A single-workflow diagnostic

The 10-node map

The bottleneck was rarely the model. It was the data underneath, and who owns the decision.

The map I use breaks a B2B order into 10 value activities: Concept → Design → R&D → Manufacturing → Pricing → Marketing → Transaction → Warehousing → Delivery → After-sales AI does different work at each node, and few firms cover all 10, so starting points differ.

Underneath sits Node 0, the foundation: • Product data: standardized SKUs, so AI can read what you sell • Know-how data: veterans' experience codified, so AI can use how the work gets done If either is thin, nothing above it runs reliably.

Public-company research

Based on public disclosures only — not clients.

Milkyway (SSE: 603713), chemical supply chain: AI is concentrated in billing and dispatch, mostly internal. Transaction-side evidence is thin.

Shanghai Ganglian, parent of Mysteel (SZSE: 300226): strongest evidence at Node 0, industry data plus IOSCO-compliant price indices. AI in trade matching is not yet proven.

Huitongda (HKEX: 9878), supplying small stores in lower-tier markets: AI lands in marketing and transaction. Marketing disclosures are mostly product features; results are unproven. Its key asset: store data being turned into reusable know-how.

The common sticking point: where the foundation is weak and accountability unclear, AI stays in assist mode. If listed leaders are there, a mid-sized B2B firm has no reason to start with full automation.

Independent diagnostic vs vendor pre-sales

Scope & disclaimer

About

Shen Fei. 12+ years in B2B marketing, brand and industry research. Based in Beijing.