The problem with most business chatbots
You’ve used one. You probably hated it. You asked a real question — something with nuance, something that required a human to actually think — and the bot confidently gave you the wrong answer, looped you through three menus, then offered to restart the conversation.
That’s not AI. That’s a phone tree with better fonts.
When we started building our enterprise chatbot platform, we asked a different question: what if the most intelligent thing a chatbot could do is recognise when it’s out of its depth and hand the conversation to a human — seamlessly, with full context, mid-sentence if necessary?
“Every chatbot vendor told us their AI could handle 90% of queries. In practice it was closer to 40%, and the 60% it got wrong were the ones that mattered most.”
— Enterprise client during discovery
What we actually built
The platform runs across two channels — web chat embedded on the client’s site and WhatsApp Business — with a unified backend. The AI handles common queries using custom training data specific to the business: product info, opening hours, service descriptions, FAQs. But the critical differentiator is the handoff mechanism.
- Confidence scoring on every response. The AI doesn’t just generate an answer — it evaluates how confident it is. Below a threshold, instead of guessing, it flags the conversation for human review and tells the customer someone will be with them shortly.
- Context preservation across the handoff. When a human agent picks up, they see the entire conversation history, the customer’s question, and what the AI considered before deciding to escalate. No “can you repeat that?”
- AI/human toggle. Agents can flip the AI back on mid-conversation for routine follow-ups — like sending a link or confirming business hours — then take back control when the conversation needs judgement.
- Custom training pipeline. The client can update the AI’s knowledge base without developer involvement. New product launched? Upload the spec sheet, the AI learns it within hours.
The architecture decisions that mattered
We built the backend in Go for concurrency — real-time chat at scale demands it. The AI layer uses OpenAI’s API with carefully engineered system prompts and retrieval-augmented generation (RAG) against the client’s own data. WhatsApp integration runs through the Business API with webhook-based message routing.
The decision to use RAG over fine-tuning was deliberate: the client’s product catalogue changes weekly. Fine-tuning would mean retraining constantly. RAG means updating a document store and the AI picks it up immediately.
What surprised us
The metric we expected to matter most was AI resolution rate. The metric that actually mattered to the client was customer satisfaction on escalated conversations. When the AI handed off gracefully — with context, without making the customer repeat themselves — satisfaction on those conversations was higher than when a human had handled it from the start.
Why? Because the AI had already gathered the basic information. The human could skip straight to the actual problem. Customers felt heard faster.
“The best conversations are the ones where the customer doesn’t even notice the handoff happened. They just feel like the response got smarter.”
— Client’s customer service lead
The takeaway for any business considering AI chat
Don’t ask “can AI handle our customer service?” Ask “what are the 30 questions we get asked every day that have clear, factual answers?” Start there. Let the AI handle those. Build a clean escalation path for everything else. Measure both sides.
The goal isn’t to remove humans from customer conversations. It’s to make sure humans only handle the conversations that actually need them — and when they do, they have full context from the start.