---
title: "Deep Dive | The AI Revolution in E-Commerce — A 10,000-Word Read"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2023-10-27T16:30:00+08:00"
canonical: "https://ffcap.cn/en/research/20231027a06f9v00"
source: "https://ffcap.cn/en/research/20231027a06f9v00"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20231027A06F9V00"
---

# Deep Dive | The AI Revolution in E-Commerce — A 10,000-Word Read

_Unique Research · Insight / deep dive · October 27, 2023 · Full editorial rewrite of ~14,482 CJK — all major sections, company profiles, founder quotes, and figures preserved (not a stub)_

**Contents:** 01 E-commerce SaaS market size · 02 E-commerce & generative AI trends · 03 AI + e-commerce: the next retail form · 04 Representative genAI + e-commerce go-global companies · 05 Outlook.

## Preface

Over the past decade-plus, AI progressed from Transformer, GPT, BERT, and GAN into real application-layer commercialization. As infrastructure, algorithms, and pretrained models matured, the bar for SaaS tools and AI-ified workflows fell. Vendors can deliver e-commerce services cheaper, faster, and better — while China’s shift from growth to stock markets forces teams to cut cost and raise efficiency. Replacing low-yield workflows with better tools is an obvious move.

Generative AI and its middle-layer apps are reshaping the e-commerce value chain. Firms that adopt deeply will displace peers that treat AI as a veneer. Unique Research deep-interviewed five companies commercializing genAI in vertical e-commerce scenes — **FOSHO, Kua.ai (Kuahai Tech), Moyin Intelligence, Unigravity (Wanyou Yinli), and Yunxiang Wisdom** — plus perspectives from GemX.ai — to explain why these players moved first.

## 01\. E-commerce SaaS market size

### China enterprise SaaS

Enterprise SaaS kept high double-digit growth through 2021. iResearch (2022): 2021 China enterprise SaaS ≈ **RMB 72.8B**. 2022 pandemic softened demand; under a base case, 2024 could reach **RMB 120.1B** (2021–2024 CAGR ≈ **18.1%**).

### China e-commerce SaaS

Vertical SaaS’s share of the pie rose for five years to **47%** in 2021; iResearch sees ~**48%** by 2024. That implies 2021 vertical SaaS ≈ **RMB 34.2B**, and ~**RMB 57.6B** by 2024. Within vertical SaaS, retail e-commerce (incl. cross-border) is the largest slice — **36%** in 2021 vs AEC’s 19%. Implied 2021 e-commerce SaaS (incl. CBEC) ≈ **RMB 12.3B**; 2024 ≈ **RMB 20.7B**.

### E-commerce services industry

Buyers judge SaaS by incremental GMV and process cost reduction. iiMedia: China e-commerce services industry ≈ **RMB 366.3B** in 2022 vs **RMB 109.5B** in 2017 — CAGR ≈ **27.3%**.

## 02\. E-commerce industry & generative AI trends

### Five e-commerce SaaS trends

Market: ~RMB 6B (2017) → ~12B (2021) → >20B expected (2024). Trends Unique Research sees:

-   **Cloudification** — mature IaaS lowers deploy cost and raises reliability.
-   **O2O** — online + offline fusion needs tighter SaaS for precision marketing and ops.
-   **Multi-platform ops** — Douyin, Kuaishou, Pinduoduo, Dewu, Xiaohongshu raise channel and rule complexity.
-   **Multi-function suites** — sales, CRM, payments, analytics on shared data.
-   **Cross-border** — beyond Amazon/DTC sites, TikTok and Kwai diversify formats; multilingual, logistics, payments, ads, customs become table stakes.

### What generative AI is — and how we got here

GenAI produces complex creative content (image, music, video, text) via deep models — autoregressive, GAN, VAE — pretrained unsupervised then fine-tuned. Landmark models: GPT, BERT, CLIP, DALL·E.

Timeline: Fei-Fei Li’s ImageNet (from 2007) and ILSVRC (2010) seeded deep vision; AlphaGo beat Lee Sedol (2016); Google’s Transformer (2017) beat CNN/RNN on capability and speed and spawned BERT, GPT, XLNet. 2018 cluster: OpenAI GPT-1, Google BERT, NVIDIA StyleGAN. Then GPT-3 API (2021), ChatGPT on GPT-3.5 (2022) — and a flood of apps for art, video, code, and writing.

