Ecommerce Chatbot Case Study: 35% Self-Serve and 17% AOV Lift

Starting point

Support was swamped with order status, returns, and sizing questions. Sales wanted guided selling to recommend bundles. The brand needed guardrails to keep the bot on-policy and escalate cleanly.

Outcomes after six weeks

• 35% of tickets fully self-served (order status, returns, product FAQs).
• 17% lift in AOV via guided selling nudges.
• 21% faster resolution time for escalated tickets thanks to better triage.
• CSAT held steady; no policy violations.

Architecture

Key design choices

• Scope first: Limit to the top 12 intents (order status, returns, sizing, shipping, discounts, materials).
• Guided selling: If the user is browsing, the bot proposes bundles and higher-margin alternatives; if in support mode, it stays utility-first.
• Retrieval-only responses: Bot answers only from approved sources; low confidence triggers escalation.
• Clean handoff: Transcript + intent + suggested macro sent to Zendesk; agents see what the bot tried.

Implementation steps

1) Knowledge and retrieval

We ingested product catalog, policy docs, and FAQ into a vector database with metadata (SKU, collection, policy type). We chunked content and tuned embeddings to reduce drift.

2) Orchestration

We used LangChain flows with tools: order lookup, return initiation, shipping status, and product recommend. Guardrails refuse anything outside scope or involving sensitive topics.

3) Guided selling flows

When the user is shopping, the bot asks 2–3 preference questions, then suggests bundles or higher-margin alternatives with links. If the user switches to support mode, it stops selling and solves the issue.

4) Escalation

Low confidence, negative sentiment, or policy topics trigger escalation. Transcript, detected intent, and context are attached to the Zendesk ticket to shorten handle time.

Rollout timeline

• Week 1: Intent selection, knowledge curation, guardrails.
• Week 2: Build flows, Shopify + Zendesk integration, pilot on web.
• Week 3: Launch WhatsApp; add guided selling nudges; monitor deflection.
• Week 4: Tune prompts and guardrails; finalize reporting.

Lessons learned

• Keep the bot scoped; policy topics escalate immediately.
• Retrieval quality beats clever prompts; invest in clean sources.
• Guided selling works when it’s contextual and optional.
• Share transcripts; agents move faster when they see prior attempts.

If you want similar results

Start with support intents that dominate volume, add retrieval over your policies and catalog, and ship guardrails before cleverness. Layer guided selling after support is stable.