Chat Logs to Conversions: Automating Pre‑Enrollment with a Messaging‑Trained Bot

We started with a careful audit of the WhatsApp export, mapping the recurring intents that drive nearly all inquiries: course packages, pricing rules, lesson counts, exam attempts, rescheduling policies, branches and timings, manual vs. automatic transmissions, and payment/refund scenarios. Before any modeling, we anonymized personally identifiable information and reconstructed threads so that answers would retain context without exposing private data. From there, we transformed the archive into a foundation for Retrieval‑Augmented Generation (RAG): concise, well‑scoped snippets designed to be pulled into an answer only when they truly matched the question at hand.

To make retrieval precise, we added a pragmatic layer of labels—just enough structure to help the system understand what a prospect is really asking, without turning the project into an over‑engineered taxonomy. Those labeled snippets were embedded and indexed in a vector store. On top, we orchestrated a chat flow that always leads with grounded information, asks clarifying questions when key entities are missing (“Which branch works for you?” “Manual or automatic?” “What dates are you targeting?”), and defers to a human when confidence is low or the topic sits outside policy.

We built the assistant into the school’s website with a conversational UX that mirrors the brand: warm, concise, and helpful. Answers are short by default, with optional detail if a prospect wants it. When a next step is obvious—book a slot, view a policy, request a call—the bot offers it without fuss. Every interaction is logged behind privacy controls so we can learn from what actually happens, not what we imagined might happen.

After launch, we ran a tight improvement loop. The school flagged both spot‑on and off‑target replies. Correct answers became positive teaching examples; the misses revealed where the index needed new snippets or where prompts should steer the model to ask for clarification sooner. We also assembled a small, realistic test set covering the top intents and a few tricky edge cases. That set gives us a baseline to check progress whenever the school updates policies or we refine prompts.

The outcome is exactly what the client wanted: most repetitive pre‑enrollment questions now receive immediate, consistent, policy‑true answers—day or night. Prospects move faster because they get the right information the first time. The director’s ad‑hoc consultations drop to the genuinely complex or high‑touch situations that deserve human attention. And because the source of truth is the school’s own labeled archive, keeping the assistant current is a matter of editing snippets rather than retraining a model from scratch. No brand dilution, no guesswork—just a reliable admissions assistant built from the conversations that made the school successful in the first place.