Architecting a full-stack conversational AI system for a 40-year-old HVAC and refrigeration distributor — replacing a one-person manual operation handling 100+ daily conversations with an intelligent, SAP-integrated agent across 7 countries.
Our client is a leading HVAC and refrigeration distributor with 40+ years of regional presence across 7 countries in Latin America and the United States. They are the primary importer for major brands in their markets, serving a diverse customer base of end consumers, independent technicians, contractors, corporate clients, and multi-country distributors.
Their WhatsApp number had become the backbone of their sales operation — receiving over 100 conversations per day in their pilot market alone, representing 75% of total inbound volume. A single person was managing all of it, manually. The cost of that fragmentation was becoming impossible to ignore.
| Metric | Current State |
|---|---|
| Daily conversations (pilot market) | ~100 chats on WhatsApp |
| Total regional volume | +36,000 conversations |
| Primary inbound channel | 75% via WhatsApp |
| Lead type breakdown | 80% new / 20% returning |
| Existing bot capability | Stalls without human handoff |
| Current sales process | 1 person "putting out fires" |
~3,000 unqualified leads per month — arriving without context, product interest, or customer type identified, requiring manual triage before any sale could begin.
30–40% estimated revenue loss from poor lead management — conversations that went cold, unanswered inquiries, and delayed quotes that competitors closed first.
900–1,200 lost opportunities per month — potential transactions that fell through the cracks due to volume the team couldn't absorb.
Zero metrics on which channel actually converts, which customer type closes fastest, or where the pipeline breaks down.
No structured follow-up — once a conversation went silent, it was gone. No recovery sequences, no re-engagement, no second chance.
CCG designed and is building a full-stack conversational AI system integrating WhatsApp Business API, SAP Business One, and a custom n8n automation layer — deployed in a Central American pilot market before regional rollout across all 7 countries.
Handles quote requests via text, PDF, image, or screenshot. Queries the SAP data warehouse for real-time pricing by customer type (End Consumer, Technician, Corporate, Distributor). Generates quotes directly in SAP B1 and delivers them via WhatsApp. Complex items — chillers, cooling towers, applied projects — are detected and escalated to Engineering.
Queries the data warehouse for real-time stock across multiple warehouses. Returns availability status per customer type, with delivery timelines: 2–4 business days in the pilot market, coordinated delivery to all major cities in-country. Express dispatch coordinated with client's own carrier.
Collects claim details from the customer, verifies purchase against SAP records, and routes the case via the confirmed handoff method — email notification, human-in-the-loop WhatsApp escalation, or SAP case creation. Warranty terms served per brand: ADINA (2yr compressor / 1yr factory), TCL (10yr compressor), McQuay/Daikin (5yr compressor).
Serves the training catalog from a vector knowledge base. Handles enrollment booking — either via link or automated n8n booking flow. ADINA-certified technician network is surfaced for installation inquiries. Training content is maintained and updated without re-deployment.
RAG-powered knowledge base covering the full product catalog across 20+ brands: Adina, TCL, Daikin, Trane, Panasonic, Elgin and more. Handles BTU sizing guidance, refrigerant type queries (R410a, R32, R404, R507), inverter technology questions, spare parts requests, store hours, payment methods, and more — with full multi-turn context handling.
When a conversation is ready for a human, the system generates a complete summary and routes it intelligently: simple requests to local sales, remote clients to televentas, complex engineering projects to the technical team. Automated re-engagement sequences activate when a customer goes silent. Every outcome is tracked across 4 states for dashboard reporting.
The system was architected across six sequential phases with intentional parallel execution — targeting a go-live deadline of 10 weeks from project kick-off, with a soft launch period built in before hard go-live.
Cloud vs. on-prem deployment confirmed per client data privacy requirements. Dev/staging/prod environments with CI/CD pipeline.
Meta Business account configuration, webhook setup, inbound/outbound message testing across all conversation types.
Service Layer / OData endpoint validation for both read (pricing, inventory) and write operations (customer creation, quote, order).
Read-only credentials, schema documentation, and query validation for products, inventory, and pricing tables.
Two instances: (1) FAQ / Knowledge Base, (2) product catalog fallback. Tooling: Pinecone / Qdrant / pgvector.
Customer types, conversation state machine, outcome tracking schema, lead storage — designed before any routes are built.
4-way classification logic: End Consumer, Technician, Corporate, Distributor. Drives all downstream pricing and routing rules.
DW price lookup, customer-type pricing rules, multi-item cart, SAP B1 quote object creation. Complex items escalated to Engineering.
Conversation summary generation, CRM/WhatsApp group push, intelligent routing: Simple → Local Sales, Remote → Televentas, Complex → Engineering.
Multi-warehouse stock queries. Claims collection, warranty verification, and SAP case creation or escalation.
RAG over knowledge base vector store with multi-turn context. Automated follow-up sequences triggered on conversation silence.
4-state logging: bot-closed sale, bot-no sale, human-closed sale, human-no sale. Feeds all dashboard KPIs.
Embedded on client website. Reuses same AI/SAP backend as WhatsApp — customer builds their own quote without contacting anyone. PDF downloadable.
Conversation volume, conversion by route, handoff rate, outcome breakdown, response times. Sales and engineering team views included.
5,000 conversations/month throughput simulation. 20 test flows (5 routes × 4 customer types). Structured UAT with client team before soft launch.
Every component was selected for its ability to handle the specific requirements of a multi-country distributor operating on SAP — not for ease of setup.
Orchestration layer connecting all systems. Custom conversation state machine, routing logic, and follow-up sequences.
Primary communication channel via Meta Business. SIP trunk routing for voice agent extension (optional phase).
Service Layer / OData for write operations: customer creation, quote generation, order creation. DW for read queries.
Conversational AI for natural language understanding, customer classification, quote parsing, and summary generation.
Two-instance setup: FAQ/KB and product catalog. Pinecone / Qdrant / pgvector evaluated per performance requirements.
Conversation state, outcome tracking, lead storage, and CRM views — designed to feed both dashboard and sales team workflows.
Deployed first in a Central American pilot market before regional rollout — transforming a manual, single-person WhatsApp operation into an intelligent, always-on system that qualifies, quotes, and routes every conversation with precision.