--- title: "Why AI Grounding Is Driving Ontology Demand for Enterprise Agents" url: "https://atlan.com/know/ai-agent/ontology/ontology-ai-grounding/" description: "Learn why AI grounding put ontologies back on the enterprise roadmap, what ungrounded agents get wrong, and why a first ontology draft is now cheaper to build." author: "Karthik Pasupathy" author_role: "Contributing Writer, AI Context & Agents" published: "2026-10-27" updated: "2026-10-27" --- > Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/ ## Why has AI grounding renewed demand for ontologies? AI grounding connects an agent's output to authoritative business definitions and source evidence. Ontologies provide this shared, machine-readable meaning, which helps agents interpret terms such as customer, revenue, or policy consistently. Ontology demand is growing because agents need consistent business meaning when they run tasks with no person checking each answer. LLMs reduce the cost of a first ontology draft, but people must still approve canonical meaning. Four properties explain why ontologies are becoming more valuable and practical: - **Precise definitions:** Establish which definition applies to critical business terms such as customer, revenue, or policy. - **Traceable evidence:** Connects each concept to its source data, governing rules, relationships, owners, and approval status. - **Shared context:** Gives different agents the same approved definitions instead of letting each system infer its own interpretation. - **LLM-assisted drafting:** Produces initial ontology structures for subject-matter experts to review. ## Key takeaways - AI agents turned ambiguous business terms into repeatable errors that spread across systems. - Ontologies ground agents by making business meaning explicit, shared, and traceable. - LLMs reduce the cost of a first ontology draft, but people must still approve canonical meaning. - Ontology demand reflects an infrastructure shift driven by agents. --- AI agents need grounding in business meaning because the terms they act on are rarely universal. Business terms like revenue, customer, and churn can each have several valid definitions inside one company. An ontology makes those definitions and their relationships explicit in a form an agent can use. That is the gap Atlan, the Context Layer for AI, is designed to address. It connects business meaning to data, lineage, owners, and approval status so agents can identify which definition applies and trace it to its source. --- ## Why is everyone suddenly talking about ontologies and AI grounding? An [ontology in AI](https://atlan.com/know/what-is-ontology-in-ai/) is a machine-readable model of business concepts, relationships, and rules. If you are new to ontologies, the [Ontology 101 explainer](https://atlan.com/know/ontology-101-explainer/) is the best place to start. Ontologies themselves are not new. What changed is who consumes the business meaning they encode. Human analysts could rely on judgment and [tribal knowledge](https://atlan.com/know/data-for-ai/tribal-knowledge/) to resolve ambiguity as they worked. They knew which dashboard finance trusted, which customer definition applied to renewals, and whom to ask when metrics conflicted. AI agents are now taking on some of that analytical and operational work, but they do not share the same institutional knowledge by default. An ontology makes that knowledge explicit, so an agent does not have to infer which definition to use before completing a task. Without it, a plausible but incorrect interpretation can spread across reports, recommendations, and other agents. Once agents became consumers of business meaning, ontology moved from a specialist modeling exercise to an enterprise AI concern. It now appears alongside [ontology-first AI architecture](https://atlan.com/know/ontology-first-ai-architecture/) and [AI-ready data](https://atlan.com/know/ai-readiness/ai-ready-data/) because the economics have changed on both sides: agents have increased demand for approved definitions, while LLMs have reduced the effort required to draft them. Understanding that demand starts with what grounding actually requires. --- ## What does AI grounding mean, and why do embeddings alone not get you there? AI grounding connects an agent's answer to authoritative business meaning and the evidence behind it. It tells the agent not only what information is relevant but also which definition, calculation, relationship, and source the organization recognizes for that context. Consider the question, "What was our Q3 revenue, and how did it change from Q2?" Depending on the team, revenue might mean bookings, billed revenue, recognized revenue, or a metric maintained by analytics. An embedding-based retrieval system can find tables, dashboards, and documents that are semantically related to the question. But without an explicit definition for that