Ontology: The Only Blueprint for Enterprise AI Agents — Forbes & China's Convergence

Forbes Article: Ontology as the Only Path for Enterprise AI Agents

A Forbes Technology Council article by Tarek Nseir, co-founder of Valliance (a European AI value-delivery firm), titled "Entering The Ontology Era: The Blueprint For Enterprise AI Agents," concludes that ontology is the unique answer for enterprise AI agent deployment. The argument rests on three pillars:

Mechanism: Agents can only reason over information they "understand." Enterprise knowledge sits fragmented across CRM, ERP, OA, and hundreds of thousands of documents, never unified. Model intelligence is the ceiling; agent understanding of the enterprise is the bottleneck. Most firms suffer "pilotitis" — buying agents first, running pilots, treating context as an afterthought — reversing the correct order.

History: Palantir's core thesis a decade ago: only by integrating scattered data, processes, and business rules into a coherent model can decisions improve; everything else (including human judgment) layers on top. This is a validated old answer, not a new AI invention.

Convergence: Today, major players independently converge: Databricks released Genie Ontology; Microsoft CEO Satya Nadella authored a piece on the "closed loop" — competitive advantage depends not on accessing frontier models but on knowledge, workflows, and decisions operating as a single system; enterprise search vendor Glean reached a $7.2B valuation and $300M+ ARR via its Context Graph; model vendors race to strengthen memory and orchestration. All point to the same requirement: make AI understand the enterprise.

The article closes: winners will not be those with the best models, but those enabling information, decisions, and actions to flow synergistically — "and ontology is the only answer." When Palantir (data origins), Databricks (data platform), Microsoft (OS), and Glean (startup) — players who never coordinated — converge, it is not a trend judgment but a confirmation .

China: Independent Convergence on the Same Wall

Chinese enterprises exhibit a "wait-and-see" stance mirroring overseas sentiment two years ago, citing three reasons: incomprehension (ontology sounds like philosophy, unrelated to this quarter's agent delivery), investment fatigue (last year RAG, this year MCP, is ontology another hype?), and lack of local cases (overseas stories are fine, but how do Chinese firms actually do it?).

Yet deeper inspection reveals Chinese industries independently hitting the same wall:

Healthcare: In scenarios where every output must be deterministically correct, probabilistic RAG and general models fail entirely. Teams that succeeded found they must restructure the entire stack — model, knowledge, engineering, delivery — with knowledge structured into a verifiable, traceable system.

Industrial: Physical laws are inviolable, process mechanisms admit no probability, tolerance approaches zero. Probabilistic models are inherently unqualified; only by structuring industrial data, mechanistic models, and business semantics via ontology can AI output become auditable and executable.

Platform & Infrastructure: Search, recommendation, and data-platform circles discuss the same need: agents require not just "where data is" but "what data means, how business entities relate, what rules constrain actions." Open semantics, ontology, and metadata are emerging as the three pillars of agent context. Comparative analyses of Palantir, Snowflake, and Databricks ontology approaches for Chinese adaptation are already underway.

Same bottleneck, different industries, each deriving the same direction — identical to the overseas convergence. China's difference: messier terminology, more fragmented systems, larger organizations, heavier semantic debt; but once built, the moat is deeper — uncopyable, unpurchasable, absent from open-source repos.

DACon 2026 Beijing: China's "Ontology Testimony" (Oct 23–24)

The conference main forum structures five keynotes as a linked narrative:

Why ontology is mandatory: Hu Haoyuan, JD Health AI VP — from the least error-tolerant medical scenarios, proves models alone are insufficient: must reconstruct the full "model → knowledge → engineering → delivery" stack, turning knowledge into a structured, verifiable system. In deterministic scenarios, loose knowledge equals no knowledge.

What ontology is and how to land it: Liu Bin, Baidu Data Intelligence GM — business ontology unifies objects, relations, logic, and actions, enabling agents to "analyze, judge, collaborate" under permissions, lineage, versioning, and process constraints. China's platform-level productized answer.

How to build the foundation: Du Junping, Apache Gravitino initiator & Datastrato founder — from data infrastructure perspective, explains construction of the three pillars of trusted agent context: open semantics, ontology, and metadata.

How to choose the technical route: Guan Tao, Yunqi Technology Co-founder & CTO — compares Palantir, Snowflake, Databricks, and other ontology routes: real difference is not "graph or no graph" but how knowledge is represented, executed, produced, and validated. Maps the path from data engineering to knowledge engineering.

Hardest industry — industrial: Feng Xingzhi, CASIC Deputy GM & State Key Lab Deputy Director — why industrial inevitably follows the ontology path, and how to build the industrial world model.

