What we count as an AI resolution, and what we don't
We bill for conversations the agent closed on its own. Here's exactly where that line sits.
Per-resolution pricing is common now: Featurebase charges $0.49 and Intercom $0.99 each. We charge $0.29 past a monthly allowance. A price per resolution is only as fair as the definition behind it, so here's ours, exactly as the code checks it.
It counts when all of these are true
- The agent answered the customer, with citations to your help center.
- Nobody on your team replied in that conversation.
- The agent didn't hand the conversation to a person.
- The conversation was closed.
It counts once. Reopen and close it again and nothing changes.
It doesn't count when
- The agent handed off. Low confidence, a topic on your block list, a question the help center doesn't cover, or the agent deciding a person should take it. The customer gets your handoff message and your team gets the thread with what the agent found.
- A teammate replied. Even one line. If a person had to step in, it's not the agent's resolution.
- It was a Copilot draft. Drafts, summaries, translations and suggested tags your team uses inside the inbox are never billed per use.
- It was a help-center search answer. Those are part of the plan.
- You run on your own key. Add your Anthropic key in Settings → AI and nothing is metered. You pay Anthropic their price and we add nothing.
Why "closed" and not "answered"
We could count every answer. That's more revenue, and it's wrong: an answer the customer came back to argue with didn't resolve anything. Waiting for the conversation to close means we sometimes miss one — a customer who got their answer and wandered off without anyone closing the thread — and that's the side we'd rather be wrong on.
How the agent earns the right to answer
The agent only answers from passages it can cite: your articles, and past conversations and posts you've allowed. Every claim has to point at a passage the customer is allowed to see. Below your confidence threshold, or with no citation, it hands off instead of guessing. A red-team suite of 50 hostile prompts runs on every change and has to produce zero uncited answers and zero leaks of private content.
What you see
Settings → Plan & billing shows resolutions this month against your allowance, and every agent run — answered, handed off or blocked — is in the review queue with its sources, so you can check the agent's work the same way you'd check a new teammate's.