context, the agent can retrieve relevant information and still choose the wrong revenue metric or join path. [Embeddings capture](https://atlan.com/know/what-are-embeddings-ai-search/) statistical similarity, which makes them useful for finding related information. [Semantic search](https://atlan.com/know/semantic-search-vs-keyword-search/) uses that similarity to improve retrieval. But similarity answers the question, "What looks related?" whereas grounding answers questions like, "What does this term mean here, how is it calculated, and which source is authoritative?" Several structures can contribute to grounding, but they are not interchangeable. A [semantic layer](https://atlan.com/know/semantic-layer/) standardizes business metrics and query logic. An [ontology differs from a semantic layer](https://atlan.com/know/ontology-vs-semantic-layer/) in that it formally defines domain concepts, relationships, and rules. A [knowledge graph](https://atlan.com/know/what-is-a-knowledge-graph/) connects specific entities and facts. In an [ontology versus knowledge graph](https://atlan.com/know/ai-agent/knowledge-graph/ontology-vs-knowledge-graph/) comparison, the ontology defines what kinds of entities and relationships can exist, while the knowledge graph records the instances and connections that do exist. A [context graph](https://atlan.com/know/context-graph-vs-ontology/) adds operational context such as lineage, owners, policies, and decision evidence. The central distinction remains simple: retrieval finds related information, while grounding resolves it against authoritative meaning. In a 2026 benchmark of 100 retail analytics questions, [Rumiantsau and Fokeev](https://arxiv.org/abs/2604.25149) gave three frontier models either the warehouse schema alone or the schema plus a 4 KB hand-written document of measures, conventions, and disambiguation rules. Accuracy rose from 45.5 to 50.5% without the document to 67.7 to 68.7% with it, a gain of 17 to 23 percentage points, and the gap between models narrowed to about one point. That document is a semantic-layer description, not a formal ontology, so the result shows the value of explicit business meaning in any form. An ontology earns its extra structure when the same meaning has to hold across domains and systems, the boundary drawn in [whether an ontology is just another word for a semantic model](https://atlan.com/know/ai-agent/ontology/is-an-ontology-just-another-word-for-a-semantic-model/). --- ## What do ungrounded agents get wrong about business terms? Ungrounded agents can produce technically valid outputs built on the wrong business meaning. They may use the wrong join, choose the wrong metric, or apply different definitions across agents. Because each result can look reasonable, it may take time for someone on the team to notice the error, often after it has been included in a report or influenced a decision. Three failure patterns recur when an agent picks the wrong business meaning: | Failure | What it looks like | Evidence | | :---- | :---- | :---- | | Hallucinated join | A sales agent joins the account and invoice tables using similarly named customer ID fields that refer to different entities. The SQL runs, but the revenue total is wrong. | [Rumiantsau and Fokeev (2026)](https://arxiv.org/abs/2604.25149) identified the need to infer business semantics from schemas as a source of incorrect answers and confident text-to-SQL hallucinations. | | Wrong metric definition | A finance agent reports bookings when the question calls for recognized revenue because bookings appear more closely related to the prompt. | Across 1,800 runs involving five industries and three LLMs, [Luong Tuan and Sanyal (2026)](https://arxiv.org/abs/2604.00555) found that ontology-coupled agents performed significantly better on metric accuracy and role consistency (p < .001 for both). | | Inconsistent answers across agents | Finance and support agents use different definitions of "customer" or "refund," leading to conflicting answers to the same question. | [Gartner (2026)](https://www.gartner.com/en/newsroom/press-releases/2026-03-11-gartner-announces-top-predictions-for-data-and-analytics-in-2026) predicts that by 2030 universal semantic layers will be treated as critical infrastructure, and says they are needed to align multiagent systems and stop costly inconsistencies before they spread. | These errors require a narrower diagnosis than simply identifying and evaluating [LLM hallucinations](https://atlan.com/know/llm-hallucinations/). An [AI agent hallucination](https://atlan.com/know/ai-agent-hallucination/) may involve fabricated information, but an ungrounded agent can also be wrong while using real data, valid SQL, and a legitimate metric. The failure lies in choosing the wrong business meaning for the task. Once an agent's incorrect choice of business meaning enters a report, tool call, or