Five speakers, five industries, five perspectives — all answering the same question. This itself is live evidence of China's "convergence."

Breakout Sessions: 30+ Ontology-Related Talks, 27 Organizations

Ontology Engineering Frontline Practices

Li Auto: "Ontology-Driven Agent Practice — From Query & Analysis to Sales Decisions": organizes enterprise data via ontology, turning objects, relations, metrics, and business rules into agent-understandable semantic structures — shifted from post-training near 100% accuracy but unmaintainable to ontology-driven.

Yuedian Technology: "Above the Semantic Layer: Turning Enterprise Tacit Experience into AI-Executable Assets": 12 years, nearly 100 enterprise ontology deployments, dual retrospective of successes and failures.

Ant Group: "Financial Data Knowledge Ontology: From 'Can Query' to 'Understands Business'": data knowledge ontology connects multi-source business semantics — anomaly attribution accuracy >85%, cross-sell opportunity identification reduced to hours.

Shopee: "Anti-Fraud Ontology Modeling & Intelligent Decision Practice on Graph Platform": ontology modeling of accounts, devices, transactions as multiple entities to combat syndicated fraud.

Tianyi Payment: "Fork in the Road of Ontology Landing: Semantic Mapping vs. Object Storage Trade-offs & Practice": ontology as "translation layer" vs. "data layer" — a near-irreversible key selection, with real rework cost retrospective.

Renmin University: "Self-Evolving Data Agents: Toward Data–Ontology–Agent Co-Evolution": academic frontier framework positioning ontology as executable semantic middle layer.

Runhe Software, NetEase Smart Enterprise, Daily Interaction (OntoOS), Mashang Consumer Finance: ontology variants in financial testing, enterprise knowledge governance, operational analysis.

Semantic Layer: Ontology's Caliber Foundation

Ant Group: "Apache Ossie Semantic Layer Landing at Ant": 100+ semantic models, 5,000+ metrics, single semantic source of truth.

Zhihu: "DataClaw: Evolution from MetricFlow to AI-Native Semantic Layer": AI-native lightweight semantic layer + knowledge graph, team penetration >65%.

Qihoo 360: "From 'Can Query' to 'Understands Business' — Data Agent Semantic Layer Evolution & Quantitative Evaluation": academic 85% accuracy in real warehouses plummets; "half-baked cold start, grow while running" path.

Datastrato (Apache Gravitino): "Making Every Data Agent Answer Traceable": trusted text-to-SQL engineering loop at thousands-of-metrics scale.

Dongchedi: "AI-Native Data Semantic Platform Construction": explicit "ontology-based data semantic" platformization.

Amap, Xiaomi, Tencent, Yiche, Guanyuan: semantic layer modeling & evaluation, semantic governance & attribution diagnosis (66,000 traceable knowledge items, 64x analysis efficiency boost), turning BI semantic layer into AI evolution fuel, data semantic library + knowledge self-evolution, three-layer foundation of semantic + ontology layers — semantic layer already in production at Chinese top-tier firms.

Context Engineering: Feeding the Built Ontology to Agents

Li Auto: "Semantic Layer as Context: Semantic Engineering Practice for Production-Grade Agents": splits knowledge scattered across prompts and tools into ontology, OSI, playbooks, and generic execution layer, forming evidence-backed root-cause judgments.

JD: "B-End Agent Context Engineering Practice": multi-layer dynamic memory lets agents understand users more with use.

JD Health (main forum): "Model + Full-Stack Engineering Is the Moat": discards probabilistic RAG, builds deterministic Agentic Search and structured knowledge body — in serious scenarios knowledge must be structured into verifiable systems.

Guanyuan Data, Tianhong Fund: decision intelligence context engineering & evidence chains, investment research multi-modal data governance into computable semantic assets.

Total: ~30 ontology-related sessions, 27 organizations — from internet giants to industrial enterprises, finance to content communities, startups to academia. Not an "overseas concept import session" but a concentrated showcase of how far Chinese practitioners have walked the ontology road.

Final Takeaway

Models will commoditize. The model you can buy, your competitor can buy; this year's leading capability becomes next year's default platform configuration.

Ontology will not. Your modeling of the enterprise business world — every object, every rule, every painful alignment meeting where "we finally agreed what this sentence means" — cannot be copied, bought, or found in open-source repositories.

Overseas has confirmed the answer. China's wall-hitters have lined up at Beijing. Oct 23–24, Beijing, DACon 2026 — Data-First, AI-Real. Come see how Chinese enterprises are doing it, then consider what you can adapt.