All use cases

Use case 04Customer support

First responses from your own documents, with a person on the risky calls.

Answers cite the passage they came from, and actions you mark as risky wait for a human.

01How it works

The mechanisms involved.

One agent, several channels
Customers reach you on WhatsApp, email, or WebChat, and the same agent answers on each.Where the docs say so: One agent, several channels
Answers from your help docs
Upload your help docs and the agent answers with citations to the exact passages, or says there is no answer when the documents don't contain one.Where the docs say so: Answers from your help docs
A person on sensitive actions
Raise a tool's risk with a per-tool override and every call to it waits for a person at the default threshold. Critical calls always wait.Where the docs say so: A person on sensitive actions
Activity by channel
Insights counts messages by agent and by channel, and approvals requested, approved, rejected, and timed out.Where the docs say so: Activity by channel

02Limits

What this doesn't cover.

  1. 01

    Insights counts activity. It does not measure resolution rates or customer satisfaction.

    Source for limit 1
  2. 02

    Customer messages are untrusted input. Prompt-injection detection is a filter, not a boundary.

    Source for limit 2

Run it on your own infrastructure.

Docker Compose, a model key, a channel. The Personal tier is free for individual, non-commercial use.