Ontology Is a Management Challenge, Not a Technical One – Enterprise AI Insights

On June 15, 2026, DataFunTalk hosted a roundtable titled “Crossing the Enterprise AI ‘Implementation Gap’ – Ontology‑Driven Agents and Knowledge Governance in Practice.” The discussion featured host Zheng Yan (Huawei) and guests Zhang Sensen (Ping An) and Ma Jinlong (Qiming Zhiyuan), who examined ontology modeling, knowledge governance, AI transformation, and organizational change.

1. What is ontology? Zhang opened by likening ontology to “new wine in an old bottle,” arguing that it is essentially hierarchical modeling—an evolution of traditional dimension tables, code tables, and master data management—enhanced by large‑model language understanding. He emphasized that the real difficulty lies not in the technical construction of classes, attributes, and relationships, but in governance: deciding who defines a class, who maintains its attributes, and who arbitrates conflicts or model errors.

2. The modeling paradox Zheng noted that while modeling feels like the biggest pain for business, IT does not share that intensity. Early on, teams cobbled together solutions using LangChain‑like or plugin‑like skill stacks. Over time, the consumption of models by large‑scale AI has created a gap compared with earlier ontology products (e.g., Palantir), making modeling increasingly diverse. He highlighted that core data‑engineering challenges—massive graph retrieval, real‑time materialized views, and unstructured data ingestion—remain hard and costly.

Ma reinforced this by stressing that regardless of the modeling methodology, the first step is data governance: de‑duplication, noise removal, and information extraction (OCR, chart parsing). Even with multimodal models that can extract semantics from images, the governance workload cannot be avoided.

3. Two “diseases” of enterprise AI Zhang described two typical corporate mindsets: (a) companies facing operational bottlenecks hope AI will be a cure‑all; (b) companies without obvious problems rush to adopt AI simply because peers are doing so. He warned that treating AI as a tool to replace a workflow often leads to layoffs, and that without a fundamentally AI‑native process, competitive advantage evaporates.

4. The hidden cost of “highlight projects” Zhang warned that flashy AI projects generate massive technical debt, especially when they involve uncontrolled unstructured data. He likened the situation to a well‑designed front‑end but a “wet‑back” backend, using an automotive example where leadership demanded rapid AI deployment despite weak data foundations, resulting in quick‑win AI‑coding attempts that ignore deeper data‑governance needs.

5. Steam engine vs. electric motor analogy Zheng compared legacy AI transformation to replacing a steam engine with an electric motor while leaving the rest of the factory unchanged. He argued that organizational inertia, legacy processes, and departmental interests are the real barriers, not the technology itself.

6. Building an AI‑Native organization from the “baby” stage Ma shared his team’s experiment: starting from scratch, they treat all documentation in AI‑readable formats (e.g., Markdown) and position AI as the primary production force, with humans as assistants. They identified domains (finance, legal, government relations, IP) that must remain human‑driven, and they kept the organization flexible to adapt as models evolve. In practice, a single person can cover front‑end, back‑end, testing, and algorithm work, with AI handling 70‑80% of tasks and humans focusing on judgment.

7. No standard answer, but you must move forward The 90‑minute dialogue concluded without a silver bullet. All three speakers agreed that while management challenges outweigh technical ones, enterprises must still act: “If you don’t try, you fall further behind; if you do, you at least push the needle forward.”