downstream agent, it becomes one of the reasons [AI agents fail in production](https://atlan.com/know/why-ai-agents-fail-in-production/). Improving [AI agent accuracy](https://atlan.com/know/ai-agent/ai-agent-accuracy/), therefore, requires [business context for AI](https://atlan.com/know/business-context-for-ai/) that specifies which definitions and relationships are authoritative. An [enterprise context layer](https://atlan.com/know/what-is-the-enterprise-context-layer/) makes that approved meaning available across systems. This is [why agents need a context layer](https://atlan.com/know/why-ai-agents-need-an-enterprise-context-layer/) instead of another definition inferred locally for each task. These failures explain why the demand for ontology grew. The next question is why enterprises tolerated this ambiguity for so long. --- ## Why did ontology stay in the back office for twenty years? Ontology work stayed concentrated in specialized domains because it was expensive to create and maintain. Experts had to interview domain teams, model entities and relationships, resolve conflicts, and keep the result current as systems changed. Analysts could also work around missing definitions. They could consult a glossary, inspect the underlying query, or ask the metric owner before using a number. People spread inconsistent metrics too, but slowly, and a person usually sat between the number and the decision. The same economics slowed many enterprise knowledge programs. Work on [why enterprise knowledge graphs stalled](https://atlan.com/know/ai-agent/knowledge-graph/enterprise-knowledge-graph-pitfalls/) and [why semantic layers stalled](https://atlan.com/know/semantic-layers-failed-context-graphs/) shows the recurring cost of narrow coverage and weak maintenance. Richard MacManus's July 2026 essay in Latent Space, ["Ontologies Are So Back"](https://latent.space/p/ontologies-agentic-systems), captures the revival from the AI-engineering side, but the bigger change is who now consumes the model. --- ## What changed when AI agents became consumers of business meaning? The rise of agentic workflows has made AI agents major consumers of business meaning alongside humans. Agents use definitions, relationships, and rules as inputs when they select data, calculate metrics, and complete tasks across enterprise systems. The timing of this shift is visible in Gartner's [2025 AI Hype Cycle](https://www.gartner.com/en/newsroom/press-releases/2025-08-05-gartner-hype-cycle-identifies-top-ai-innovations-in-2025), which named AI agents and AI-ready data the two fastest-advancing technologies and placed both at the Peak of Inflated Expectations. The release does not mention ontology. The broader signal was that enterprise attention was shifting from model capabilities to the data and context agents require to deliver reliable, consistent outcomes. Gartner made the need for semantics more explicit in 2026. Its [release on semantics and AI agents](https://www.gartner.com/en/newsroom/press-releases/2026-05-11-gartner-says-lack-of-semantics-causes-inaccurate-artificial-intelligence-agents-and-wasted-spending) states that agents need context at each step of an agentic workflow. It also predicts that, by 2027, organizations that prioritize semantics in AI-ready data will increase their agentic AI accuracy by up to 80% and reduce costs by up to 60%. Those are forecast ceilings, not measured results. Rita Sallam, Distinguished VP Analyst at Gartner, summarized the business case this way: *"Context with semantic coherence will become a cost-control and trust strategy, not a nice-to-have."* Business meaning is therefore no longer documentation that an analyst consults only when something is unclear. It is an input that agents consume while they work. This is why [systems of semantics](https://atlan.com/know/ai-agent/data-for-ai/systems-of-semantics/), [context graphs](https://atlan.com/know/gartner-context-graphs/), and [AI context platforms](https://atlan.com/know/ai-agent/context-layer/gartner-impact-radar-ai-context-platform/) have entered the enterprise AI infrastructure conversation. Agents raised the cost of ambiguous meaning. At the same time, LLMs began lowering the cost of creating the ontologies needed to make that meaning explicit. --- ## What changed on the supply side? Why are ontologies cheaper to build? LLMs can now draft ontology structures from material an enterprise already has: SQL, schemas, lineage, BI logic, usage patterns, definitions, and annotations. This reduces blank-page modeling, but it does not remove review. Two recent studies show both the opportunity and the limit. In a question-to-report system deployed at China Unicom, [Duan et al. (2026)](https://arxiv.org/abs/2609.27257) reported that ontology construction across 27 tables and thousands of metric types took a few hours, compared with about one week of manual effort. The pipeline