The problem
Shared inboxes (bookings, service, billing) fill up with the same dozen questions: hours, availability, balances, rescheduling, a job update. When it gets busy, replies slip a day or two, and customers notice.
How it works
- An automation watches the inbox and skips anything it has already processed, using the message ID as a de-duplication key.
- The message is sent to the agent which returns structured output: intent, key details, a draft reply, a confidence score and an escalate yes/no.
- The agent looks up facts through shared tools like open balances from accounting, bookings from your scheduler, and hours or pricing from a sheet your staff maintain.
- Routing rules decide what happens next so escalations ping a person, low-confidence drafts land in a review folder, and the rest sit as ready-to-send drafts in the thread.
What it needs from you
- A list of your most common email types and how you normally answer them
- Read access to the inbox and the systems that hold the answers
- One named owner who reviews drafts during the pilot
- 50–150 past emails we can label to build a test set
Why it's built this way
The lookup tools are built once and reused. Adding a second inbox, SMS or voicemail later is mostly configuration, not a rebuild.
Prompt or model changes are re-run against the labeled test set before release. If accuracy drops, the change waits.
How it stays under control
- Draft-only at launch; a person sends every reply
- Confidence threshold routes uncertain drafts to review
- Escalate flag sends sensitive topics (billing disputes, complaints) to a person
- Labeled test set with a minimum pass rate before each release
- De-duplication so retries never double-send
The path to autonomy
After confidence scores hold steady and the review folder is measured, narrow low-risk intents (hours, directions, simple confirmations) can be switched to auto-send. Everything else stays supervised. The business decides, not the vendor.