Why drafts first
In our scorecard, drafted replies pass all three tests: messy input, cheap mistakes, and a human sees the output before it matters. That makes them the ideal first AI feature for a team that is nervous about bots. Nothing reaches a customer without a person choosing to send it.
The setup
When a new inbound message arrives, a workflow sends the recent conversation, the contact's key fields (service interest, stage, last appointment) and the approved knowledge base to the model, and writes the suggested reply into the conversation as an internal note or a pre-filled draft. The rep sees it beside the message, edits, and sends. Total added latency: a few seconds.
The knowledge base is the same one the booking bot uses, so facts stay consistent across channels. The prompt is a cut-down version of the bot template: role, knowledge, rules, examples, format.
Keeping the voice
Two inputs shape tone. A one-page style guide: greetings, sign-offs, words to use and avoid, how formal to be. And thirty of the team's best real replies, chosen by the owner, included as examples. The drafts start sounding like the team on day one instead of like a help-desk template.
For small teams, one shared voice. For larger teams, each rep can have their own example set, so Emily's drafts sound like Emily. The knowledge base stays shared either way.
Team habits
The feature works as long as people keep editing. We ask reps to change at least one thing in most drafts for the first month, which keeps them reading closely. Anything factually wrong gets flagged with a reaction so the knowledge base can be corrected. Drafts that are perfect as-is are fine to send, but "send without reading" is the habit that eventually embarrasses someone.
What to measure
Minutes from inbound message to reply, and the percentage of drafts sent with edits. In the first month replies get about twice as fast and around 70 percent of drafts are edited. Over time the edit rate falls as the knowledge base improves. If it falls to near zero while customers start complaining about tone, the team has stopped reading, and it's time for a reminder.
Key takeaways
- Drafts are reviewed before sending, so mistakes are caught and the risk is low.
- Feed the draft engine the conversation, the contact's fields and the same knowledge base the bot uses.
- Style comes from a short guide plus the team's own past replies.
- Edit, don't just send. The moment people stop editing, the voice drifts.
- Measure minutes per reply and edit rate. Falling edits with steady quality is the goal.