Retail Chain Private AI Customer Service and Store Network Upgrade

Challenge

1,800 stores ran on a single broadband line and POS / loyalty systems dropped at peak; 400 contact-centre agents searched 30+ policy and product documents by hand with a 6-minute average handling time; member data and conversations could not leave the company, ruling out public-cloud LLM APIs.

GSS scope of responsibility

GSS owned store dual-link SD-WAN, HQ inference-cluster and RAG platform infrastructure, nationwide field engineering and 24/7 operations. Knowledge-base content, service scripts and model evaluation criteria were owned by the customer service and IT teams.

Solution

GSS delivered dual-link SD-WAN (broadband + 4G/5G backup) with centralised security policy to every store, deployed an on-premises LLM inference cluster and enterprise knowledge RAG platform at HQ, launched an AI agent assistant and a store-staff Q&A bot, and provides nationwide field engineering plus 7×24 managed operations.

Background

The retailer operates 1,800 stores nationwide and a 400-seat contact centre. Stores connected to HQ over a single local broadband line, and POS / loyalty drop-outs during promotions were the top store complaint. Agents searched more than 30 policy, promotion and product documents to answer each contact, averaging six minutes, and new hires needed four weeks of training.

The company wanted large language models to lift service productivity but required that member information, transactions and conversations never leave its environment — so public-cloud LLM APIs were off the table.

Solution

Store network (traditional network business)

• Dual-link SD-WAN at all 1,800 stores: primary local broadband plus 4G/5G backup, with POS and loyalty traffic prioritised and millisecond failover when the primary degrades.
• Centralised security policy: uniform store egress rules, isolated guest Wi-Fi, separate segments for IoT (electronic shelf labels, cameras) — pushed once from HQ and applied nationwide.
• Nationwide roll-out by GSS field engineers in five regional waves at 40–60 stores per day, with zero trading downtime (work before opening / after closing).
• Operations: all stores monitored 7×24 by the GSS NOC; failed devices replaced in-store within 8 hours by the nearest field engineer.

Private AI customer service (AI business)

• Inference cluster: a GPU cluster in the company’s HQ data centre running an open-source LLM fine-tuned on company data; all inference stays on the corporate network.
• Enterprise knowledge RAG: 30+ policy documents, promotion rules, the product catalogue and historical tickets cleaned into a vector knowledge base; every answer cites its source so agents can verify in one click.
• Agent assist: listens to the live conversation, retrieves knowledge and pushes suggested replies and procedures; member identity and order lookups run through internal system APIs.
• Store-staff Q&A bot on WeCom for promotion rules, returns processes and system how-tos, replacing phone calls to regional supervisors.
• Quality loop: agent acceptance / edits of AI suggestions are logged for monthly model and knowledge refreshes; sensitive topics (complaints, legal) escalate automatically to humans.

Results

Metric	Before	After
Store outage events	~2,300 / month	~200 / month (−91%)
POS availability at promotion peaks	~97%	99.9%
Average handling time	6 min	2.5 min
AI first-contact resolution	—	68%
New-agent training	4 weeks	1.5 weeks
Store calls to supervisors	~900 / day	~250 / day
Data leaving the company	—	Zero

Fully live in January 2025; GSS continues managed operations for the store network and model-iteration support for the AI platform.