An agent starts from your documents, policies, past tickets or calls. Before it talks to a customer we write an evaluation set: real questions with the answers your team would give. The agent must pass it, and we show you the score.
In production every answer the agent is unsure about goes to a review queue for a person. Hand-offs carry the whole conversation. Nothing is trained on your data without your written consent, and your data stays in your own accounts.
The run plan pays for hosting, model usage up to a cap, and a tuning session each month where reviewed conversations become better answers. You see conversations, outcomes and cost on a dashboard.