Title: Entexis AI Assistant — Our Website Had 97% Bounce Rate. Then We Gave Visitors Someone to Talk To. Client: Entexis Systems (Internal) Industry: SaaS URL: https://entexis.in/case-studies-saas-development-company/entexis-ai-assistant-website-chatbot-agent --- Our website was getting traffic but losing visitors to unanswered questions. The contact form was a barrier. The FAQ page was stale. Live chat needed humans 24/7. So we built an AI assistant that knows every service, case study, and engagement model — answers in natural language, captures leads when it detects buying intent, and adapts based on which page the visitor is on. Four iterations from FAQ bot to production agent. Test it right now — click the chat icon on this page. The Problem Space ## Every Software Company's Website Has the Same Problem Our website was getting traffic. People visited service pages, read case studies, explored our process. And then they left. The contact form sat there like a barrier — too much commitment for someone who just wanted a quick answer. The FAQ page was outdated before we finished writing it. And live chat meant someone on our team had to sit and wait for messages that might never come. We knew visitors had questions. We could see it in the analytics — people spending 4-5 minutes on a service page, scrolling up and down, then bouncing. They wanted to know something specific. They just did not want to fill out a form to find out. 97% Visitors Leave Without Converting The average website converts 2-3% of visitors. The rest leave with unanswered questions. We were no different — until we gave those visitors someone to talk to. 6+ Hrs Average Response Time A visitor who asked a question via the contact form waited hours for a human response. By then, they had already contacted two competitors. Stale FAQ Pages Are Outdated on Day One We updated our FAQ page twice in two years. Meanwhile, visitors asked questions we had never thought to include. Static content cannot keep up with dynamic curiosity. "The first month of conversation logs taught us more about what potential clients actually want than two years of Google Analytics." — What we learned after launching v1 The Impact ## What Changed AfterWe Stopped Ignoring Visitors Four iterations. Each one fixing problems the previous version created. Here is what the production agent delivers today. Instant Response to Every Visitor Question No more waiting for a human to check email. The agent answers immediately, 24/7, in natural language with links to relevant pages. New Market Intelligence Channel Every conversation is a window into what the market wants. Competitor mentions, feature requests, pricing expectations — data we never had before. Zero Off-Topic Responses Since v4 Guardrails block poems, homework help, and off-topic requests. Every API dollar goes to genuine business conversations. Live Proof of Our AI Capabilities When a client asks "Can you build an AI agent?" we say "Open the chat on our website and try it." Nothing sells AI like a working demo. Before Visitors bounced with unanswered questions. FAQ page was stale. Contact form was a barrier. Zero visibility into what visitors actually wanted. After Every question answered instantly. Leads captured contextually. Page-aware responses. Conversation logs revealing market intelligence we never had. The Proof The agent is live right now on this page. Click the chat icon. Test it. Ask about our services, try off-topic questions, see how it handles pricing. It is the demo. How It Works ## From Visitor Questionto Intelligent Response Four layers. One seamless conversation. The entire flow happens in under 2 seconds. 01 ### Visitor Asks a Question The chat widget captures the message, detects which page the visitor is on, loads their conversation history, and sends everything to the intelligence layer. Session persists across page reloads. 02 ### Knowledge Retrieval The RAG pipeline searches 63 knowledge sources — crawled web pages, manual entries, pricing models, FAQs — and injects the most relevant content into the AI's context. The agent answers from our actual content, not hallucinated data. 03 ### Guardrails Check Before the response reaches the visitor, 20+ rules check it: no specific pricing, no off-topic answers, no confidential information. The agent includes links to relevant service pages and case studies automatically. 