全球数千零售网点、多品牌多云：一家头部奢侈品厂商的数据底座改造 交付效率提升约70%

The most tangible takeaways from DACon: a group of Chinese companies have already moved ontology technology into production systems, with verifiable results. Ant Group reduced cross-opportunity identification from 3–5 days to hours using a knowledge ontology, with anomaly attribution accuracy above 85%. A leading global luxury goods manufacturer revamped its multi-cloud data foundation, improving end-to-end data delivery efficiency by roughly 70%, and raising code generation accuracy from 30% to 90%.

1. To avoid pitfalls, start by remembering five reasons ontology adoption fails: no unified standards, over-engineering into big-bang projects, irreversible implementation paths, conflicting data definitions, and knowledge that depreciates after creation. These are not theoretical issues—they are real obstacles every enterprise will encounter.

2. Three replicable methods stand out: first, scenario-first lightweight ontology modeling rather than attempting full-scale deployment from day one; second, dual-path hybrid routing that gets systems running before gradually accumulating semantic assets; third, a three-tier evaluation and feedback flywheel combining programmatic checks, LLM validation, and human review to ensure trustworthy results.

3. Cases visible to the general public: in manufacturing, quality traceability dropped from several people taking days to one person taking five minutes, while yield improved by 5%. Shopee used a graph platform to identify fraud rings and abnormal transaction chains. Li Auto enabled agents to move from data querying to sales decision-making. These examples show ontology is not esoteric—it is essentially converting business experience into assets AI can use.

For brand owners, the key insight is that the data foundation is no longer just back-end engineering—it directly shapes competitiveness in store operations, user engagement, and anti-counterfeiting traceability. The multi-cloud data foundation case of a leading global luxury goods manufacturer is highly illustrative: amid data silos, compliance barriers, and fragmented AI, they built an Agent-Ready foundation that lifted end-to-end data delivery efficiency by about 70%, raised code generation accuracy from 30% to 90%, and supported scenarios including store optimization, user operations, demand forecasting, and anti-counterfeiting traceability.

1. In brand marketing and channel building: with a unified semantic layer, brands can break down data scattered across multi-cloud and multi-brand systems, forming more complete insights into user behavior, enabling more precise store site selection, layout optimization, and omnichannel operations.

2. In user behavior observation: Li Auto's sales decision case shows that reorganizing objects, relationships, metrics, and business rules with ontology allows agents to progress from data lookup and analysis to sales decisions, helping brands understand customer purchase intent and deliver more effective recommendations and follow-up strategies.

3. In price competition and consumption trends: Shopee's anti-fraud ontology modeling protects promotional campaigns from fraud rings, avoiding losses from abnormal transaction chains. Meanwhile, data-driven demand forecasting helps brands plan inventory and pricing in advance. Brands should recognize that data governance and knowledge assetization are becoming infrastructure—early investment pays off in long-term operational efficiency and decision quality.

For sellers, the most relevant opportunities lie in risk control and efficiency gains in cross-border e-commerce and retail. Shopee's fraud-fighting ontology modeling on a graph platform turns accounts, devices, and transactions into relational entities to identify organized fraud and abnormal transaction chains. This is a direct warning for sellers running cross-border and promotional campaigns: fraud rings can coordinate in disguise, traditional tabular models cannot stop them, and smarter relational analysis is required.

1. In growth markets and consumer demand: the luxury retail multi-cloud data foundation case shows that store optimization, user operations, demand forecasting, and anti-counterfeiting traceability are all being reshaped by AI and data. If sellers can tap into similar data capabilities, they can sense shifts in consumer demand earlier and optimize stocking and marketing cadence.

2. In incident response and opportunity signals: China Telecom Bestpay's review of rework costs in ontology implementation shows that choosing a technical path carefully matters, because some paths are irreversible. For sellers, this signals that relevant technology service providers are maturing; in the future they can buy more reliable data intelligence tools rather than building everything from scratch.

3. In learning points and collaboration approaches: 360's progressive semantic layer methodology offers a “run first, grow with use” mindset. Sellers can adopt this small-step-fast-iteration strategy: use lightweight AI question-answering to solve business problems first, then consolidate recurring experience into long-term assets. Meanwhile, the partnership between Shuxin New Intelligence and the luxury manufacturer demonstrates that building an Agent-Ready foundation with professional service providers is a viable path for traditional retail and e-commerce to acquire new capabilities.

The most direct takeaway for factories is the ontology practice in manufacturing quality traceability: traceability across CRM, MES, ERP, and PLM systems was shortened from multiple people spending days to one person taking five minutes, while the yield rate improved by 5%. This shows that structuring business objects, relationships, and rules with ontology can break system silos, quickly locate quality issues at their source, and reduce communication and rework costs.

1. In product production and design needs: factories starting digital transformation should first build a unified model of business objects such as equipment, processes, materials, and quality metrics, forming a semantic layer that AI can understand and execute. YueDian Technology emphasized “scenario-first lightweight ontology modeling”—do not pursue comprehensiveness, or it will become obsolete first. Factories should enter from high-value scenarios like quality traceability and predictive equipment maintenance.

2. In digitalization and e-commerce insights: Li Auto uses ontology to organize enterprise data, enabling agents to advance from data lookup to sales decisions. This shows factory data should not be used only for reports; it must become knowledge assets that support decision-making. A continuous iteration mechanism of “LLM-assisted extraction plus expert review” is needed to maintain the ontology and prevent knowledge depreciation.

