Skip to main content

Autonomous LLM AI Agent Configuration

Summary​

The Wabixo AI Agent provides generative conversational intelligence powered by modern Large Language Models (LLMs). Unlike rule-based keyword bots that follow rigid decision trees, the AI Agent comprehends open-ended customer messages, interprets intent, and formulates natural, human-like answers strictly grounded in your business documentation (product PDFs, service catalogs, website links, and FAQ databases).

Organizations can deploy multiple specialized AI Agents, establish strict safety guardrails, enable automated human escalation, capture sales leads, and test interactions in real time using the interactive sandbox.

All AI Agent settings are managed under Automations → AI Agent (/dashboard/automation/aiAgent).

Workspace Layout​

The AI Agent module features a split-pane interface designed for simultaneous configuration and verification:

  • Left Workspace (Configuration & Knowledge): Configure agent personas, connect LLM API keys, manage uploaded knowledge base files, set escalation rules, and inspect knowledge gaps.
  • Right Workspace (Interactive Sandbox): A live WhatsApp chat simulator that allows you to chat with your configured AI Agent, test edge-case inquiries, and inspect source document citations before publishing live to customers.

Step-by-Step AI Agent Setup​

1. General Persona & LLM Calibration​

Navigate to Automations → AI Agent and select the General tab:

  1. Enable Agent: Switch the toggle to ON to activate the AI Agent for inbound customer messages on your active WhatsApp line.
  2. Agent Name: Assign an internal identity (e.g., Customer Care Copilot or Sales Inquiries Agent).
  3. LLM Model Selection: Choose your preferred Large Language Model (LLM) version optimized for speed and conversational reasoning.
  4. Temperature Slider: Calibrates response creativity versus deterministic precision:
    • Range: 0.0 (strictly factual) to 1.0 (highly creative).
    • Recommended Setting: We recommend 0.2 for business customer support. A low temperature ensures the model adheres faithfully to your knowledge documents and does not invent unverified information.
  5. Context Window Size: Specifies the number of preceding conversation turns remembered by the AI (default: 15 turns) to maintain natural conversational context.
  6. System Prompt (Agent Persona): Define your agent's identity, guidelines, and operational boundaries: 
    You are the official customer service assistant for [Your Company Name].
    Your goal is to answer customer questions politely, concisely, and helpfully.
    Always base your answers strictly on the provided Knowledge Base documents.
    If the answer cannot be found in the knowledge base, state that you do not have that information and offer to connect the customer with a human team member.
    Never invent pricing, discounts, return windows, or technical capabilities.
  7. Thinking Message Status (Optional): Enter a brief interim status message (e.g., "Checking our catalog...") displayed to the customer while the AI retrieves and processes complex multi-page documents.

2. API Key Management (BYOK vs. Managed AI)​

Under the API Keys tab, choose how your organization connects to the LLM infrastructure:

ModeBilling & CostSetup Process
Bring Your Own Key (BYOK)Incur zero platform token markup. Token consumption is billed directly by your LLM provider account.Enter your LLM provider API key in the input field, click Test Connection to verify validity, and save.
Managed AINo external API accounts required. All token costs are deducted automatically from your Wabixo prepaid wallet balance.Activate Managed AI with a single click without managing third-party cloud credentials.

3. Knowledge Base Ingestion (Training Your AI)​

The Knowledge Base (KB) tab provides Retrieval-Augmented Generation (RAG) grounding tools to ensure your AI speaks with factual accuracy:

  1. Document Uploads:
    • Upload .pdf, .docx, or .txt files containing your official warranty policies, product spec sheets, pricing menus, or store guidelines.
    • Wabixo automatically parses, chunks, and indexes document text for fast semantic retrieval.
  2. Website Link Ingestion:
    • Enter your public website URLs or online help center links. Wabixo indexes the published page content so your agent remains aligned with current web offerings.
  3. Structured FAQ Pairs:
    • In the FAQs tab, enter exact Question & Answer pairs for high-frequency inquiries requiring verbatim, standardized responses (e.g., return policies, operating hours, payment methods).