04 ### Lead Intelligence The system monitors for buying signals — pricing questions, timeline mentions, project requirements. When intent is detected, a lead capture form appears naturally. Every conversation is logged for market intelligence. Platform Features ## What Makes This AgentActually Production-Ready Every feature exists because a real conversation revealed a gap. Nothing was built speculatively — every capability was added after we watched visitors struggle without it. Feature 01 ### RAG Knowledge Base Every page on our website is crawled, chunked, and stored as searchable knowledge. When a visitor asks a question, the agent retrieves the most relevant content and generates an answer grounded in our actual information — not hallucinated from general training data. Knowledge Sources 50 Auto-crawled pages 13 Manual knowledge entries URLService pages (9) URLCase studies (12) TextPricing models, FAQs, team info Feature 02 ### Page-Aware Context The agent knows which page the visitor is currently on and tailors its responses accordingly. The same question gets a completely different answer on the CRM page vs the homepage vs the contact page — just like a human would. Same Question, Different Context On /services/crm-development "We build industry-specific CRM systems with custom pipelines, role-based access, and..." On /contact "Ready to discuss your project? Share your requirements and our team will..." On homepage "We offer SaaS, CRM, AI, and custom software development across 5 continents..." Feature 03 ### Guardrail System Over 20 rules define what the agent must not do. It refuses off-topic requests, never quotes pricing, redirects competitor comparisons diplomatically, and blocks attempts to use it as a free general-purpose AI. Every guardrail exists because something went wrong without it. Guardrail Rules BlockOff-topic: poems, code, homework BlockSpecific pricing or dollar amounts RedirectCompetitor comparisons AllowServices, case studies, process AllowGeneral software questions The poem incident: v1 wrote a user a poem about the moon. On our API bill. Feature 04 ### Contextual Lead Capture Instead of interrupting after a fixed number of messages, the lead form triggers when the conversation shows buying intent — pricing questions, timeline discussions, specific project requirements. It feels like a natural next step, not an annoying popup. Buying Intent Keywords pricing timeline budget get started proposal requirements When intent detected → Lead form appears. Name + email. "Continue" or "Skip". No pressure. v2 triggered after 3 messages. v4 triggers on intent. Conversion feel is completely different. Feature 05 ### Conversation Memory If a visitor starts a conversation, leaves the page, and comes back hours later — the conversation picks up where it left off. No repeating yourself. Session persists in the browser and full history loads from the database. Session Persistence Monday 2:15 PM Visitor asks about CRM development... — visitor leaves, comes back next day — Tuesday 10:30 AM Conversation resumes with full history intact 24-hour session window. History restored from database on reload. Feature 06 ### Admin Dashboard & Analytics Full management interface — conversation logs with full transcripts, knowledge base editor, bot configuration, lead tracking, and analytics. Read every conversation. Update knowledge weekly. Improve the agent based on real data. Conversations 142 This month Leads Captured 23 Via chatbot Conversations vs Leads — 30 Days - Conversations Leads Under the Hood ## A Closer Look atWhat Powers the Agent Deep Dive 01 ### The Four Iterations This agent was not built in one sprint. It evolved through four major versions, each one fixing problems the previous version created. Evolution — v1 to v4 v1 Retired FAQ Bot One-sentence prompt. Hallucinated pricing. Wrote poems on our bill. v2 Retired RAG + Leads Knowledge base. Lead form. Too aggressive — popped up after 3 messages. v3 Retired Page-Aware Context by page. Links in responses. Started being genuinely useful. v4 LIVE Production Agent Guardrails. Intent-based leads. Memory. Proactive. Rating. Retry. Each version fixed problems the previous one created - v1 to v4 in under 6 months of iteration - Production-tested with real visitor conversations - Zero off-topic responses since v4 guardrails Deep Dive 02 ### Guardrails — What the Agent Must Not Do Version 1 had zero guardrails. It answered anything — poems, homework, coding problems. Every off-topic response cost money and delivered zero value. Here is what v4 enforces. Guardrail Rules — What Gets Blocked vs Allowed BLOCKED ✕ Off-topic requests (poems, code, advice) ✕ Specific pricing or dollar amounts ✕ Confidential internal information ✕ Pretending to be anything else ✕ Writing code or completing homework ALLOWED ✓ Services, case studies, process ✓ General software development topics ✓ Engagement models (no specific prices) ✓ Team, working hours, contact info ✓ Redirect to team for pricing/proposals THE POEM INCIDENT v1 wrote a user a poem about the moon. Beautifully. On our API bill. This guardrail exists because of that. - Blocks