Letterbook Docs
Configure AI replies
Control when Letterbook drafts replies and what sources it should use.
Last updated June 25, 2026
What this guide covers#
Configure AI replies so Letterbook drafts useful responses from the conversation, your knowledge base, and connected customer data.
Start with AI drafts as a reviewed workflow. The goal is to make the first draft accurate enough that a teammate can approve or lightly edit it, while keeping risky actions and sensitive cases under human control.
Setup steps#
- Choose the channels where AI should draft replies.
- Connect the customer data sources the AI needs for those tickets.
- Add internal knowledge for policies, tone, and escalation rules.
- Tune Draft Eagerness so only useful drafts appear.
- Decide whether drafts require human review before sending.
- Decide which suggested actions require explicit approval.
- Send realistic test tickets through each channel.
- Review the draft, sources, and proposed actions before enabling the workflow for production.
What the AI uses#
AI replies can be grounded in several kinds of context:
- The customer message and thread history
- Customer records from connected databases or billing systems
- Published or internal knowledge base articles
- Past support decisions captured in your playbook
- Channel metadata from email, forms, API-created tickets, or Discord
- Approved custom actions and integration context
If the AI does not have the data source or policy it needs, it should ask for review or escalation instead of guessing. Add missing guidance to the knowledge base when you see a repeated gap.
Review requirements#
Use review requirements based on risk:
| Workflow | Recommended review setting |
|---|---|
| General product questions | Human review at launch, then loosen only after consistent quality |
| Account-specific questions | Human review until customer matching and data integrations are reliable |
| Refunds, cancellations, credits, or plan changes | Always require approval for the action |
| Security, privacy, legal, or abuse issues | Route to a human owner |
| Low-risk acknowledgements | Consider a lighter review workflow after testing |
Quality checklist#
- The answer cites the right source material
- The answer follows tone and policy guidance
- The answer includes account-specific context when available
- The answer avoids unsupported promises
- The answer does not expose internal-only notes to the customer
- The proposed action matches the policy and the customer record
- The draft escalates when the source material is missing or ambiguous
Improve bad drafts#
When a draft is wrong, identify the cause before editing the symptom:
- If the policy is missing, add or update a knowledge base article.
- If the customer data is missing, fix the integration or identifier mapping.
- If the action is risky, require approval or narrow the action scope.
- If the tone is wrong, add tone examples and phrases to avoid.
- If the ticket should not have used AI, update channel or triage settings.