What We Were Solving
How We Built It
Custom AI chatbot built on OpenAI API with RAG, integrated with website, WhatsApp Business API, and HubSpot CRM. 240-entry knowledge base trained on 6 months of enquiry data. Automated document fulfilment, Calendly scheduling integration, intelligent human handoff with full context. 3-month optimisation expanded knowledge base to 312 entries with 18 automated workflows.The Challenge
The client is a Johannesburg-based B2B professional services firm with approximately 400 active clients. Their customer service team of three handled everything: service enquiries, onboarding questions, document requests, scheduling, billing queries, and general support.
The volume was unmanageable. The team received an average of 340 inbound enquiries per month across email, WhatsApp, and a basic website contact form. Response times averaged 4.2 hours during business hours. After hours and weekends, enquiries sat unanswered until Monday morning.
Client satisfaction scores were slipping. A post-interaction survey showed a satisfaction rating of 6.8 out of 10 — dragged down primarily by slow response times and the need to repeat information across channels.
The firm couldn't justify hiring a fourth support person at a cost of approximately R25,000–R35,000 per month including benefits. They needed a solution that could handle routine enquiries immediately while freeing the human team for complex work.
What We Built
We built a custom AI chatbot integrated with their website, WhatsApp Business API, and their existing CRM (HubSpot). This wasn't a template chatbot with scripted responses. It was a context-aware AI assistant trained on the client's specific knowledge base.
Phase 1: Knowledge Base and Training (Week 1–2)
We audited 6 months of customer enquiries. We categorised them by type and frequency:
82% of enquiries were low-to-medium complexity — perfect candidates for AI handling.
We built a knowledge base of 240 question-answer pairs, covering every common enquiry, product detail, process explanation, and policy.
Phase 2: Build and Integration (Week 2–3)
The chatbot was built on a custom stack using OpenAI's API with retrieval-augmented generation (RAG). The bot understood context within a conversation, integrated with HubSpot for client data, triggered automated workflows for document requests, connected to Calendly for scheduling, and featured intelligent handoff to human agents with full conversation context.
Phase 3: WhatsApp Integration (Week 3–4)
Over 60% of inbound enquiries came via WhatsApp. The chatbot was integrated with WhatsApp Business API, providing the same AI-powered experience within the app clients already used.
Phase 4: Optimisation (Months 1–3)
We monitored every conversation for three months. We expanded the knowledge base from 240 to 312 entries and added 18 new automated workflows.
The Results
The chatbot handles 207 of 340 monthly enquiries without human involvement. The support team spends 60% less time on enquiries, redirecting capacity to revenue-generating activities.
Cost Analysis
Total first-year cost: R105,600 | Annual savings: R456,000 | ROI: 4.3x in year one
Client Testimonial
Key Takeaways
The chatbot succeeded because it was trained on the client's specific data, integrated with their specific systems, and optimised over time. For B2B companies handling high volumes of routine enquiries, a custom AI chatbot is a competitive advantage. R105,000 investment, R456,000 annual savings, and happier clients.