AI Pre-Sales Assistant for Home & Field Services
Designed and deployed "Ada," an AI Pre-Sales Assistant designed to act not as a generic chatbot, but as a competent Digital Receptionist. Unlike standard bots that hallucinate answers or push for sales too early, Ada uses a strict architectural framework to answer questions accurately, qualify prospects based on specific business rules, and route structured data directly to a CRM (Airtable) for human follow-up.
Technical Brief
Discuss implementing this for your team
1. Executive Summary
For home and field service businesses (HVAC, Security, Plumbing), the pre-sales process is often plagued by friction. High volumes of repetitive questions ("How much is it?", "Do you serve my area?") clog up communication channels, leading to delayed responses and missed opportunities.
This project deployed "Ada," an AI Pre-Sales Assistant designed to act not as a generic chatbot, but as a competent Digital Receptionist. Unlike standard bots that hallucinate answers or push for sales too early, Ada uses a strict architectural framework to answer questions accurately, qualify prospects based on specific business rules, and route structured data directly to a CRM (Airtable) for human follow-up.
System In Action
2. The Challenge: The "Pre-Sales Noise" Problem
Home service businesses face a specific set of challenges that traditional "dumb" chatbots fail to address:
- Inconsistent Qualification: Human agents might forget to ask about property size or budget, leading to site visits for unqualified leads.
- Premature Selling: Most AI bots try to "close" the deal in the first message, which feels robotic and pushes customers away.
- Hallucination Risks: AI models often invent prices or inventory (e.g., promising "Solar Cameras" when none exist) to please the user.
- Data Silos: Leads from WhatsApp, email, and chat often end up in messy inboxes rather than a structured database.
The goal was to build a system that behaves with the discipline of a trained senior receptionist: helpful, guarded, and process-driven.
03 Solution Architecture
The solution was engineered as a modular automation system using n8n as the orchestrator. It consists of three distinct layers:
A. The Knowledge Layer (Ingestion Workflow)
To prevent hallucinations, the AI is not allowed to rely solely on its training data. Instead, a dedicated Retrieval-Augmented Generation (RAG) pipeline was built.
- Process: Company policy documents, pricing guides, and service descriptions are ingested via n8n.
- Chunking Strategy: Data is split into semantic chunks (1000 characters with 150-character overlap) to ensure complex concepts like "Full Smart Home Packages" are retrieved in their entirety.
- Storage: These chunks are embedded using Google Gemini and stored in a Pinecone vector database, creating a "frozen" source of truth that the AI must cite.
Figure 1: Knowledge Ingestion Pipeline
B. The Reasoning Layer (The "Gatekeeper" Logic)
This is the core innovation of the project. Rather than a simple Q&A bot, the AI Agent is governed by a strict System Prompt that enforces business logic:
- The Recommendation Gatekeeper: The AI is legally forbidden from recommending a package until it has collected four specific data points: Property Type, Size, Primary Goal, and Preference Level. If a user asks "What should I buy?" early, the AI politely deflects and pivots back to qualification.
- The "Soft Pivot": When users ask about out-of-scope items (e.g., "Do you have solar cameras?"), the AI does not hallucinate a "Yes." Instead, it flags the item for human review ("That’s a great question for the specialist...") and immediately returns to the qualification flow.
- Memory Management: A sliding context window (10 messages) ensures the AI remembers details provided early in the chat, preventing repetitive questions.
Figure 2: The Logic Engine & Gatekeeper
C. The Operations Layer (CRM Integration)
Once a lead is fully qualified and requests a consultation, the system transitions from "Chat" to "Data Entry":
- Data Formatting: The AI converts conversational dates ("Next Tuesday at 2 PM") into structured ISO timestamps (2023-10-24 14:00).
- CRM Handoff: The lead is saved to Airtable via API. The record includes not just contact info, but the full qualification profile (Goal, Budget, Property Size), giving the human sales team a "warm" lead to close.
Figure 3: Structured Lead Data
4. Operational Outcomes
The system was deployed with a custom web-frontend connected via secure webhook tunnels.
| Metric | Improvement |
|---|---|
| Response Time | Reduced from hours/minutes to Instant (24/7) |
| Lead Quality | 100% of CRM entries contain required qualification fields |
| Hallucinations | Eliminated via strict RAG + "Soft Pivot" logic |
| Staff Efficiency | Sales team only speaks to leads who have passed the "Gatekeeper" |
5. Why It Works (The "System" Approach)
The success of this project lies in treating AI as infrastructure, not a novelty. By separating Knowledge (Pinecone) from Reasoning (Gemini + n8n) and Storage (Airtable), the system avoids the common pitfalls of "black box" AI.
It does not try to replace the human expert; it replaces the friction of getting to the expert.
6. Future Applicability
While built for Sentinel Smart Homes, this architecture is domain-agnostic. By simply swapping the PDF documents in the Ingestion Workflow, this exact system can be redeployed for:
- HVAC & Plumbing: Qualifying emergency vs. maintenance calls.
- Legal & Consulting: Intake forms for new client case evaluations.
- Real Estate: Pre-screening tenants before scheduling viewings.
This project demonstrates that with the right guardrails—specifically the Gatekeeper and RAG patterns—AI can be safely and effectively trusted with the front door of a business.
Ready to Automate Your Front Desk?
If you want a system that defends your time and qualifies your leads automatically, let's discuss your requirements.
Book System Review