## 03\. AI + e-commerce: the next form of retail

### GenAI application scale & SaaS penetration

Forward-looking / Zhongguancun Big Data Alliance: China genAI application scale to **RMB 207B** by 2025 (2020–2025 CAGR ≈ **84%**).

**Gong Yi** (Kua.ai CEO): most go-global e-commerce teams are still in small experiments — ChatGPT for SEO/blogs, Midjourney for PDP images. Single-function AI tools may touch **50–60%** of the market; end-to-end AI-optimized workflows <**5%**.

**Li Jiayi** (FOSHO CEO): ~**80%** of go-global marketers have tried AI tools, but mostly shallow features. Affiliate marketing is a small share of marketing spend — AI that lowers the affiliate ops bar can unlock adoption.

**Sha Tao** (Yunxiang): penetration hinges on business contribution and compliance/IP. SMBs adopt when tech lifts results; large players move slower on compliance. Cross-border AI penetration exceeds domestic — partly lighter policy friction. Demand is rigid; tech is iterating fast.

### Impact by scene

**Smart customer service (Moyin / Yang Kai):** Pre-GPT stacks used hand-built Q&A + Elasticsearch — hundreds to thousands of pairs, stiff answers; buyers demand humans when they smell a bot; shops only enable bots off-peak or for simple questions. Embedding product knowledge bases + GPT yields accurate answers in “gold CS” tone with context and emotion — robots handle most pre-sales; few elite agents take exceptions.

**Product design (Unigravity / Zhong Wei):** AI speeds designers (mass options → filter/tune) and lowers the bar for buyers, ops, stall couples, small factory owners. Classic path Stretch → Photoshop → Blender was too steep. Agents (e.g. AutoGPT chaining tools) can close loops from banner design → click tests → analytics. AI will sit in new-SKU launch, hero push, and assortment adjusts.

**Content generation (Sha Tao):** Content/social commerce makes quality and emotional value as important as product function. GenAI lands first on: ops-side video/image/copy cost; marketing on Xiaohongshu/Douyin; design-side new SKUs (especially apparel).

### Drivers and blockers

**Qin Shuai** (GemX.ai): disruptive features (e.g. AI video still early), UX friendliness, and ROI drive B adoption; in go-global e-commerce, B payment habit is less of a blocker.

**Gong Yi — drivers:** AI’s orchestration flexibility shortens workflow build and validation vs prior digitization waves; cross-border firms are pragmatic with less legacy debt.  
**Blockers:** undefined B standards (unclear selling points, keywords, voice → random gen → random conversion); high bar for scene engineering/data/software — without them you only solve point problems and ROI looks weak.

**Li Jiayi:** affiliate SaaS growth is blocked by weak domestic cognition (traffic concentration killed independent-site affiliate culture) — needs simple one-stop intelligent solutions + best practices.

**Sha Tao blockers:** compliance (foreign foundation models vs regulation), unclear IP of AI-generated content (large firms wait), and controllability (models must connect to real SKUs; tuning is hard).

### Do foundation models create instability?

**Smart CS:** Yang Kai wants near-GPT-4 quality plus Claude-class long-summary tools; apps must talk to programs/APIs with stable outputs. GPT-4 still immature for full production; domestic models lag; whether widely used LLMs open-source will shape the industry.

**Affiliate:** FOSHO leans on exclusive knowledge graphs — less exposed than thin wrappers on public data for copy.

**Cross-border:** Gong Yi — content for foreign consumers, production often offshore → lower geo risk; domestic GPT-3.5-class models will appear; private deploy keeps data local for China scenes.

**VOC / GemX:** open-source proliferation reduces model geopolitics as the main risk; closed models may be pressured by open ecosystems (unless geopolitics hits open source itself).

### Understanding LLM call cost

Gong Yi: Moore-like dynamics — each major model upgrade can cut same-volume generation cost by ~**10×**. GPT-3.5 Turbo ≈ **$0.0002 / 1K tokens** (~0.75 words/token) — prompts returning a ~700-word article. Image gen: fractions of a cent to a few cents per word of prompt. A non-native English writer for a 700-word SEO piece might cost ~$20; a U.S. native writer ~$300. AI can be thousands of times cheaper and faster.

### Competition, moats, customers

**Yang Kai:** future competition in narrow scenes may be less brutal — many AI app companies creating distinct value customers will pay for.

**Sha Tao:** first crowded tracks use general LLMs + light prompts for text/dialogue cost cuts (assistants, copy, titles) — painful demand + mature tech. Multimodal (video/image) with owned pretrained models or domain RLHF is less occupied — more opportunity, higher tech bar (controllable SKU video/image; public multimodal marketing content).

