Alibaba 1688 Digital Employee: A Business Teardown

I. Background and Operating Results

Alibaba’s AI-driven merchant-operations capability is deployed mainly across three properties: 1688, its international B2B site (ICBU), and Taobao/Tmall’s seller workbench (Qianniu). The 1688 and ICBU teams have converged on a similar approach — a “digital employee” that combines conversational Q&A with agent-executed skills — while the Qianniu team is still concentrated on content/creative features, with agent-style features still under evaluation.

1688 grew out of B2B trade: most of its merchants have strong offline manufacturing capability but weak online operating skills, which maps closely onto Meituan’s own merchant base — so this piece focuses on 1688’s AI operations product specifically.

The core pain point: 1688’s merchant base shows weak understanding of platform rules and a real skills gap in running their own operations. The large majority of merchants have never participated in an official promotional campaign, and most lack the ability to build their own marketing engine. Millions of listings sit in a persistent “no traffic — no sales velocity — no active management” dead zone, and long-tail merchants are getting squeezed out faster as traffic concentrates around top sellers.

Why the project exists, and where it started: 1688’s AI-driven operations product was built to close that gap. The team began preparing after the 2023 wave of large-model breakthroughs, started working with merchants in early 2024, and shipped its core capability by mid-2024. It targeted long-tail merchants with meaningful order volume but no third-party service provider — merchants who lack operating skill but have real willingness to pay for a simple, well-priced AI package, since it lowers their labor cost through “foolproof” automation. Mid-to-long-term goal: replace full-stack third-party account management with “AI for everyone,” automating the entire workflow apart from financial operations.

Current performance (a category leader):

• Activity: the AI operations feature’s DAU peaks meaningfully above its weekday average, itself already running well ahead of its original internal target.
• Retention and output: weekly retention sits under half; cumulative optimized-listing count has reached a six-figure total.
• Commercialization: merchant paid-conversion sat at a modest share during the early validation window (the feature is now free for all merchants).
• Adoption: the internally tracked AI-recommendation adoption rate runs meaningfully lower than the figure quoted externally.

---

II. Technical Implementation

• Base model: primarily Qwen, with some locally-deployed DeepSeek. 1688 doesn’t restrict itself to internal or external models — it plugs in whichever model leads in a given sub-capability. For example, a proprietary MoE commerce model has a lead in traffic optimization, while Taobao’s proprietary image-generation model leads in creative generation.
• Data sources: training data draws on merchants’ existing operating data, internal category-specialist playbooks, and performance data from skills already deployed. The team runs a regular cadence (roughly every one to two months) reviewing high-frequency questions and collected requirements for use in model training and iteration. Requirements come from three sources:

Source Team	Nature of Requirements
Category operations	Relatively well-defined (industries differ a lot — e.g. steel-trading inventory management is fairly standardized), though not yet down to the feature level. Category-specific customer interviews plus sampled KA customization work have surfaced common pain points across top- and mid-tier sellers.
Sales/business development	Relatively vague and qualitative, needing follow-up clarification.
Merchant-facing product managers	Mostly complaint-driven.

2.1 Technical Approach

• Process control: most current scenarios rely on engineering-defined constraints, with the model executing a human-authored standard operating procedure. For long-tail scenarios, the model is given an API plus an attached knowledge base and left to complete the task on its own. The team has already hand-enumerated 50–60 strategies and is close to exhausting that approach — future strategy discovery will rely on data mining instead.
• Multi-agent state of the art: the team built its own MCP-like (multi-agent communication protocol) coordination mechanism to orchestrate different AI roles. Multi-agent architecture is still at an early stage overall — coordination is mostly hand-coded rather than running on a mature framework.

2.2 Memory Capability

1688’s digital employee gives merchants an “invisible” memory experience — built on product data, conversation history, and order details, integrating a merchant’s sourcing preferences and customer transaction preferences to handle supply-chain lookups and order execution on their behalf. Preferences are visible in the merchant’s workbench.

1) Tiered memory storage architecture: Different agents can share data from the same memory store (mostly long-term memory), enabling cross-scenario continuity.

Type	Storage	Duration	Notes
Real-time memory	Held immediately in VRAM	Same day, same turn	/
Short-term memory	Real-time conversation accumulated into a Redis store	On the order of a few months (not firmly defined)	Anything older than two weeks moves into RAG
Mid-term memory	RAG vector database	Roughly three months to half a year	/
Long-term memory	RAG vector database	Beyond the model’s context limit	Historical info, product info, conversation history, merchant profile

2) Evaluation framework (with a human-annotation team of meaningful scale):

• Latency
• Accuracy: share of retrieved passages that are actually correct
• Contextual consistency: whether generated content logically matches historical memory
• Satisfaction: measured via A/B testing
• Coverage: share of product types and scenarios the system can actually handle

3) Current limitations and roadmap:

• Current limitation: merchants frequently change promotional strategy during a sale, and this constant “noise” disrupts memory coherence.
• Roadmap: build a multimodal product-feature library, replacing pure-ID-based SKU matching with image-text mapping — moving from search that requires an exact [product ID + spec] to a simple [photo + short description] cross-modal search (currently succeeding a large majority of the time, with room to improve).

2.3 Overall Evaluation Mechanism

• For high-frequency, deterministic scenarios (rewriting a listing title, optimizing the primary image), a small-traffic A/B test validates impact.
• For fuzzier, diagnostic-style scenarios, the focus shifts to feature-adoption rate and churn.
• Supplementary evaluation: intent testing (whether the model correctly identifies high-frequency user intents) and human evaluation (by category-operations experts).

