2026-08-23. The origin of this was a desire to evaluate a direction: AI + E-commerce ERP for a startup. Starting from the question “What systems do China’s largest e-commerce sellers actually use?”, it eventually landed on a list of software copyright application materials. This post lays out the entire research chain across four parts: market research, startup strategy, platform API barriers, and software copyright execution.

Part 1: ERP Market Research — Where the Money Is, Who’s Earning It

This section used 107 research agents across approximately 7.6 hours. Every key finding was checked against three independent perspectives, and any claim with 2 out of 3 rebuttals was discarded. The full report is at ERP_RESEARCH_REPORT.md (50KB); here I only cover findings that withstood verification.

1.1 What Do China’s Large E-commerce Sellers Use?

The answer is more concentrated than expected: Jushuitan (聚水潭), Wangdiantong (旺店通), and Jikecloud (吉客云), segmented by order volume:

Jushuitan went public on the Hong Kong stock exchange in 2025, processing nearly 100 million orders per day. This isn’t marketing fluff — IPO data carries legal liability.

1.2 Vendor Marketing Numbers Are Basically Unreliable

Adversarial checking eliminated a number of widely circulated figures:

• “Kuaimai serves 4.5 million merchants, 1.4 million paying customers, 80 million parcels daily” — 0/3 votes, pure vendor self-promotion, no third-party audit
• “Wangdiantong has 400,000 customers, 50,000 orders per second during peaks” — 0/3 votes
• Various media reports on Jushuitan’s financial data contradict each other — only IPO-grade data retained

Lesson: Most ERP market data on the Chinese internet is just PR reprints. Whenever you see “XX million merchants,” ask who audited it first.

1.3 Open-Source ERP and Chinese E-commerce Are Two Parallel Worlds

Odoo (Community Edition, LGPL-3) and ERPNext (GPL-3.0) are both mature open-source ERPs, but here’s the critical fact: neither natively supports any of Taobao, Tmall, JD.com, Pinduoduo, or 1688. The e-commerce connectors in Odoo’s official app store only cover Amazon, Shopee, Lazada, and TikTok.

This is precisely why the sellers who actually make money don’t use Odoo — not because open source is bad, but because the gritty, tedious work of integrating platform APIs has already been done for ten years by domestic e-commerce ERP vendors, forming a moat. Traditional ERPs are organized around finance and production; Chinese e-commerce ERPs are organized around order flows. The product structure differs at the root.

1.4 AI ERP: L3/L4 Simply Don’t Exist Yet

Looking at the market through a maturity lens in 2026:

• L1 (chatbox added on top): Everywhere
• L2 (AI data querying): SAP Joule and Dynamics Copilot are working on it, but independently verified cases are rare
• L3 (AI analysis + API execution): No product has stood up to verification
• L4 (autonomous discover-decide-execute-verify closed loop): Doesn’t exist

Every marketing claim of “AI Agent operating an ERP” currently sits at L1/L2. This is a real gap — but honestly, a gap doesn’t automatically equal an opportunity, and I’ll explain why below.

Part 2: Six Startup Routes — Why We Chose B

2.1 Six Routes at a Glance

2.2 Conclusion: B, but the platform combo was rewritten by reality

For a 2–5 person team, B is the only route that satisfies “move fast, validate early”: zero licensing cost, start charging seed users after building just one module, and failure is affordable.

But the original plan to start with “Taobao + Pinduoduo + 1688” was forced to change after checking each platform’s open API:

MVP = Pinduoduo + Douyin + 1688; Taobao and JD.com will come later.

See Part 3 for reasons.

Part 3: Platform API Barriers — The Death Gate for the Entire Route

This section checked each platform’s latest 2025–2026 policies. It was the most time-consuming and most valuable part of the research.

3.1 Five-Platform Comparison

Three counterintuitive findings:

1. The largest platform has the highest barrier. Intuition says start with Taobao, but it’s exactly the platform individuals can’t enter at all.
2. “Start with self-use, move to tool-type later” is the only viable path, and it only works for Pinduoduo / Douyin / 1688: first verify the full pipeline with your own store’s authorization, then apply for ISV marketplace listing once you have real cases. Tool-type applications have a chicken-and-egg problem — they may require evidence of already-listed apps and customer counts — so keep thorough records of real usage during the self-use phase.
3. 1688 is critical on the procurement side. Its distribution/dropshipping APIs are free at the basic tier, but rate-limited. You must stress-test the throttling strategy in week one of development.

