Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.lightwork.co/llms.txt

Use this file to discover all available pages before exploring further.

​
In one minute

Felicity is LightWork’s AI agent. She reads incoming messages, understands what the sender needs, finds the right information, and responds - or escalates when she cannot handle something confidently. This page explains how she works under the hood so you can understand her behaviour and get the best results.

​
Message processing pipeline

When a message arrives, Felicity processes it through a series of steps:
1

Preprocessing

The message is normalised and prepared. Felicity analyses the tone and sentiment, checking for indicators that might need special handling (frustration, legal language, urgency).
2

Intent classification

Felicity determines what the sender is trying to do. Is it a general enquiry? A maintenance report? A lettings enquiry? This step routes the message to the correct module:
  • Resident Desk - general tenant enquiries and questions.
  • Prospects - prospective tenant screening and viewings.
  • Maintenance - issue reporting and work order coordination.
  • Compliance - certificate and renewal queries.
3

Context loading

Felicity loads all relevant context: the sender’s identity, their property and unit, previous conversation history, and any open cases or work orders.
4

Knowledge base search

For enquiries that need an answer, Felicity searches your knowledge base for relevant articles. She uses the scope hierarchy (global > property > unit) to find the most specific answer available.
5

Planning

Felicity determines the next actions: what information to send, who to message, and what workflow steps to progress. Each module has its own planning logic tailored to its workflow.
6

Response generation

Felicity composes the response in a clear, professional tone. She adapts her language to the channel (shorter for WhatsApp, more detailed for email) and the context of the conversation.
7

State transition

The workflow state is updated to reflect what has happened. This tracks where each conversation and case sits in the overall process.

​
Conversation management

​
Channel lock-in

When a conversation starts on a particular channel (email or WhatsApp), Felicity continues on that channel to avoid confusion. If someone contacts you on both channels about different topics, separate conversations are created.

​
Chase cadence

When Felicity is waiting for a response - from a tenant, contractor, or other party - she follows a consistent follow-up pattern:
  1. Initial message sent with the request or question.
  2. First reminder sent after a set number of days if no response.
  3. Second reminder sent after a further wait.
  4. Escalation - if no response after multiple attempts (typically 3), Felicity escalates to your team or closes the conversation depending on the context.
This keeps conversations moving without sending excessive messages.

​
Multi-party coordination

In workflows like maintenance and prospects, Felicity communicates with multiple people (tenants, contractors, landlords, letting managers). She keeps each party informed with only the information relevant to them and coordinates between them without requiring your team to relay messages.

​
Tone and style

Felicity writes in a clear, professional, and friendly tone. She:
  • Uses plain language and avoids jargon.
  • Keeps messages concise and action-oriented.
  • Adapts message length to the channel (shorter for WhatsApp).
  • Maintains a consistent voice that represents your organisation.
  • Never makes promises she cannot keep (e.g. she will not guarantee resolution times unless specified in your knowledge base).

​
Escalation triggers

Felicity escalates a conversation to your team when she detects any of the following:
TriggerDescription
No confident answerThe knowledge base does not contain a match for the question.
Legal languageThe sender mentions solicitors, court proceedings, Section 21, or similar legal terms.
ComplaintThe sender is making a formal complaint or expressing significant dissatisfaction.
Human requestedThe sender asks to speak to a person, manager, or “someone real”.
EmergencyThe issue involves immediate safety or security risks.
Vulnerability indicatorsThe sender mentions health issues, disability, financial hardship, or other vulnerability factors.
Frustration detectedThe sender’s tone indicates high frustration after multiple exchanges.
Multiple unanswered questionsThe sender has asked several questions Felicity cannot address.
Repeated failureFelicity has attempted to resolve the issue multiple times without success.
When escalating, Felicity provides your team with a summary of the conversation, the reason for escalation, and suggested next steps.

​
What Felicity does not do

  • Guess - if she does not have a confident answer, she escalates rather than making something up.
  • Make commitments - she does not promise specific outcomes, timelines, or financial amounts unless they are in your knowledge base.
  • Bypass human approval - for decisions that need your team’s judgement (contractor selection, landlord approval, complex cases), Felicity pauses and waits for your input.
  • Store sensitive data - she processes information to help with conversations but follows data handling policies. Financial details and sensitive personal data are handled according to your data protection requirements.

​
Improving Felicity’s performance

The single most impactful thing you can do is improve your knowledge base:
  • Add articles for questions Felicity escalates frequently.
  • Review AI suggestions - Felicity proposes new articles based on questions she could not answer.
  • Keep articles current - outdated information leads to poor answers or escalations.
  • Use specific scoping - property-level and unit-level articles give more accurate answers than generic global articles.
See Knowledge Base for a full guide.