You can stop at any phase. We define success criteria with you up front, and the deployment tears down cleanly if it is a no.
Compiled has a fully air-gapped mode that runs offline in an isolated environment — no external calls, no cloud dependency, nothing to attest before you start. Run it on a sample of your own traffic and see real verdicts on your own machine. This deliberately sidesteps the "we can't touch a vendor without SOC 2" blocker for a first look.
What you get: a working recognizer library running locally, real verdicts on your sample data, and a concrete basis for deciding whether to proceed.
Deploy as an Azure Marketplace managed application into your own tenant. Analysis runs on your own Azure OpenAI resource at full quality, so content never leaves your boundary. Scope to one use case — one desk's communications, or one agent fleet. Your analysts calibrate on your benign traffic and confirm or dismiss findings.
What you get: real detection results on your actual traffic, a calibrated threshold tuned to your noise level, and examiner-ready output samples.
An agreed readout against the success criteria defined at the start: detection results on your data, the false-positive rate at your real base rate after calibration, coverage versus your current tooling, and examiner-ready output samples. Success criteria are set with you before we begin.
If the numbers are not there, you get a clean teardown and an honest explanation. If they are, you have everything you need to make a decision.
You retain full control at every step. There is no dependency on Compiled that you cannot remove.
The in-tenant phase has four requirements. None of them are unusual for a scoped technical evaluation.
Email us to begin the air-gapped phase or the in-tenant pilot. We will confirm within one business day with everything you need to get started.