How AI-drafted replies work

Lettr drafts a reply for you the moment a customer message arrives. The draft appears pre-filled in the composer when you open the conversation. You can send it as-is, edit, or discard.

What goes into a draft

A draft is generated from three sources:

• The current customer message and the conversation history.
• Your saved replies and any examples from previously-resolved tickets in your enterprise.
• A generic-tone fallback when there isn't enough context yet.

When the draft is good

Drafts are most reliable on common, repeated questions — order status, password resets, refund policy, business hours. The model has seen similar phrasings many times and produces a confident, on-brand answer.

When to ignore the draft

Edge cases, complaints, anything legally or financially sensitive, or any conversation that depends on real-time data the model can't see. When in doubt, write the reply yourself or escalate.

Improving drafts over time

The more you reply, the better the drafts get. After about 30–50 resolved tickets, drafts noticeably reflect your team's tone and reasoning. To accelerate this, seed your common-questions list under Bot Agent → Examples.

Privacy

Drafts are generated server-side; the customer's message and your saved replies are passed to the model and not retained beyond the request. Every message — inbound and your reply — also passes through norse3 AI safety to screen threats, abuse, and PII leaks.