AI agent and copilot
AI that says what it is and knows when to stop
The AI agent answers customers on your digital channels from the sources you publish, discloses that it is an AI first and hands over to a person whenever it is asked or unsure. The copilot drafts for agents; people decide what is sent.
The AI agent
Discloses first
The agent says it is an AI before anything else, and its messages are labelled AI-generated in the console and the API.
Answers from your sources
Answers come from your published knowledge base, with citations. Below your confidence threshold, it hands over instead of guessing.
Hands over on request
"Talk to a human" works at once. The agent leaves a summary as an internal note for the person who takes over.
Tools with guard rails
Only the tools you define, on the channels you allow, with confirmation before any action. Arguments such as the customer's verified phone number are bound by the platform, not the model.
Prompt-injection defences
Untrusted text is kept apart from instructions; answers are checked for prompt leaks, secrets, unapproved links and claims to be human. A red-team suite of injection attempts is part of the automated tests.
Tested before it goes live
Your own test conversations must reach your pass mark before a flow with the AI agent can be published.
The copilot
Help for agents, never a substitute for their judgement.
Suggested replies
Drafts with their sources. Sent as written, a reply is labelled AI-generated; edited first, it is marked AI-assisted.
Summaries and translations
A conversation's summary or a translation as a draft the agent can use or discard.
Budgets, masking and your choice of model
Each account picks its model provider, with failover to a second one, on the platform's keys or its own. An account's own keys are kept as references, never as values in the database, and shown only as configured; typing a key into the console needs the secret vault, which is planned.
- A monthly AI budget: every request is checked against it before it is made.
- Card numbers, IBANs, passwords and keys are masked before a request leaves the platform.
- When the budget or the balance runs out, AI stops and conversations go to people.
Questions about AI
Will customers know they are talking to an AI?
Yes. The disclosure comes first in every conversation, and you set its wording.
Can the AI agent take actions in our systems?
Only through the tools you define, with a schema for each, on the channels you allow, with the customer's confirmation before any action and at most two tool calls a turn.
Does AI score our agents?
Automatic QA can apply your QA form to closed conversations, but it is advisory: a person reviews every score before it counts, and an agent never sees an unreviewed score.