paired LLM generation with expert-authored business skills, question verification, and selected expert review. The ontology-grounded system achieved 95.0% strict accuracy on real business questions, compared with 72.5% for document RAG. It is one deployment at one company, and the time figure covers building the first ontology. Keeping it current is a separate cost. [Lippolis et al. (2025)](https://arxiv.org/abs/2503.05388) evaluated two prompting techniques and three LLMs on a benchmark containing ten ontologies, 100 competency questions, and 29 user stories. One combination, OpenAI o1-preview with the Ontogenia technique, produced ontology drafts of sufficient quality to meet ontology engineers' requirements and significantly outperformed novice ontology engineers in modeling ability. However, the authors also observed common mistakes and variations in quality, supporting the use of LLMs as ontology-authoring aids rather than autonomous decision-makers. The study measures quality, not time or cost. So the evidence supports a narrower claim than "ontologies are cheap now": the first draft is cheaper to produce. Getting finance, sales, and support to agree on one canonical definition, and keeping it current as schemas change, remain the larger costs, and neither study measures them. This changes the starting point for [building a knowledge graph for AI](https://atlan.com/know/ai-agent/knowledge-graph/knowledge-graph-construction-for-ai/) and [designing an ontology for AI](https://atlan.com/know/ai-agent/ontology/ontology-design-for-ai/). Teams can begin by [mining context from existing systems](https://atlan.com/know/ai-agent/context-engineering/context-mining-for-ai/) and use a [context development lifecycle](https://atlan.com/know/ai-agent/context-engineering/context-development-lifecycle-for-ai-data/) to test, review, approve, and maintain it. That is [context engineering](https://atlan.com/know/what-is-context-engineering/) applied to semantic work. ### How does Atlan help enterprises build and manage ontologies? Atlan, the Context Layer for AI, helps enterprises respond to the supply-side shift by building an ontology layer from signals already present in their systems. Atlan's Context Agents analyze column-level lineage, SQL, BI semantic logic, usage patterns, and annotations to propose terms, relationships, and mappings. Once reviewed and approved by domain experts, this ontology layer provides shared business meaning that agents can use consistently. Atlan supports each ontology task with a different capability: | Ontology task | How Atlan supports it | Human role | | :---- | :---- | :---- | | Creating the first draft | Orion, the Ontologist [Context Agent](https://atlan.com/context-agents/), maps the relationships between domains, terms, and data assets from the existing enterprise context. | Domain experts confirm that the proposed model reflects how the business operates. | | Resolving conflicting definitions | Sage identifies competing metric definitions, while Nexus connects technical fields to business terms. | Metric and domain owners decide which definition and mapping apply in each context. | | Reviewing and approving the ontology | [Context Engineering Studio](https://atlan.com/context-engineering-studio/) provides a workspace for testing, reviewing, approving, and releasing the proposed context. | Domain teams determine whether the context is accurate and ready for use by agents. | | Sharing approved context | [Context Lakehouse](https://atlan.com/context-lakehouse/) makes the approved context available through MCP, SQL, and APIs. | Admins and owners set the personas and policies that scope it. Every [MCP](https://atlan.com/know/what-is-atlan-mcp/) call runs under the requesting identity's existing Atlan permissions. | Joe DosSantos, VP, Enterprise Data and Analytics at Workday, describes that last step, where shared language reaches AI through MCP: *"Atlan captures Workday's shared language to be leveraged by AI via its MCP server. As part of Atlan's AI labs, we're co-building the semantic layer that AI needs."* Building an ontology is only the starting point. An [operational ontology](https://atlan.com/know/ai-agent/ontology/what-is-an-operational-ontology/) connects approved meaning to governed actions, while an [active ontology](https://atlan.com/know/what-is-active-ontology/) keeps that meaning current. For organizations working across data platforms, the guide to [extending an ontology beyond one warehouse](https://atlan.com/know/ai-agent/ontology/extending-ontology-beyond-snowflake/) explains how to make that context available across the enterprise. --- ## Real stories from real customers: Shared meaning for AI agents "Context is the differentiator. Atlan gave our teams the shared vocabulary and lineage to move from reactive data management to proactive AI enablement." Kiran Panja, Managing Director, Cloud