How Liya AI works

Liya is designed around one core principle: every AI action should be explainable, grounded, and controllable. This page walks through exactly what happens inside Liya from the moment a ticket lands in your inbox to when an agent reviews the suggested reply.

The triage pipeline

Every inbound message — regardless of channel — passes through the same six-stage pipeline before an agent sees it:

Stage 1: Ingestion and normalisation

Liya receives the raw message (email, chat, widget, or API) and normalises it: stripping email signatures and quoted history, collapsing whitespace, extracting attachments, and identifying the customer's identity from headers or prior conversation history.

Stage 2: Intent classification

Liya classifies the core intent of the message — what the customer is actually asking for. Examples: billing_inquiry, order_status, password_reset,feature_question, complaint, escalation_request.

Intent classification uses a combination of fine-tuned models trained on support conversation patterns and your workspace's own ticket history as it grows. You can view and edit detected intents, and define custom intent categories in Settings → AI → Intent taxonomy.

Stage 3: Signal extraction

Alongside intent, Liya extracts a set of structured signals from the message:

• Urgency — how time-sensitive the customer's situation is (urgent / high / normal / low)
• Sentiment — the customer's emotional tone (frustrated / neutral / positive)
• SLA risk — whether the conversation is approaching or has breached an SLA threshold
• Escalation signal — whether the message contains language indicating a desire to escalate (threats to cancel, requests to speak to a manager, legal language)
• Key entities — order numbers, account IDs, error codes, product names mentioned in the message

Stage 4: Knowledge retrieval

Using the classified intent and extracted entities, Liya performs a hybrid searchacross your knowledge base: combining dense vector search (semantic similarity) with keyword search to retrieve the most relevant articles.

The top-ranked articles are then re-ranked using a cross-encoder model that scores each article against the full message text. Only the highest-quality matches are passed to the next stage.

If no article exceeds the relevance threshold, the conversation is flagged as a knowledge gap and the draft may be withheld (depending on your confidence threshold settings).

Stage 5: Draft generation

The retrieved articles, the customer message, conversation history, and your workspace's tone settings are assembled into a structured prompt and sent to the generation model.

The model generates a reply that:

• Directly addresses the customer's specific question
• Draws only from the retrieved knowledge articles (no hallucination)
• Matches your configured tone (formal / conversational / brand voice)
• Is channel-appropriate (email replies are more formal than chat responses)
• Includes any customer-specific details extracted in Stage 3 (e.g. order number)

Stage 6: Quality evaluation

Before the draft reaches the agent, it is evaluated by a separate quality model that checks:

• Grounding — every claim in the draft is traceable to a retrieved article
• Safety — the reply does not contain harmful, misleading, or legally sensitive content
• Completeness — the reply addresses the full question, not just part of it
• Tone consistency — the reply matches your configured tone settings

The output of this evaluation produces the confidence score (0–100) that agents see on each draft. Drafts below your configured threshold are withheld.

What Liya does not do

Understanding the boundaries of Liya's AI is as important as understanding its capabilities:

• Liya does not fabricate information.If your knowledge base doesn't contain the answer, Liya will not invent one. It will either withhold the draft or generate a reply that acknowledges the limitation and offers to escalate.
• Liya does not auto-send without your approval (unless you explicitly enable auto-send for specific intents). Every draft is a suggestion — agents always have the final word.
• Liya does not access external systems (e.g. your CRM, order management, or billing platform) unless you explicitly connect an integration. Without integrations, all AI context comes from the conversation and your knowledge base.
• Liya does not learn from sent replies by default. Your drafted and approved replies are not used to retrain the model without explicit opt-in.

Model infrastructure

Liya uses a combination of purpose-trained classification models (intent, urgency, sentiment) and frontier large language models for draft generation and quality evaluation. All processing occurs on infrastructure within your selected data residency region.

Customer message content is never used to train shared models. See Data residency for details on where your data is processed.