New competitors may not be old peers: strong-tech/weak-industry players, or strong-industry players who learn AI fast.

**Moats:** Yang — tech, cost control, first-mover (even “simple” ideas take long to productionize). Gong — clear boundaries, great UX, cost discipline (GPT plans workflows + cheaper tools; don’t bill customers for unused frontier model calls); vertical pretrain cost/efficiency for strong tech teams. Zhong — industry cognition (ops logic, data, design paradigms, feedback loops) + three tech layers: control models (e.g. garment–pose consistency), toolchains (multi-model pipelines, e.g. photo → realistic LEGO), function/mid-platform (stable callable APIs/UI). Qin — market-data-to-design: analytics + product-dev, now with design images not just charts. Li — LLM + exclusive affiliate knowledge graph for creator match and personalized creative.

**Customers:** complex pre-sales categories (beauty, 3C, home) need bots more than snacks/apparel; replacing humans beats vague “copilot efficiency.” Sha: cost-cutters (ops agencies, CBEC, domestic agencies) want image/video/copy/multilingual CS now; sales-lifters care about human-or-better conversion ROI. Pre-LLM martech prioritized revenue then cost then efficiency; post-LLM, SMBs chase cost first, large buyers chase compliance + revenue. Gong: almost all CBEC pros have tried AI; small sellers solve point problems, large sellers want closed loops (selection, supply, R&D, content) and fear AI-armed rivals. Zhong: design talent scarce/expensive near factories; sellers want cheap model shoots; large sellers (>100 people) chase efficiency, SMBs chase revenue; generation must match physical product detail. Li: brands at $100–200M GMV or new go-global have thin affiliate teams; global brands already have affiliate teams/agencies and buy growth/cost cuts — CBEC benefits first from AI because of multi-scene, multi-country, multi-culture content demand.

## 04\. Representative genAI + e-commerce go-global companies

### 1\. FOSHO — AI affiliate marketing for brand globalization

Team DNA: global brands, top internet firms, international 4As — marketing + tech + localization know-how. Stack: big data + proprietary AI engine + global channel ecosystem covering insight → decision → transaction → management → optimization.

**FOSHO AFF AI Alliance Cloud** — first AI-driven one-stop affiliate cloud linking global affiliates, creators, partners:

-   **Intelligence** — strategy brain on a proprietary ~100B-scale global affiliate knowledge graph; 400+ tag dimensions; search/match **1.5M+** affiliate channels; live competitor strategy intel.
-   **Marketplace** (beta) — reach active affiliates, task publish, AI creatives, order tracking, auto commissions.
-   **Manager** — multi-platform auth, multi-country/project dashboards, commission forensics, suspicious-order flags.

Tech: enterprise NER for the graph; style-transfer AIGC (OSII) trained on graph creatives + conversion outcomes; AAL + LLM “AI Commander Engine” that learns product SOPs and automates multi-step workflows via RPA.

### 2\. Kua.ai (Kuahai Tech) — cross-border AIGC content

Team from Amazon, Tencent, Nielsen, CMU, Sun Yat-sen; partners with IDEA Institute, Narrators, SYSU, Amazon, Meta. On GPT-4 + AIGC for CBEC sellers and go-global brands:

1.  Marketplace ops — selection, VOC, keywords, listing create/optimize (text+image).
2.  Independent-site bootstrap — brand/domain ideas, meta, banners.
3.  Daily ops — listing gen/publish, insights, SEO posts+images, social.
4.  Image tools — text-to-image, white background, background swap, variants, contour, AI retouch.

Serves marketplace sellers, DTC sites, and export factories across marketing scenes.

### 3\. Moyin Intelligence — knowledge base + smart CS

Enterprise knowledge base and smart CS for mid/large SaaS, IT, retail, finance. ChatGPT-class tech on company knowledge — no heavy corpus config — so staff and customers query product knowledge fast. Clients include Moka, Beisen, Fadada, Shanghai Seer, Yio Education, Joyvio, Tezign, pingpong, and others across internet, retail, biopharma, manufacturing, e-commerce, education.

### 4\. Unigravity (Wanyou Yinli) — AIGC for jewelry & apparel productivity

AIGC production + ops for CBEC, sellers, manufacturers: trend+sales → creative new SKU models; remodel hero SKUs; multi-angle human display + scene fusion; virtual models and outfit scenes.

