An AI support agent should show its work

The fastest way to lose a customer's trust is an AI agent that answers confidently and wrongly. The second fastest is one that says "I don't know" to questions your docs answer perfectly.

Grounding is a product decision

Cairn's agent only answers from your workspace's knowledge — captured guides, published articles, resolved conversations. Every answer carries its sources, so the customer (and your team) can see exactly where a claim came from.

When retrieval comes up empty, the agent doesn't improvise. It says so, offers to connect a human, and files the gap. Those gaps become a work queue: the questions your customers actually ask, ranked by how often your knowledge fails to answer them.

The escalation contract

• Confidence below threshold → hand off to the inbox with full context
• Policy or billing questions → always a human
• Frustration signals → skip the dance, get a person

An agent that knows its limits deflects more in practice, because customers keep using it.

Credits, not surprises

AI usage is metered in credits you can see — a monthly grant with your plan, top-ups when you want them, and a hard stop instead of a surprise invoice.