Business Context in AI: Definitions, Ownership and Ontologies
Actian's Ole Olesen-Bagneux and LexisNexis Risk Solutions' Manca Vitorino explained why context is not a solved problem, why every business definition needs one accountable owner, how ontologies guide AI agents and where AI can take routine work off governance teams.
By The AI Enterprise Magazine Team6 min read
Key takeaways
- Context isn't solved. Generic models can't infer industry conventions built up over decades, Vitorino said.
- One owner per definition. Many can contribute, but distributed ownership "just diluted accountability."
- Ontologies map relationships. Unlike flat glossaries, they show an agent how a domain fits together.
- AI can draft, people decide. Agents can write metadata descriptions and map data models so teams don't start from a blank page.
- Scale magnifies what's there. Leaders should ask whether they understand what they are scaling and can prove it.
A question as simple as "how is revenue doing?" depends on what the organization means by revenue, over which period, compared with what and for which part of the business. When those definitions vary across teams and systems, AI tools can return confident answers that disagree with finance. That problem framed "Business Context in AI Initiatives: Why It Matters More Than You Think," an AI Connect webinar sponsored by Actian.
Brian Mink, co-founder and president of AI Connect, introduced the session. Jennifer Jackson, Chief Marketing Officer at Actian, moderated a conversation with Ole Olesen-Bagneux, Vice President and Chief Evangelist at Actian, and Manca Vitorino, Director of Data Ecosystem at LexisNexis Risk Solutions. Olesen-Bagneux closed with a short demo of Actian's data intelligence platform.
Is business context a solved problem for AI?
Vitorino, who moved into data leadership from journalism, said the industry has come full circle. After years of focus on data volume and the potential of unstructured data, quality and meaning are back in focus. "Volume for me isn't as important as intent," she said.
She has seen claims that context is becoming a solved problem, and said that may hold for common concepts such as customer, sale or churn. When she asked one of LexisNexis Risk Solutions' heads of data governance whether context could simply be inferred, he replied, "Inferred from where?" A generic model cannot infer conventions and exceptions that an industry built up over decades, she said, which is where domain expertise matters most. Olesen-Bagneux agreed that context is far from solved. The word has become central through terms such as context window and Model Context Protocol, he said, but its real meaning is much broader than a technical category of compute.
Who should own business definitions?
Vitorino said many people can contribute to a definition, and committees can debate it, but one person has to be accountable for reaching the decision, recording it and maintaining it. LexisNexis Risk Solutions found that distributed ownership "just diluted accountability," so it named accountable owners across its data domains.
What does context mean in practice?
Asked by the audience to define context, Vitorino called it "the boundary within which we're answering a specific question." Her example came from ICIS, the chemicals and energy commodity information brand she looks after. For some commodity price questions Mexico counts as North America, while in some trading terms it is Latin America, so the right regional grouping depends on the question being asked. Olesen-Bagneux described context as something closer to reading a room or having a sense of place. Its absence shows up when humanoid robots respond oddly or self-driving cars do not behave as their engineers intended, he said.
How do ontologies differ from glossaries and vector search?
Vector similarity can help a language model find related words, Olesen-Bagneux said, but graph structures that represent ontologies give an agent a formal, objective map of how a domain fits together. That is what lets it act with a deeper understanding of a specific universe of data rather than the general probabilities it learned in training.
Vitorino called ontologies "all about relationships." Glossaries are flat, while ontologies are hierarchical and dynamic and model the interdependencies in an industry. That matters for questions such as how a weakening polar vortex will affect gas storage next month and how a buyer should diversify supply if stocks drop. Answering it takes prices, weather patterns, gas and temperature forecasts and current storage levels, and an ontology helps a model or agent find and combine each strand. She said LexisNexis Risk Solutions is working toward this rather than claiming to have solved it.
Who is to blame when AI and finance give different numbers?
Vitorino first asked why two apparently authoritative answers were allowed to exist. The cause could be data quality, two definitions maintained in separate systems, permissions that hide information from the employee, or a poorly formed question, which she treated as part of data and AI literacy. What she does not want is AI giving a confident answer without enough provenance to show where the number came from. "AI for me is just a tool. Data is the material," she said.
How do you keep glossaries and ontologies current?
Olesen-Bagneux said glossaries only get used if they help, and that maintaining them takes lifecycle management policies, which regulated industries tend to have more of. Because companies everywhere are experimenting with AI, he said, semantic gaps now surface faster than they used to. Vitorino said an ontology in a mature industry should evolve slowly rather than change daily, and expects AI to help detect when a definition has drifted and flag it for review.
Where can AI take work off data and governance teams?
Olesen-Bagneux said governance, data management and engineering teams never had enough people, and AI can now help write descriptions, apply tags, assess quality, detect anomalies, trace lineage and respond to observability alerts. It can even help build the semantic structures used for governance. He rejected the phrase "human in the loop" in favor of a human who is in charge, and suggested people monitoring AI output take regular breaks because attention fades during long stretches of review.
Vitorino described what she called first hurdle analysis. Projects often stall before work begins because someone has to gather evidence and frame the problem. Her teams build agents that draft metadata descriptions and map data models, not to make governance decisions but so people are not "staring at a blank page." An imperfect draft is easier to correct than an empty one, she said.
What did the Actian demo show?
Olesen-Bagneux showed an ontology for a fictional life sciences company in Actian's data intelligence platform, generated with AI and then curated by people. Starting from the definition of a patient, he expanded a graph of related concepts such as allergies. He then queried the same definition from Claude through MCP and asked it to explain "patient" as an ontology, returning the full structure. An agent working with patient data across systems could use the same structure to understand what a patient is, he said. Answering an audience question, he said the platform pulls technical metadata from many systems through connectors and ties it together in a graph, and recommended that teams treat building the ontology as human work that AI can help scale.
What should leaders ask before scaling AI?
Olesen-Bagneux told leaders to respect metadata and semantic structures, be realistic about what AI can do and use it to build those structures faster than was possible a few years ago. Vitorino said scale "just magnifies whatever is already there." Clear definitions, ownership and context put a company in "the land of opportunity," while weak ones mean it is "just amplifying your problem." Her question for leaders was whether they understand what they are scaling and can prove it.
On how much context an AI project needs, Olesen-Bagneux said teams can often do with less than an ontology of the entire enterprise. Vitorino said LexisNexis Risk Solutions builds ontologies around specific customer problems rather than one knowledge graph meant to solve everything, and that deciding which problems are worth solving is often the hardest part.
Frequently asked questions
Why does business context matter for AI?
Questions such as revenue or active customers depend on definitions, periods and comparisons. When those vary across teams, AI can give confident answers that disagree with finance.
Who should own business definitions?
Many people can contribute and debate, but one accountable owner should reach, record and maintain each definition, according to LexisNexis Risk Solutions' Manca Vitorino.
What is the difference between a glossary and an ontology?
A glossary is a flat list of terms. An ontology models the hierarchy and relationships between concepts, giving AI agents a map of how a domain fits together.
Can AI help with data governance work?
Panelists said AI can draft descriptions, tags and data model mappings, assess quality and detect anomalies, while people stay in charge of governance decisions.
This article is based on remarks by Jennifer Jackson, Chief Marketing Officer, Actian (moderator); Ole Olesen-Bagneux, VP and Chief Evangelist, Actian; Manca Vitorino, Director of Data Ecosystem, LexisNexis Risk Solutions at AI Connect webinar on November 11, 2026. More coverage: Webinars.
Filed under: Context, Semantics & RAG, AI Governance & Trust, AI Strategy & ROI
