A services business with a busy WhatsApp and website chat presence had a problem that looks like success: message volume growing faster than staff could answer. Response times stretched to hours; after-hours messages waited overnight; and two support staff spent their days retyping the same fifty answers - prices, availability, process steps, requirements - in English, Nepali, and the mixture customers actually use.
The Goal - and the Fear
The target was simple: instant, correct answers to routine questions, 24/7, in the customer's language - with humans freed for the conversations that need judgment. The fear was equally simple: every business owner has seen chatbots confidently inventing prices and policies. The client's first sentence was, "It must never make things up."
How the Bot Was Built to Not Lie
That requirement dictated the architecture (the same pattern across our chatbot builds):
- A curated knowledge base, not an open mind. Services, prices, policies, and process docs were cleaned and structured. The bot retrieves relevant passages and answers from them - the RAG pattern - rather than from the model's imagination.
- An engineered "I don't know." Questions outside the knowledge base trigger a polite handoff, never improvisation: "Let me connect you with the team for that."
- Hard rails on sensitive topics. Complaints, negotiations, and refund requests route straight to humans by rule - with full conversation context attached.
- Bilingual by test, not assumption. The bot was evaluated against real customer message logs - including romanized Nepali and code-switching - before launch.
The unglamorous secret: most of the project was knowledge-base work, not AI work. The bot is only as correct as the documents behind it - which forced the client to finally write down policies that had lived in staff heads for years. That documentation alone had value.
Rollout: Trust in Stages
- Weeks 1-2 - shadow mode: the bot drafted answers that staff approved before sending. Accuracy was measured, gaps in the knowledge base were filled.
- Weeks 3-4 - supervised autonomy: routine categories answered directly; everything else drafted for approval. Staff spot-checked transcripts daily.
- Month 2 - steady state: full autonomy on routine questions, instant human escalation paths, and a weekly transcript review habit that continues today.
Results and Honest Limits
- Roughly 70% of incoming questions are now resolved by the bot without human involvement - measured over months, not a launch-week snapshot.
- Response time for routine questions: hours → seconds, around the clock. After-hours inquiries - previously lost overnight - convert at rates comparable to daytime.
- Support staff went from two overwhelmed to one focused - the second redeployed to sales follow-ups, which the owner considers the project's best ROI.
- Escalations work: customers reaching humans arrive with history attached; the "talking to a wall" complaint pattern never materialized.
The limits, honestly: the bot still occasionally misreads ambiguous phrasing and hands off conversations a human might have salvaged; the knowledge base needs updating whenever prices or policies change (that discipline is part of our AI maintenance service); and the 70% figure reflects this business's highly repetitive question mix - yours may differ.
Want to know your automation ceiling? Send us a sample of your actual customer questions - we will tell you honestly what share a grounded bot could handle. Start here.