off-topic requests — poems, code, homework - Never quotes specific pricing or dollar amounts - Redirects competitor comparisons diplomatically - Allows services, case studies, and general software topics Deep Dive 03 ### What the Conversation Logs Revealed The most valuable output of this project is not the chatbot itself — it is what the conversations taught us about our market. Insights From Real Conversations - Competitive Intelligence We Never Had Multiple visitors asked how we compare to specific agencies we had never heard of. Market intelligence you cannot get from any analytics tool. Questions We Never Anticipated Team size, remote work policy, NDA process, post-launch support, whether we work with startups or only enterprises. None were on our FAQ page. All of them are now in the knowledge base. Conversion Prediction The first question someone asks predicts whether they will convert. Visitors who start with a specific problem convert at dramatically higher rates than those who start with general exploration. Intent Patterns by Page Visitors on the homepage explore. Visitors on case study pages evaluate. Visitors on the contact page are ready to act. Page awareness makes this actionable. Competitive intelligence from visitor questions - Questions we never anticipated — now in knowledge base - First question predicts conversion likelihood - Intent patterns differ by page — homepage explores, contact converts Technology Stack ## What Powers the Agent Node.js + Express Backend server handling chat API, knowledge retrieval, session management, and SSE streaming. Backend Multi-LLM Support Provider-agnostic architecture supporting Claude, GPT-4, Gemini, and open-source models. AI Layer Shadow DOM Widget Self-contained chat widget with complete style encapsulation. Works on any website without CSS conflicts. Frontend SSE Streaming Server-Sent Events for real-time token-by-token response delivery. Response appears as the AI generates it. Real-Time MySQL + Pooling Conversations, messages, knowledge chunks, rate limits, and leads — stored with efficient connection pooling. Database Admin Dashboard Full management — conversation logs, knowledge editor, bot config, analytics, and lead tracking. Operations Cheerio Web Crawler Automated page crawling that extracts, cleans, and chunks website content into searchable knowledge entries. Re-crawls on schedule via node-cron. Knowledge Pipeline Anthropic Claude SDK Primary AI provider integration via the official Anthropic SDK. Handles streaming responses, system prompts, and conversation context management. AI Provider Rate Limiting & Sanitization Per-IP and per-session rate limiting prevents abuse. Input sanitization strips injection attempts before they reach the AI layer. Security Helmet + Encryption HTTP security headers via Helmet. Conversation data encrypted at rest. Session tokens rotated on each interaction. Infrastructure EJS + Server-Side Rendering Full server-side rendering for SEO and performance. No client-side JavaScript framework — the widget loads independently via Shadow DOM. Rendering What's Possible ## Where This AgentCan Go Next An agent like this is built to extend. Here is where the technology can go next — each one a natural evolution of what already exists. Live ### RAG Knowledge Base 63 sources auto-crawled and manually curated. The agent answers from real content, not hallucinated data. Live ### Page-Aware Context & Links Responses adapt based on which page the visitor is on. Every mention of a service or case study includes a clickable link. Live ### Guardrails & Contextual Lead Capture 20+ rules prevent off-topic and pricing leaks. Lead form triggers on buying intent, not arbitrary message counts. Next ### Voice Input Add a microphone button for voice-to-text input. Visitors speak their question, Whisper transcribes, the agent responds. Especially powerful for mobile and regional language users. Next ### Semantic Search Replace keyword-based knowledge retrieval with vector embeddings. The agent finds relevant content based on meaning, not just matching words. Better answers for ambiguous questions. Next ### CRM Actions Let the agent take actions beyond capturing leads — create CRM records, schedule follow-ups, assign to team members, and trigger automated email sequences directly from the conversation. Next ### Multi-Language Support Respond in Hindi, Tamil, and other regional languages. Detect the visitor's preferred language automatically and switch context accordingly. ### Try It Right Now The AI assistant is live on this page. Click the chat icon in the bottom-right corner. Ask about our services. Try off-topic questions. See how it handles pricing. It is the demo. [Build One for Your Business →](/contact-saas-crm-development-company)