Retail & CG

Support ticket deflection rate achieved
CSAT score maintained (out of 5), unchanged from pre-launch baseline
₹16 Cr
Five-year NPV unlocked from support cost reduction
About
An e-commerce marketplace deflected 38% of support tickets with a RAG AI agent, kept CSAT at 4.4 out of 5 and gained a ₹16 crore five-year NPV.
Industry
Retail & CG
Company size
5,000 – 10,000 employees
Founded
2017
Business Context
Business context and operating model
An e-commerce marketplace scaling faster than its support desk The client runs one of the region's largest e-commerce marketplaces, spanning retail, grocery, and quick commerce across a fast-growing customer base. Support ticket volume rose roughly 40% year over year as the platform scaled, driven by order tracking, returns, and payment queries across a widening product and seller base. Support was staffed and scripted the traditional way — tiered agent teams working from FAQ scripts — which meant answer quality and tone varied by whoever picked up the ticket, and OPEX scaled roughly linearly with ticket volume rather than sub-linearly.
Problem Definition & Framework
Current State
Failure Modes
Decision Criteria
KPI Baseline
Problem definition and framework
Scripted answers could not keep pace with rising volume With ticket volume outpacing headcount growth, average resolution times were creeping upward and first-response quality was inconsistent — the same question could get a materially different answer depending on which agent, and which script version, handled it. Complex or high-sentiment tickets — an angry customer over a failed refund, a seller dispute — needed to be identified and escalated quickly, but the triage process itself relied on agents self-flagging rather than any systematic sentiment detection, so escalations were frequently late.
Decision framed: Which customer issues can be resolved automatically, which require clarification, and which need immediate human escalation with full context?
Framework used: RAG answer quality, sentiment triage, and escalation taxonomy
Baseline mapped: current process, failure modes, handoffs, data owners, and decision latency.
Success criteria tied to business KPIs, adoption behavior, and operating cadence changes.
How it works
Before / After

Architecture

Process Flow

Data Pipeline

Results

Delivery Timeline

Engagement Method & Process Diagram
01 Discovery
02 Data Audit
03 Baseline
04 Prototype
05 Pilot
06 Production
07 KPI Monitoring
Engagement method and process diagram
Process diagram: Support KB + policy docs + historical tickets -> retrieval index -> GPT-4 response agent -> sentiment and ambiguity triage -> human handoff with context -> CSAT and deflection monitoring
Engagement method: Ticket taxonomy; knowledge-base readiness audit; retrieval evaluation; guarded response design; escalation testing; phased deployment by support category.
Delivery pattern: discovery, data audit, baseline measurement, prototype, pilot workflow, production rollout, and KPI monitoring.
A RAG pipeline over 60,000 knowledge-base documents We built a retrieval-augmented generation pipeline over the client's full support knowledge base, more than 60,000 documents spanning policy, product, and seller-operations content, powering a GPT-4 agent that answers routine queries directly, cites the underlying policy it drew on, and asks a clarifying follow-up rather than guessing when a query is ambiguous. A sentiment-triage layer runs on every incoming ticket, scoring urgency and emotional tone so high-risk cases are escalated to a human agent immediately, with the full conversation context and a suggested resolution path already attached rather than starting the handoff from scratch. The whole pipeline is instrumented with Prometheus for real-time monitoring of deflection rate, response latency, and escalation accuracy.
Architecture, Results & Measurement
Solution architecture, results, and measurement
38% ticket deflection with CSAT held steady Ticket deflection reached 38% — more than a third of all incoming support volume now resolves without a human agent touching it — while CSAT held at 4.4, meaning the deflection did not come at the cost of customer satisfaction. The combination unlocked an estimated ₹16 crore in five-year NPV through reduced support headcount growth relative to ticket volume growth. Sentiment-based triage meant high-risk tickets reached a human agent with full context and urgency already flagged, cutting escalation lag and giving agents a materially better starting point on the cases that most needed a human touch.
Measured outcome: 38%: Support ticket deflection rate achieved
Measured outcome: 4.4: CSAT score maintained (out of 5), unchanged from pre-launch baseline
Measured outcome: ₹16 Cr: Five-year NPV unlocked from support cost reduction
Measurement loop: compare weekly operating behavior against baseline, surface exceptions, and review model/workflow drift with business owners.

The agent didn't just cut our costs — it made our escalations better, because the tickets that do reach a human now arrive with the full picture already attached.
Head of Customer Experience, e-commerce marketplace
KEEP READING
Explore more studies on AI, analytics, and enterprise intelligence.