and Data Engineering, CME Group "Atlan is our context operating system to cover every type of context in every system including our operational systems." Sridher Arumugham, Chief Data Analytics Officer, DigiKey --- ## What does this mean for your AI roadmap? Ontology belongs on the AI roadmap when agents repeatedly disagree about business terms, choose plausible but incorrect joins, or return the wrong version of a metric. Those symptoms show that the model is being asked to infer meaning that the enterprise has not made explicit. The next question is whether recurring ambiguity creates enough cost to justify ontology work for a particular use case. Ontology becomes a shared infrastructure investment when the same ambiguity affects multiple agent use cases. LLMs can accelerate the initial model, while domain experts focus on resolving conflicts and approving the definitions agents will use. --- ## FAQs about ontology and AI grounding ### 1. What does grounding mean for an LLM or AI agent? Grounding connects an LLM or agent's output to authoritative information and explicit business meaning. In an enterprise, this includes approved definitions, source data, relationships, rules, and provenance. ### 2. Why are ontologies suddenly popular again? AI agents need explicit, machine-readable business meaning to complete tasks consistently. At the same time, LLMs can produce initial ontology drafts from existing SQL, schemas, lineage, and BI logic. The need for ontologies increased just as the effort required to start building them began to fall. ### 3. Why can't embeddings or a bigger model resolve business terms on their own? Embeddings help retrieve similar information, and larger models can reason more effectively over the context they receive. Neither determines which definition an organization has approved for a particular task unless that decision is made explicit. ### 4. What goes wrong when an agent is not grounded in business definitions? An ungrounded agent may select a plausible but incorrect join, use the wrong metric, or apply a different definition from another agent. The output can look credible because the query runs and the explanation is fluent. The error may not be detected until someone compares the result with an authoritative business answer. ### 5. Is the ontology revival real, or is it hype? The evidence points to a real infrastructure need, although market forecasts and early studies should not be treated as universal proof. Gartner has identified AI-ready data and semantic context as requirements for reliable agents, while recent studies have reported accuracy improvements from explicit business meaning. ### 6. Can LLMs build ontologies, and how good are the drafts? LLMs can propose entities, relationships, constraints, and definitions from requirements and existing enterprise material. Recent evaluations found that some drafts met expert requirements, and one enterprise deployment reported building its first ontology in hours rather than a week. Quality still varies, so every draft needs testing and expert review. ### 7. If LLMs can draft an ontology, what is left for humans to do? People decide which definition is canonical, which source is authoritative, and which exceptions matter. Domain experts resolve conflicts, approve changes, and determine whether the ontology reflects how the business operates. LLMs reduce drafting effort, but people remain accountable for the meaning agents use. --- ## Sources 1. [Gartner Says Lack of Semantics Causes Inaccurate AI Agents and Wasted Spending, Gartner, 2026](https://www.gartner.com/en/newsroom/press-releases/2026-05-11-gartner-says-lack-of-semantics-causes-inaccurate-artificial-intelligence-agents-and-wasted-spending) 2. [Gartner Announces Top Predictions for Data and Analytics in 2026, Gartner, 2026](https://www.gartner.com/en/newsroom/press-releases/2026-03-11-gartner-announces-top-predictions-for-data-and-analytics-in-2026) 3. [Gartner Hype Cycle Identifies Top AI Innovations in 2025, Gartner, 2025](https://www.gartner.com/en/newsroom/press-releases/2025-08-05-gartner-hype-cycle-identifies-top-ai-innovations-in-2025) 4. [Semantic Layers for Reliable LLM-Powered Data Analytics: A Paired Benchmark of Accuracy and Hallucination Across Three Frontier Models, Rumiantsau and Fokeev, arXiv, 2026](https://arxiv.org/abs/2604.25149) 5. [Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents, Luong Tuan and Sanyal, arXiv, 2026](https://arxiv.org/abs/2604.00555) 6. [UniDataAgent: An Ontology-Grounded Agent for Enterprise Question-to-Report Automation, Duan et al., arXiv, 2026](https://arxiv.org/abs/2609.27257) 7. [Ontology Generation using Large Language Models, Lippolis et al., arXiv, 2025](https://arxiv.org/abs/2503.05388) 8. [Ontologies Are So Back: Why AI Agents Are Reviving the Semantic Web, Latent Space, 2026](https://latent.space/p/ontologies-agentic-systems)