GraySmith.io

Questions we get on the first call.

How much do companies usually save?
It depends on where the waste is. Companies whose spend is mostly production inference typically see 30 to 50% from routing and caching. Companies whose spend is spread across teams, seats and agents often find a further 10 to 25% in the first scan alone, spend with no owner. Multi-provider deployments save the most.
What do you actually need access to?
Read-only access to provider billing and usage APIs, plus whichever of the following you're willing to share: expense and card feeds, SSO logs, egress data. The more sources, the more complete the inventory. The scan works with provider access alone.
Do you store our prompts?
No. By default GraySmith stores metadata and cost data only: model, token counts, latency, price, and the routing decision. Content-level storage is opt-in and only for teams that want quality scoring against stored samples.
Where does our data live?
EU by default, with US and other regions available. Data residency can also be enforced per request through policy routing.
Will routing hurt quality?
That's the thing we test hardest. Every route runs a minimum seven-day A/B against your incumbent model with quality scored continuously. If it regresses past your threshold, it reverts automatically. We report the levers that don't work for you as well as the ones that do.
How long until we see something?
The scan returns findings within a day. Attribution and dashboards are live inside a week. Router savings show up in weeks three to four, on the invoice at the end of that cycle.
We already use a cloud FinOps tool. Why this too?
Cloud FinOps tools are built for infrastructure: compute, storage, Kubernetes, GPUs. They're good at that and poor at API-based AI spend, unprovisioned tools and per-token attribution. GraySmith sits beside them, not instead of them.
We're small. Is this for us?
Yes. There's no spend minimum. Small companies tend to have proportionally more grey-zone spend, because nobody has the job of watching it.
Are you replacing our AI tools?
No. We don't care which models or tools you use. We care that each one has an owner, a budget and a reason. The router chooses between the providers you've already approved.

Still not sure? Run the scan.

Free, read-only, one hour. If we find nothing, you've lost an hour.