AI Agent was first implemented and scaled in e-commerce scenarios

We read the report "Agentic Shoppers Are Coming…

Who Could Win or Lose?" released by Morgan Stanley on November 17 and gained some confirmation: The first to benefit from AI will undoubtedly be the giants, making the strong even stronger. The diffusion of AI will further widen the disparities within industries, making the Matthew effect more pronounced.

Here are some ongoing cases that analysts have already observed:

1. AI Agent from "Tool" to "Colleague"
Problem: Traditional AI can only handle single-point tasks (recommendations, customer service) and cannot cover the entire transaction chain.
Solution: Turn LLM into an "Agent Middle Platform" capable of calling 20+ internal system APIs, allowing the model to break down goals, adjust tools, and make decisions on its own.
Cases:
- Taobao's "Wanxiangtai" Agent autonomously generated 120 million graphic materials during Double 11, with ROI increasing by 38% compared to manual groups.
- JD.com's "Galaxy" Agent reduced the cold start period for new products from 14 days to 3 days, with single-item traffic increasing fourfold on average.

2. Let Agents Help Buyers "Screen" Global Source Goods
Case: 1688's "Source Selection Agent"
- Problem: Small and medium-sized store owners don't know how to choose factory-source goods, and 1688 has 20 million SKUs of noise daily.
- Solution: The Agent crawls factory production capacity, Taobao downstream sales, and TikTok trends in real-time, recommending "ship within 48 hours + guaranteed downstream popularity" combinations to store owners.
- Effect: Within 30 days of launch, a 10-person store in Dongguan placed a single order of 800,000 yuan, with inventory turnover days decreasing by 60%.

3. AI Agent Adjusts Prices 196 Times a Day, Still Cheaper Than Humans
Case: SHEIN's "Dynamic Pricing Agent"
- Problem: Simultaneous promotions in 150 countries globally, with exchange rates, tariffs, inventory, and competitor prices changing by the second.
- Solution: The Agent scrapes 50 competitor sites and 30 exchange rate interfaces every 5 minutes, using a reinforcement learning model to automatically bid under the hard constraint of "gross margin ≥18%".
- Effect: The same dress sold for $28 on other sites, while SHEIN sold it for $23 with a 22% gross margin; Q2 financial report showed a 3.4 percentage point year-on-year increase in gross margin.

4. Supply Chain: Writing Typhoons, Strikes, and Public Sentiment into the Model
Case: Cainiao's "Logistics Risk Agent"
- Problem: Last year's typhoon "Plum Blossom" delayed 3 million orders in East China, costing the platform 110 million yuan in compensation.
- Solution: The Agent connects to the meteorological bureau, port Twitter, and Weibo sentiment, marking high-risk routes in red 72 hours in advance and automatically relocating stock to western warehouses.
- Effect: During this year's typhoon "Hai Kui," the order fulfillment rate in affected areas remained at 96%, with only 9 million yuan in compensation, saving 92%.

5. Customer Service: Agent Directly Approves Refunds Without Manual Review
Case: Douyin E-commerce's "After-Sales Decision Agent"
- Problem: The return rate for clothing is 55%, manual review costs are high, and buyers wait 2 days for refunds.
- Solution: The Agent obtains return logistics records, image recognition, and user credit scores, automatically approving refunds and triggering courier pickups when criteria are met.
- Effect: Refund time reduced from 48h to 11min, platform dispute rate dropped by 27%, and merchant capital turnover sped up by 1 day.

6. Marketing Copy: Turning "Li Jiaqi's Tone" into Adjustable Parameters
Case: Pinduoduo's "Spokesperson Agent"
- Problem: Directly sourced fruits need "down-to-earth" language, but small merchants can't write it.
- Solution: The Agent has 12 built-in sales styles (Li Jiaqi-style, rural-style, Gen Z slang), generating 100 short video scripts in 3 seconds from an SKU number, automatically matching subtitles and voiceovers.
- Effect: A Yunnan pomegranate merchant sold 1.2 million jin in 7 days, with a script completion rate of 42%, 18 percentage points above the platform average.

7. Live Streaming: 24-Hour "Digital Human" Without Scandals
Case: Intime Department Store's "Sales Assistant Digital Clone"
- Problem: No live streaming during off-hours for brand counter guides, wasting traffic.
- Solution: A 3-minute recording of a real person's speech sample generates a 1:1 digital human, automatically explaining 300 inventory items and answering bullet comments in real-time.
- Effect: Midnight (0-6 AM) sales reached 3.8 million yuan, accounting for 14% of daily sales, with user repurchase rates matching real-person live streams.

8. Compliance: Review Before Broadcast, Agent Kills "Extreme Words" on the Spot
Case: Xiaohongshu's "Live Compliance Agent"
- Problem: Hosts mistakenly saying "cheapest" or "absolutely effective" resulted in 500,000 yuan fines.
- Solution: The Agent converts live audio to text in 300ms, popping up alerts in the admin backend for banned words under advertising law, with a 1-second mute option.
- Effect: Within 3 months of launch, platform administrative penalty cases dropped by 76%, and host complaint rates fell by 41%.

9. Organizational Change: The Emergence of "AI Operations" Roles—How to Set KPIs?
Solution:
- Alibaba upgraded its original "Express Driver" to "Agent Trainer," changing KPIs from "ROI" to "AI adoption rate + human-machine hybrid ROI."
- JD Logistics added an "Agent Supervisor" role to feed abnormal scenario data to Agents, with performance tied to Agent-saved work hours.
Effect:
- Within six months, 3,000 Alibaba operations staff were certified, with an 85% AI adoption rate and a fivefold increase in per capita budget management; JD's per-order labor costs dropped by 22%.

AI's tremendous value in improving efficiency can be perfectly applied in trade and service sectors, but its application in production will take time.

Long-term bullish on service and trade industry giants.