4. Human Escalation & Safety Guardrails​

To prevent customer frustration, the Escalation tab provides safety controls:

  1. Keyword Escalation Triggers:
    • Configure words that immediately trigger human takeover (e.g., agent, human, person, representative, help).
    • When a customer sends any of these triggers, the AI Agent stands down immediately.
  2. Customer Confusion & Sentiment Detection:
    • Set the confusion sensitivity threshold (Low, Medium, or High).
    • If the AI detects repeated customer confusion or negative sentiment over multiple turns, it automatically initiates escalation.
  3. Escalation Target & Team Handover:
    • Choose which Team Group or department queue receives the escalated conversation in the Team Inbox.
    • Configure an automated handover message (e.g., "I am connecting you with a member of our customer care team. One moment please.").
    • The conversation transitions from the AI queue into the human Open Queue, alerting active online agents.

5. Sales Intent & Callback Scheduling​

Turn automated conversations into qualified revenue opportunities:

  • Sales Intent Capture: Under the Sales Intent tab, configure the AI to recognize commercial buying intent (e.g., inquiries about bulk orders, enterprise plans, or custom quotes). When triggered, the AI automatically applies a designated tag (e.g., #High-Intent-Lead) and routes the chat to your sales group.
  • Callback Booking: Under the Callback tab, enable automated callback scheduling. The AI can collect the customer's preferred call window and discussion agenda, creating a structured task for outbound agents.
  • Dormancy Nudges: Under the Nudges tab, configure automated follow-ups if a customer stops responding mid-consultation.

6. Testing in the Interactive Sandbox​

Before publishing your AI Agent to live customer lines:

  1. Locate the Interactive Sandbox panel on the right side of the screen.
  2. Type test questions (e.g., "What is your refund policy?" or "Can I book an appointment for tomorrow?").
  3. Review the AI's generated response:
    • Check the Knowledge Sources badge to see which specific uploaded document or FAQ entry the AI referenced.
    • Verify that tone, accuracy, and formatting align with your brand standards.
  4. Test human escalation by typing "I want to speak with an agent" to confirm that the handover prompt and routing trigger smoothly.

7. Knowledge Gap Analysis​

Under the Gaps tab, Wabixo records customer questions that the AI was unable to answer with high confidence:

  • Review the list of unanswered inquiries to identify missing topics in your documentation.
  • Click Add to FAQs next to any recorded gap to provide the verified answer, permanently eliminating that knowledge blindspot with a single click.

Customer Troubleshooting & FAQs​

Problem: The AI Agent answers that it does not know the answer to a question​

  • Why it happens: The necessary information does not exist in your uploaded Knowledge Base documents or FAQs, or the semantic similarity score was below the minimum RAG confidence threshold.
  • How to fix it: Add the missing information into your Knowledge Base as a document or add a direct entry under the FAQs tab.

Problem: How do we prevent the AI from hallucinating or inventing details?​

  • Why it happens: Setting a high temperature or using vague system prompts allows LLMs to generate creative approximations.
  • How to fix it:
    1. Set your Temperature to 0.2 under the General tab.
    2. In your System Prompt, explicitly instruct the agent: "Answer strictly based on the provided Knowledge Base documents. Do not assume or invent information not explicitly stated in the documents."

Problem: The AI Agent stopped replying in the middle of a customer conversation​

  • Why it happens:
    1. A human agent opened the thread or typed a reply, activating the Human Takeover safeguard.
    2. The customer triggered an escalation keyword or expressed confusion.
    3. The customer's 24-hour WhatsApp messaging window expired.
  • How to fix it: Check the conversation in the Team Inbox. If human takeover is active, the agent can resolve the query or release takeover in the chat header.