Telecom · RetailB2CTop-5 Telecom RetailNDA

An AI Chatbot Trained on Veteran Salespeople — Purchase Conversion Up 300%

A chatbot trained on the tone and manner of veteran salespeople. Purchase conversion up 300% over the previous Kanana AI bot, absorbing the workload of 10+ CS and sales staff.

Top-5 telecom retailer in Korea · sales-trained AI chatbot (NDA · anonymized)

After launch, here’s what changed
+300%
purchase conversion (vs. previous Kanana AI bot)
10+
CS & sales roles absorbed
24/7
uninterrupted chat coverage
Sales AI Chatbot for a Top-5 Telecom Retailer — Top-5 telecom retailer in Korea · sales-trained AI chatbot (NDA · anonymized)
Before

Where things stood
before the project

  • 01The existing AI chatbot stopped at one-line answers and rarely carried conversations to an actual purchase.
  • 02CS and sales staff couldn’t cover late-night and weekend inquiries, so revenue leaked through the same time slots every single day.
  • 03Contract terms, device subsidies, plan-discount options, and gift combinations differed every time — and handset market prices moved constantly — so the old bot’s answers frequently diverged from actual pricing.
  • 04Veteran salespeople’s techniques never transferred to new hires, leaving wide gaps in response quality.
Approach

How we solved it

“How would a veteran answer this, right here?” — that became the bot’s starting point. We fine-tuned on veteran call recordings and chat logs to replicate their tone, pacing, and deal-closing style exactly, then layered on a quote RAG that indexes plans, contracts, and device subsidies across Korea’s three major carriers in real time. A purchase-intent classifier automatically tells “just asking” apart from “about to buy” and switches the response strategy, escalating only pre-payment and complex cases to human agents. The structure never loses trust — and brings people in only at the moments they’re needed most.

Inside

How it works on the inside

Work that one person used to handle is split across multiple agents. The automation runs across the entire flow — not just a few steps.

01 · Agent

Sales Tone & Manner Learner

Fine-tuned on veteran agents’ call recordings and chat logs — replicates their friendly, zero-pressure consultative style exactly

02 · Agent

Quote RAG

Real-time Vector DB indexing of the three major carriers’ plans, contracts, device subsidies, and plan discounts — price changes reflected automatically

03 · Agent

Purchase-Intent Classifier

Auto-detects “just asking” vs. “about to buy” → pushes near-purchase customers to the payment page

04 · Agent

5 Quick Actions

“Recommend a phone”, “lowest price”, “business phones”, “upgrade vs. carrier-switch comparison”, and “gift check” entry points

05 · Agent

Comparison Quote Cards

Simultaneous quotes for the same model across all three carriers + instant monthly-payment simulation

06 · Agent

HITL hand-off

Payment-stage and complex cases auto-escalate to human agents — a boundary designed to never lose trust

07 · Agent

Live Inventory & Promotion Sync

Internal ERP API integration surfaces sold-out stock and limited promotions instantly

08 · Agent

Conversation Quality Monitoring

Dashboard for response satisfaction, drop-off points, and the purchase-conversion funnel

After

What changed
after launch

  • +300% purchase conversion vs. the previous AI chatbot (90-day cumulative average).
  • The bot absorbed the workload of 10+ CS and sales staff — people now focus solely on handed-off payment and complex cases.
  • 5.2-second average first response, 24/7 with zero downtime — late-night and weekend conversion rates now approach business-hours levels.
  • Comparison quote cards and purchase-intent scores flow straight into sales meetings as input for designing the next promotion.
Runtime

Project details

Duration
3 weeks design + 8 weeks build + 4 weeks fine-tuning & evaluation
Team
TreeSoop team of 5 (1 PM / 2 AI engineers / 1 backend / 1 prompt engineer)
Operating mode
5-channel integration (web · Kakao · Naver TalkTalk · in-app · callback) · automatic HITL hand-off
Stack

Tech stack

STACK
LangChain · DeepAgents · FastAPI
RAG
Quote data · veteran call logs
STATUS
Live B2C · Top-5 telecom retail

Trained directly on veteran salespeople’s tone, pacing, and deal-closing style. It converts the late-night and weekend inquiries humans can’t take — 3x the purchase conversion of the previous AI bot, absorbing the work of 10+ full-time staff.

More

Other work in a similar vein

Facing a similar challenge?
Let’s talk it through

Excel workflows, chatbots, AI marketing — whatever the domain, it can be redesigned the same way.The initial assessment and quote are free — the contact form is the fastest way to reach us.