3. In business opportunities: ontology technology creates a traceable semantic network across suppliers, production, quality inspection, and after-sales. This improves internal yield and strengthens responsiveness to downstream customers, becoming a competitive advantage in winning orders and delivering. Factories should pay attention to multi-cloud data foundations and semantic layer platforms, and position themselves early with Agent-Ready data capabilities.

Service providers can see a clear technology demand map: enterprise clients are moving from data platforms to knowledge systems, urgently needing to turn business semantics, relationships, and rules into AI-consumable assets. Multiple companies shared reusable approaches and reflections, forming a solution library for service providers.

1. In industry trends: there is no unified ontology standard; Palantir, Snowflake, Databricks, Apache Ossie, and Agent Memory represent five co-existing paths. When helping clients select technology, service providers cannot just ask whether a graph exists—they must clarify how knowledge is represented, run, produced, and validated. China Telecom Bestpay offered selection criteria at the instance layer: query patterns, latency requirements, whether state and write-back are needed, and change frequency.

2. In client pain points: conflicting definitions, the big-bang failure, depreciation after creation, and irreversible paths are the most frequently cited pitfalls. Ant Group's solution preserves differences and uses state machines plus reverse extraction to auto-build the ontology. China Telecom Bestpay proposed separating the concept layer from the instance layer. 360 uses dual-path hybrid routing to solve cold-start problems, with confidence scores determining which results can be automatically written to reports and which need human confirmation.

3. In solutions: Datastrato proposed a three-tier progression from unified metadata to an open semantic layer to Ontology. Renmin University proposed a self-evolving ontology, forming a closed loop through failure attribution, candidate changes, evaluation gates, and version rollback. Shuxin New Intelligence delivered a complete architecture with a multi-cloud AI-native platform, enterprise semantic layer, sandbox protection, and iterative closed-loop. Service providers should focus on building measurable, acceptance-ready delivery capabilities, for example a five-part template: ontology snippets, rule bases, case libraries, validated queries, and evaluation sets.

Platform providers can identify clear platform demands and product directions from this conference: enterprise clients no longer need just a data catalog or semantic layer—they need complete infrastructure that enables agents to genuinely “understand the business.” Multiple enterprise cases mention the “Agent-Ready data foundation” as a critical need, which is exactly the opportunity window for platform providers.

1. In commercial platform requirements: high-end retail manufacturers face pain points such as data silos, compliance barriers, and fragmented AI. They need a multi-cloud data intelligence foundation plus an enterprise-specific semantic layer, along with sandbox security and fine-grained permission control. Platform providers should make “unified metadata management” and “ontology modeling” core capabilities, not just simple data connectivity.

2. In the latest platform approaches: Datastrato is evolving from a data catalog to an AI Context Platform, using a three-tier progression of unified metadata → open semantic layer → Ontology to connect business concepts, data assets, and governance rules. Yunqi Technology suggests enterprises start from high-value business domains and build a closed loop of knowledge production, evaluation, certification, and renewal. 360 demonstrated the path of semantic assets growing naturally through use: thin API → thick Skill → ontology metadata, eventually with Skills degrading into orchestrators while knowledge is fully provided by the ontology layer.

3. In operations management and risk avoidance: China Telecom Bestpay emphasized that the concept layer and instance layer must be separated. For the instance layer, whether to build mappings or object libraries must be decided in advance based on four criteria; otherwise rework costs are enormous. Platform providers should design products that support hybrid approaches and clearly define boundaries to help clients avoid irreversible selection errors. Additionally, LLM-generated SQL accuracy in production can drop from 85% to 50%. Platform providers must build in three-tier evaluation and confidence-based routing so clients feel safe putting AI into production.

Researchers can observe the latest industry trends and real problems in ontology adoption in China from this conference. Core argument: ontology is widely discussed but rarely successfully deployed, due to lack of unified standards, a tendency toward big-bang projects, irreversible paths, conflicting definitions, and post-build depreciation. This provides a clear problem list for academic research.

1. In new industry trends: Ant Group, Li Auto, Shopee, China Telecom Bestpay, 360, and Shuxin New Intelligence collectively showcased production-grade ontology applications across seven industries: finance, automotive, e-commerce, payments, retail, and industrial software. This means ontology is moving from concept validation to large-scale empirical validation. Researchers can distill reusable engineering patterns from these cases.

2. In new problems and theoretical frameworks: China Telecom Bestpay revealed that the division of responsibilities between the concept layer and instance layer is a core difficulty, and the choice between mappings and object libraries is irreversible. Renmin University proposed a self-evolving data intelligent agent that uses Ontology as an executable semantic layer, unifying business concepts, tool capabilities, and data models, and forming a Data-Ontology-Agent co-evolution closed loop through failure attribution from interaction traces, knowledge correction, and evaluation gates.

3. In policy and business model implications: the article identifies structural reasons for the scarcity of ontology deployment in China, and stresses that enterprises’ trade-offs under their own constraints matter more than standard answers. Researchers can focus on: how to establish unified yet evolvable specifications? How to design continuous iteration mechanisms to prevent knowledge depreciation? How to quantify ontology ROI? The comparison of five technical routes—Palantir, Snowflake, Databricks, Apache Ossie, and Agent Memory—offers a clear comparative sample for industry research.