See every use of AI.Govern every call.Spend less.

Installation is a single command pasted from the dashboard. The scanner runs on any host that stays on and can observe egress, whether that is a router, a SPAN port, a small Linux box or a laptop in device mode. It sweeps once within minutes, then settles into continuous sixty second windows that survive reboots.

Detection reads DNS queries and the TLS Server Name Indication field only. Kairos learns that a device contacted a known AI endpoint and how much data moved; it never terminates TLS, inspects a payload or stores a prompt. Volume is modelled into token estimates and priced against published provider rates, which is how a byte count becomes the dollar figure in Figure 1.

Attribution follows IP, MAC and hostname, covering laptops, phones, servers and unattended devices alike. Scanners sharing a LAN reconcile so nothing is counted twice, and any network can be ignored outright.

Data loss prevention runs before the provider, not after. More than 152 categories of personal and sensitive data, including names, national identifiers, medical record numbers, dates of birth, payment details, credentials and API keys, are detected and replaced in the outbound request. What reaches the model is a redacted document; what reaches your audit log is tamper evident proof of exactly which fields were stripped.

Redaction is deliberately surgical. Regulated meaning is preserved while identity is removed, so the model still reasons about a high risk diabetes record and returns clinically useful guidance without ever receiving a patient name. Twenty two frameworks, including the EU AI Act, NIST AI RMF, OWASP LLM Top 10, ISO 42001, HIPAA, GDPR, CMMC Level 2 and 3, FedRAMP aligned controls and named mandates such as the Colorado AI Act, are evaluated per request rather than sampled after the fact, and organizations can add their own.

The intent firewall handles adversarial input. Jailbreaks, prompt injection, roleplay exploits and attempts to extract credentials or personal data are classified by meaning as well as by pattern, then blocked, masked, warned on or logged according to your policy. Repeat offenders can be quarantined automatically by source, category, severity and rolling window, and high stakes requests can be routed to a human reviewer before any model sees them, with reviewer corrections exportable as JSONL training data.

Cost control shares the same pass. Context compression removes at least a quarter of each payload through summarization, pruning and isolation without degrading answer quality, and can be bypassed when a workload demands verbatim context. Routing weighs cost, sending trivial work to inexpensive models and reserving premium capacity for requests that need it. Budgets are enforced rather than announced: when a key, agent or Command Center reaches its cap, the request is refused.

Pricing is a slider rather than a negotiation. Choose the monthly token volume you expect, watch the price update, and change it whenever your usage changes. The effective rate improves automatically as volume grows, so scale is rewarded without anyone asking you to sign a longer contract first. Paying annually takes a further twenty percent off.

Forecasting is what keeps the number predictable. Usage and Cost tracks spend as the period progresses and projects where it will finish, so a busy month is visible on day ten rather than in the invoice. Set a warning threshold, a hard cap, or both, at the level of an API key, an agent or a whole Command Center.

Platform cost is also only one side of the ledger: because compression removes at least a quarter of every payload while routing steers work to the cheapest capable model, the provider invoice underneath usually falls by more than Kairos costs.