---

III. Core Features and Use Cases

1688’s primary push is behind AI operations and the “digital employee,” targeting scenarios merchants already touch frequently, prioritized using an internal 5-tier framework (L1–L3 cover the large majority of usage). E-commerce is a supply-chain workflow that breaks into roughly 10 major stages and 50 sub-stages — AI penetrates deepest wherever a stage is already the most digitized.

Capability tier system:

Tier	Function
L1	Basic storefront setup
L2	Data analysis, report generation (the hard part is tailoring analysis to industry specifics)
L3	Image generation
L4	Digital human/avatar
L5	Strategic recommendations (currently low user trust)

3.1 AI Operations (Foundation Layer)

Spans onboarding, transaction management, marketing/promotion management, and back-office operating-data management end to end. Performance is a category leader; internally, the team mainly tracks the AI-recommendation adoption rate, which meaningfully lowers small-merchant operating cost.

Function Type	Description
Onboarding (priority)	Storefront setup, storefront design, listing publication, CTR/CVR analysis and tuning — lowering the barrier for new sellers.
Transaction management (core)	Scores platform listings on demand-fit and intelligently recommends them to buyers, generating a matched-listing shortlist.
Marketing/promotion management	A standalone dashboard across six modules: 1) product management; 2) traffic promotion; 3) order management; 4) CRM (e.g. city-level analysis); 5) image/video generation tools; 6) customer messaging. Note: module 5 is mostly provided by the cloud arm; the rest is largely delivered via 1688’s own service marketplace, mostly on a monthly-subscription API model.
Back-office operating-data management	Customer management, order fulfillment, an export-customs-clearance virtual assistant, and an AI decision assistant (analyzing order rate, average category score, ratings, highlights, and competitor products).

3.2 AI Digital Employee (Specialized Layer)

Built on top of the AI operations foundation, the “virtual employee” persona makes module-by-module management easier and lands better with merchants. The three most-clicked features are storefront design, listing publication, and the “operations assistant” (embedded in the merchant app for daily use — monitoring performance, optimizing bundles, and tracking competitors).

Interaction model: dialogue leans toward consultation and Q&A (chatbot-style, emphasizing professional efficiency); skills are executed automatically via the agent. Open-ended questions make up over half of queries (split between “how do I,” “what’s my number,” and casual chat).

Digital-employee role breakdown:

Employee Type	Function Walkthrough and Details
1. Business Manager	Q&A: answers high-frequency/open-ended questions, can route action items to the right specialist, pull data, or surface trend opportunities. Skills — operations optimization: automated solutions across both the product and buyer dimensions, each powered by a dedicated sub-agent. Product side: 1) getting a new listing its first order; 2) pushing a new listing past its early sales threshold; 3) racing a promising listing to “gold crown” status; 4) boosting promotional volume on a high-potential listing Buyer side: 1) converting inquiries into sales; 2) winning back repeat customers Workflow example (set a promotional price): Step 1 [Merchant]: select a strategy, click “View now” Step 2 [AI]: combine category, data, and target-buyer signals to estimate impact and recommend the best product/discount combination Step 3 [Merchant]: click “Apply now” Step 4 [AI]: execute with one click; merchant reviews the result
2. Product Specialist	Q&A: diagnoses data on live listings, flags recently underperforming ones for focused analysis. Skills: campaign sign-up currently requires a pop-up click, with a conversational flow planned.
3. Customer Specialist	Q&A: supports customer-profile analysis, churn analysis, performance diagnosis. Skills: lightweight CRM tooling (product seeding, smart bundling) and automated chat handling.
4. Promotion Specialist	Q&A: answers advertising/marketing questions, generates a promotion-performance recap. Skills: competitor ad-spend analysis, automated ad placement, ROI optimization.

---

IV. Team Structure

1688’s AI team runs a meaningful bench of algorithm engineers (split between frontier exploration and specific projects), alongside front/back-end engineering and product management — with no dedicated operations role. The digital-employee team itself peaked at over a hundred people and now runs a much leaner steady-state headcount.

• R&D team: focuses on engineering — agent flow orchestration, tool chaining — with responsibility extending as far as the RAG layer.
• Algorithm team: owns model training (distillation, SFT, end-to-end agent construction, prompt tuning).
• Product team: focused on finding the right scenarios and product-market fit, designing how large-model products land best with merchants.

---

Appendix: Glossary and Definitional Notes

• Qianniu: the Taobao/Tmall seller workbench, used by both Taobao and Tmall merchants, covering messaging, chat, order management, listing management, and more.
• The internal vs. external adoption-rate gap: the internally tracked figure is measured on a unique-visitor basis; a page-view basis would read meaningfully lower still.
• Customer interviews and sampling: covered merchants of varying size, aligned via spreadsheets/screenshots; deep customization sampling focused on high-GMV industries like steel trading, industrial goods, and export rubber products.
• Third-party service marketplace: the platform’s in-app app marketplace; typical listers are long-established Taobao/1688 stores capable of delivering the service themselves.
• Common query-type breakdown: “how do I” (AI matches a knowledge-base article and breaks down the steps), “what’s my number/report” (e.g. margin, 21-day sales, inquiry rate), and casual chat (follow-up or low-signal questions get bucketed here rather than being forced into a strict answer).
• STF (SFT): supervised fine-tuning — further training a pretrained model on labeled data to improve performance on a specific task.
• The operations-optimization team’s mandate: reportedly the first team in the industry willing to directly commit to business outcomes for merchants, rather than just delivering process work (image optimization, listing publication) — accountable for lifting GMV, order count, and marketing ROI as core outcome metrics.