Question 1

How does this actually save money?

Accepted Answer

Two ways. It makes the spend visible and pinned to a team, so the obvious waste (a premium model on trivial requests, a GPU left running overnight) stops hiding in one big invoice. And it flags cheaper routes for the calls where the extra cost buys you nothing, with the evidence to check before you switch. You apply those through the gateway you already run. Nothing to rip out.

Question 2

Do you sit in front of our production traffic?

Accepted Answer

No, not today. The recommendations apply through your own gateway; we ship config for LiteLLM and Portkey. Inline routing shipped in August 2026, and it is off by default on every policy: a policy cannot arm until it has accumulated shadow proof on your own traffic. It is fail-open by design: if anything is unclear, the request goes to the provider you originally picked.

Question 3

Which providers and clouds?

Accepted Answer

OpenAI, Anthropic, Bedrock, Azure OpenAI and Vertex for managed inference, plus your own GPUs through an agent. We reconcile each provider's real bill against what we saw, so multi-cloud doesn't mean flying blind on one of them.

Question 4

What do you need access to, and what happens to our data?

Accepted Answer

For the replay, read-only billing access and an IAM role. No production access, no code changes, nothing in the request path. Tenants are isolated at the database level with Postgres row-level security, not just in app code.

Question 5

SOC 2? SSO?

Accepted Answer

No SOC 2 today. We have not engaged an auditor and we are not going to name a quarter we might miss. You can hold us to that rather than to a roadmap slide. SSO is live now: Okta, Azure AD, any SAML 2.0. There is a public trust portal documenting the controls we run today, and we will complete your security questionnaire as part of onboarding.

Question 6

How do I prove the savings to finance?

Accepted Answer

Every change we recommend and you accept is logged against a control cohort of traffic we deliberately leave untouched, so the before-and-after isn't just calendar drift. Thirty days in, we re-read the same slice of spend and record the real delta against that control, which can come out negative if adoption slipped. The methodology is public, down to the request IDs we used, so your finance team can recompute the number themselves.

Question 7

Can we set guardrails, not just watch dashboards?

Accepted Answer

Yes. Write a short YAML policy: a trigger like "GPU under 5% for two hours", a scope and an action. Preview what it would have done over the last week against real numbers before anything goes live. Nothing touches your cloud without a human approving it in the queue.

Question 8

How much work is onboarding?

Accepted Answer

The replay is two weeks and read-only, and it ends with a written report. Your first snapshot shows up within 48 hours of connecting.

Flat fee. Prove it first.

Pricing that's on your side

Common questions

See what the last 30 days would have saved.