3.2 Confirming the Competitive Gap

On the open-source side: there is no existing solution for “Pinduoduo + 1688 one-click dropshipping full pipeline” on either ERPNext or Odoo. The strongest SaaS competitor, Dianxiaomi (店小秘), has a fairly complete free tier, but the dropshipping step requires manual operation — automation depth and data sovereignty (open-source base allows self-hosting) are the differentiators.

Part 4: Software Copyright Application — The Only Tangible Thing Worth Pushing During the Freeze

Why this detour: Pinduoduo’s tool-type ISV application lists software copyright as a hard prerequisite, and the approval cycle runs 1–3 months — which fits perfectly into the project freeze period. Below are the latest requirements verified as of 2026-08.

4.1 Checklist (4 Core Document Categories)

① Registration Application Form: Fill out online at the Copyright Protection Center website. Software name must follow the “XX System / XX Software” convention, version like V1.0, development completion date must precede publication date, and all fields must match the code and manual exactly.

② Software Manual: ≥15 pages (30 recommended), must include real running screenshots, clearly explaining features and operational flows with illustrations, plus runtime environment details. For backend systems with no UI, supplement with API docs and processing flowcharts.

③ Source Code Document (highest rejection rate, accounting for 35% of corrections):

• First 30 pages + last 30 pages of continuous code, 60 pages total
• ≥50 lines per page (counting after stripping blank lines and comments)
• Recommended total: 6,000–10,000 lines; select business-unique logic
• Header: software name + version number (matching the application form) + page number in the top right
• First line is the starting statement, last line is the ending statement
• No other parties’ Copyright/Author notices permitted anywhere
• If open-source code is used: mark the source + attach the license text + document modifications

④ Identity Proof: Individual = photocopy of ID card (front and back); Enterprise = business license copy stamped with company seal.

4.2 Process and Cost

Fully online (paperless reform in June 2023 means only electronic certificates are issued now):

1

Real-name verification (1-3 business days) → Online pre-review → Formal examination (30-40 business days standard) → Electronic certificate

The truth about fees: the official registration fee was cancelled in 2017; doing it yourself costs ¥0. The hundreds to thousands charged by agents are purely service fees; “expedited” processing is an agent value-add, not an official channel. If time is tight, paying for an agent once is fine; otherwise, do it yourself.

4.3 2026’s “Strictest Review Year” — Red Lines

Rejection rates rose 40% year-over-year in 2026. Eight categories of behavior trigger immediate rejection and placement on a credit blacklist. The four most relevant to us:

1. Pure AI-generated submissions will be rejected — Using AI to assist coding is fine, but there must be substantive human modification, and records of that modification must be retained
2. Code similarity checking system is now live — Full-text comparison against the historical registration database and all public open-source projects; shell-wrapping open-source code will fail
3. Boilerplate manual language = 100% rejection — Phrases like “the system is stable and easy to operate” will get you rejected; every claimed feature must map to a specific code module
4. Submit ≤3 applications per day, don’t fragment to pad numbers — One real system, one application, that’s enough

The best protection is actually just good habit: git commit history, project initiation documents, test reports — these authentic records of the development process are what give you confidence when facing any verification.

4.4 Our Execution Plan

• Applicant: The independent integration service from Route B, named something like “XX E-commerce Order & Procurement Integration System V1.0”
• Timing: Compile materials immediately after developing a real, runnable feature; all screenshots come from actual running interfaces
• During the freeze period: Individual real-name verification + git record keeping — just those two items, no new project initiation

Appendix

• Full market report: ERP_RESEARCH_REPORT.md
• Route B implementation plan: erpnext-route-b-plan-2026-08-23.md
• Software copyright checklist: ruanzhu-checklist-2026-08-23.md
• Methodology note: The market section was verified via multi-agent adversarial checking (3-vote system). The platform API and software copyright sections are first-hand information verified online on 2026-08-23; please cite the date when referencing.