from the same event that carries the identity, the policy, the redactions and the verdict. Spend can be refused before it happens. And because the cost row and the governance row are one row, a question about money and a question about compliance are the same query.
Every governed call runs the same sequence, in the same order, inside your own cloud. Nothing in it leaves your environment.
Your application points at Raidu instead of the provider. No rewrite, no SDK swap, no change to how the prompt is built. The endpoint changes and nothing bypasses it.
Identity, team, project, agent, client application and remaining budget are resolved for this specific call. Budget enforcement runs here, before the provider is contacted, so an overrun is refused at zero cost rather than reported next month.
The prompt is evaluated against your policy. Sixty-plus entity types detected at 99.2% accuracy, data-residency rules applied, permitted models checked for this data class, prompt-injection screened.
The call is allowed, redacted, escalated for human review, or blocked. Token cost is computed at that same instant, from that same event, and written with the verdict attached.
One record carries the cost and the governance decision together, signed with RSA-4096, appended to a SHA-256 hash chain, timestamped under RFC 3161, written to WORM storage with ten-year retention.
Discover, attribute, optimise and control is the industry loop, and Raidu runs all four. It adds the two a read-only tool cannot reach: enforce, at the moment of the call, and prove, in a record that survives an audit.
Of forty-two thousand dollars that reached a provider this month, fourteen thousand carried personal data that was redacted before the model saw it. A further seven thousand eight hundred never reached a provider at all: four thousand six hundred refused at the budget, three thousand two hundred held for a human to approve.
Every clause in that sentence comes from one table. A FinOps platform can answer the first number. It has no column for the rest, because the prompt never reached it.
Finance and compliance stop reconciling two systems and start reading the same row.
One integration surface for both jobs. The same interception that redacts a prompt is the one that prices it, so coverage for governance and coverage for cost are never two different numbers.
See the full list on the integrations page.
Compared by architecture rather than by vendor, because within each architecture the answers are the same. For named comparisons, see the comparison library.
Assessed from public product documentation, September 2026.
Much of this category prices as a percentage of the AI spend it manages, which means the vendor's revenue grows with the problem the customer bought it to shrink. Raidu is a platform licence. Bring your own provider keys, keep your own contracts, pay nothing on the tokens.
