The agent is working for the business.
Last 30 days. Volume, speed, cost, and capacity returned to the team. Customer-satisfaction numbers below are proxies — derived from our own data, not customer survey responses.
Per-draft cost assumes Haiku 4.5 (classifier) + Sonnet 4.6 (drafter) model-family pricing. Cache-read tokens counted as regular input for demo-grade accuracy.
The minutes-per-reply assumption is adjustable on the endpoint. A conservative 4 min / draft means 0.0h/day of reviewer capacity on average.
% of inbound with angry/rude tone. Lower = customers arrive calmer.
% of reviewed drafts the reviewer flagged as good. Rising = agent is learning the right voice.
% of threads where customer wrote back after our reply. High = first-reply didn't resolve.
Roadmap
Real CSAT / NPS requires post-send survey capture. Planned for V2 — one-click inline rating in the reply itself.
- refund_request1 · 33%
- sales1 · 33%
- account_access1 · 33%