From “New Bottle, Old Wine” to AI‑Native Organizations: What Ontology Governance Really Means for Enterprise AI

The discussion opened with Zhang Sen‑sen asserting that ontology is essentially a “new bottle with old wine,” a layered data‑modeling approach that now incorporates semantic capabilities thanks to large language models. He emphasized that the real difficulty lies not in building classes or relationships, but in defining ownership, maintenance, and arbitration of those entities—making ontology a management problem as much as a technical one.

Zhang highlighted a modeling paradox: while business teams feel the pain of modeling the most, the IT side does not experience the same level of distress. Early attempts relied on ad‑hoc solutions like LangChain‑style skill pipelines, but the rise of large models has widened the gap compared to earlier ontology products such as Palantir’s. Meanwhile, fundamental data‑engineering tasks—massive graph retrieval, materialized views, and unstructured data ingestion—remain hard problems that AI models cannot yet solve.

Ma Jin‑long added a practical framework, dividing ontology work into three blocks: (1) defining ontology and its governance boundaries, (2) leveraging large models for efficient ontology construction and extraction, and (3) ensuring quality through evaluation, whether the work is manual or model‑driven.

The panel then diagnosed two common “diseases” of enterprise AI transformation. The first is a symptom‑driven approach where companies hope a single AI solution will cure business bottlenecks. The second is a herd‑mental‑driven rush to adopt AI simply because competitors are doing so, often leading to superficial “AI‑coding” projects that replace programmers without addressing deeper workflow issues.

Highlight projects were described as a double‑edged sword. While they showcase rapid front‑end results, they generate massive technical debt on the back‑end, especially when dealing with uncontrolled unstructured data. An automotive‑industry anecdote illustrated how pressure to deliver AI quickly forces teams to pick low‑hanging‑fruit solutions like AI‑assisted coding, neglecting the reality that most enterprise work is communication‑heavy rather than code‑heavy.

Using a steam‑engine‑to‑electric‑motor analogy, Zheng Yan argued that swapping the engine without changing the organization yields little impact; deep organizational restructuring is far harder than building an AI application. He noted that even well‑funded firms struggle to start from scratch, whereas smaller “baby‑stage” teams may find it easier to experiment.

Ma described his team’s “AI‑Native” experiment: they abandoned traditional front‑end/back‑end/testing/algorithm silos, allowing a single person empowered by AI to handle up to 80% of a project’s tasks, with humans focusing on judgment and verification. However, they also warned that once AI simplifies creation, the hardest problem becomes identifying the right problem to solve.

In closing, the speakers agreed there is no universal answer, but enterprises must move forward. Zhang reiterated that governance—who defines and maintains ontology—is the biggest hurdle; Zheng reminded that legacy structures cannot be overturned by merely changing technology; and Ma demonstrated that a lean, AI‑native organization can be built from the ground up, albeit with many open questions that will evolve as models improve.