### 5\. Yunxiang Wisdom — controllable, knowledge-augmented models for fashion retail

Founded 2018; core from EF, Ctrip. Product 1: image platform — garment → garment model; person photo/desc → model; scene photo/desc → scene; fuse into on-model product shots for e-comm/social. Product 2: LLM search on owned storefronts (mini-program/app/web) — smart search, guide, recommend, CS. Moat: knowledge-graph-enhanced base + years of fashion retail delivery. Buyers: anyone needing model photography; brands with owned e-comm.

### 6\. GemX / GenMatrix (Jingmai) — data intelligence for global product innovation

From demand insight to product innovation: continuous scrape/clean of global e-comm + social; multimodal AI for market/product/consumer insight.

-   **Market insight** — Amazon et al. trends, TAM, competition, new-SKU opportunities.
-   **Product insight** — competitor ops, history, traffic entry; CV for patent screening.
-   **Consumer insight** — NLP on reviews: sentiment, selling points, motives, pros/cons, expectations, scenes → personas.
-   **Product design** — auto opportunity points → genAI concept drafts.

Customers: factories, brands, traders, CBEC sellers innovating for overseas. Delivery: SaaS + KA private deploy. ~1 year live: **20k+** SaaS registrants, **2k+** paid (RMB 1,999/yr → raising to 3,999); **6** KA clients + 10+ in POC (RMB 200k–1M/yr).

## 05\. Outlook

### Scene opportunities

**Smart CS:** Yang — realism now near human; best fit for SKUs priced hundreds to ~RMB 1,000; higher AOV / in-store scenes (e.g. cars) harder to replace.

**Cross-border:** Gong — agencies that don’t embrace AI risk obsolescence. Pricing shifts toward usage/outcome-based — more use, more value created.

**Affiliate / brand globalization (Li) — five stages:** (1) 2010–16 dump/supply-driven CBEC+trade; (2) 2016–17 pricing-power manufacturers (esp. 3C); (3) 2017–20 DTC/owned sites & apps; (4) 2020–21 COVID acceleration; (5) 2022+ brand globalization, multi-channel precision as developed markets leave platform-dividend era.

**Product design (Zhong):** shelf logic → interest/match logic (virtual humans, live). People–goods–place keep getting rebuilt; live commerce elevates “place.” Consumers less brand-loyal in the old sense (China domestic brands; global cultural affinity). AI design → click-test loops → production can cut inventory for non-standard goods (apparel, home, toys) — possibly one of e-commerce’s biggest shifts this decade. Marketing tools today are still shallow “shells” for creatives+ads; deeper mid-platform rebuilds matter.

### Industry trajectory

Yang: LLM gains won’t compound forever — tech dividend → growth → commercialization → maturity; high odds AI replaces repetitive labor. Gong: in 3–5 years (or less), >**50%** of high-quality keyword articles AI-written/assisted with >**90%** process automation; info overload may birth B2C AI filters; AI assistants may unify marketplace + DTC + brand sites for discovery; near-term selection/VOC/keyword tools need real data to fight hallucination. Zhong: platforms never fully fixed people–goods–place match — AI’s opening; design-from-demand at the front end creates incremental revenue via personalized assortments. Foundation models without core industry apps are unsustainable — partner with industry for 2B apps in the next six months; reshape mid-platforms (process + data + people–goods–place), not just tools; AGI-style general models still need specialized training per vertical.

### Early-company opportunity

Sha: medium-term ceiling = conversion at or above human levels. China lagged on foundation models but can catch up in apps. Chip/model layers have opportunities but suit startups poorly; real startup opportunity is **application**, especially scenes AI can transform. Prefer fragmented industries with lots of public knowledge and many potential customers (apparel, e-commerce, F&B, fashion, industrial energy) over concentrated, already-AI-armed verticals (finance, insurance, auto, call centers). Li: vertical app startups win on deep scene cognition + private/public data fit; incumbents have products but less flexibility; model labs win on base tech but not first-party scene data. Zhong: GPT-3.5/4 is a dimensional strike on prior solutions — startups back on the same starting line; problem understanding beats legacy matching logic that compounds delay. Qin: AI-native startups should attack pains traditional software never solved — and win on price/performance where features overlap.

_21 CN figures (market charts, product screens, company materials) — attach on import._

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20231027a06f9v00
On-site reading page: https://ffcap.cn/en/research/20231027a06f9v00