Pricing question
Scenario: A lead asks about one-bedroom pricing.
System behavior: AI searches the active pricing document for that property and drafts a response.
Result: Agent reviews and sends an accurate answer.
AI communication assistant
AI Response Drafting
AI drafts SMS and email responses using the guest card, recent communications, detected intent, approved templates, and Document Center knowledge.
Human reviewDocument-groundedAudit trailConfidence-based actions
1. Feature Overview
Response Drafting reduces typing effort while keeping agents in control. AI should not invent pricing, policies, or availability. When the answer depends on property facts, it should retrieve from the Document Center / AI Knowledge Base.
2. Drafting Workflow
flowchart TD
A[Inbound SMS/Email/Form] --> B[Intent Recognition]
B --> C[Load Guest Card Context]
C --> D[Retrieve Document Center Content]
D --> E[Generate Draft + Confidence]
E --> F{Mode + Confidence}
F -->|Assistive AI| G[Pending Review Task]
F -->|Automation AI + High Confidence + Low Risk| H[Eligible for Auto-Send]
F -->|Low Confidence / Sensitive| G
G --> I[Agent Approve/Edit/Reject]
H --> J[Send Message]
I --> J
J --> K[Audit Log]
3. Human-in-the-Loop and Pending Task List
- Assistive AI mode routes drafts to a pending review task list.
- Low-confidence outputs require human approval.
- Sensitive/high-risk content requires review regardless of confidence.
- In Automation AI mode, high-confidence and approved low-risk actions may be eligible for auto-send based on customer configuration.
- Audit trail records AI-generated, human-edited, auto-sent, and manually sent actions.
4. AI Learning Loop
AI should improve from agent behavior without bypassing governance.
- Capture agent edits to AI drafts.
- Track approved vs rejected drafts.
- Detect repeated successful responses as knowledge suggestions.
- Use confidence scores to identify where AI needs better documents or prompts.
- Keep learning scoped to PMC/community context to avoid cross-customer mixing.
5. Document Center Connection
- Develop Document Center UI and workflow plan.
- Define Document Center versioning/update plan.
- Include S3 + vector search + PMC/community isolation.
- Retrieve only active approved documents for the correct community/property.
- Fallback when approved information is missing: ask human to confirm rather than inventing an answer.
6. User Cases
High-confidence after-hours response
Scenario: A visitor asks a simple pet policy question at 2 AM.
System behavior: In Automation AI mode, if the active policy document is found and confidence is high, response may be eligible for auto-send based on configuration.
Result: Prospect gets immediate answer while action is audited.
Pending approval
Scenario: A lead asks about complex care or unclear pricing.
System behavior: AI creates a draft and routes it to pending task list for approval.
Result: User approves/edits before send.
7. Acceptance Criteria
- AI drafts SMS/email responses from recent communication context.
- AI retrieves approved Document Center content.
- Drafts include confidence and audit metadata.
- Agent can approve, edit, reject, or send.
- Pending review task list captures low-confidence/sensitive drafts.
- Optional auto-send only applies in configured Automation AI mode.
- Agent edits and outcomes feed the learning loop.