--- title: "AI Ontology: Bottom-Up Doesn't Mean Starting From Scratch" url: "https://atlan.com/context-and-chaos/issue/ai-ontology-bottom-up-doesnt-mean-starting-from-scratch/" description: "We mined 32,237 claims into a business ontology. Every claim carried a source and a confidence score. Not one could say when it stopped being true." keywords: "Ontologies, Context Engineering, Knowledge Graphs, Data Governance" --- > 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/ A Context & Chaos deep dive by Emily Winks (Data Governance Expert, Atlan), published September 4, 2026 (13 min read). It reviews a design, not a working system: an experiment to induce a business ontology bottom-up from a company's own data, which recorded where every definition came from and how strongly the evidence supported it, but had no way to record when that definition was valid. Every claim carried a source and a confidence score. Not one could say when it stopped being true. ## Why bottom-up - The author came to data from library and information science, where the ontology problem ran the other way: models built in the abstract that nobody could reconcile with a real database. An ontology describes how a business is supposed to work; the data records how it actually ran. - So the team worked strictly bottom-up: no hand-authored schema, no seeded business concepts, everything from systems the business already used, designed for agents to consume first and people second. Nobody consulted a standard or a specialist on the way in. - "We over-corrected." Bottom-up answered where definitions should come from; the team treated it as an answer to how they should be represented too. ## The failure we were trying to correct Ask an agent for last quarter's win rate and it answers with complete confidence, but inside a company there is usually no single definition. One team computes closed-won over all closed opportunities, another excludes deals that never reached qualification, a third counts only new-business pipeline. The definitions sit in analyst SQL, dashboard formulas and a deal-desk playbook, and disagree without any being obviously wrong. The goal: mine the definitions in use, keep each attached to its source, score the evidence, and show disagreement instead of quietly reconciling it. Contested cases go to a person who can approve, change, split or reject them. Column-level lineage fills in which columns a mined formula runs on. ## What the experiment actually produced - Pointed at four internal systems: a cloud warehouse, a CRM, a BI tool, and Atlan's own metadata catalog. - Yield: **32,237 claims**, reconciled into **14 entities, 79 metrics, 61 properties and 35 relationships**. - Output lands in two places: an exploration tool where the candidate ontology is browsable as a graph, and the catalog the claims were mined from, where anyone can edit results like any other entry. The exploration tool is closer to the intended destination; getting there means learning from real use what people remove, edit and find useful. - Some of it runs today; the parts meant to close the loop do not. It has not been implemented or tested as a system and is still in the design phase (the "[squiggly part](https://thedesignsquiggle.com/about)"). It is early enough to throw any of it out. ## Three old arguments already in the design The design was never checked against ontology engineering's unsettled questions. Read back against them afterwards, the mapping is uncomfortable. (Kurt Cagle made the broader case in [Four Arguments Ontologists Never Finished](https://atlan.com/context-and-chaos/issue/four-arguments-ontologists-never-finished/).) - **Open and closed worlds.** The design keeps competing claims during discovery, then narrows what can be served through human review, which is a formal implementation of neither. It settles the governance question underneath both (when several definitions are defensible, who decides which one an agent may use?) without drawing on the literature. - **Reification.** Treating a claim as an entity you can make statements about. The claim records are close in spirit to RDF-star, now in RDF 1.2, whose own illustration of what to attach to a statement is its source, confidence level and a timestamp. The team arrived at a similar mechanism without the standard, and took the source and the confidence. - **Partitioning and authority.** The design says what is in scope and logs what falls outside it: a boundary with no source authority, no access rules and no query-time visibility. It resembles part of a named graph. Named graphs are now under consideration, but what their boundaries mean for authority and access would still need defining. ## The missing dimension was time Every claim carries a source. None carries a validity period: no validity intervals, no effective dates, no supersession. A win-rate definition is mined, scored, reviewed and served, and nothing records the period it applied to. - Detecting that a source changed is not the same as representing the interval during which the previous definition was valid. Without it, a question about last quarter silently gets today's definition. - Reviewer feedback could make it worse: scoring weights adjust from earlier decisions, so if an approved definition can expire but its approval cannot, judgments about a past state of the business keep steering the current one. - The fix may be small: claims already carry evidence and confidence; a validity date may need only two dates and a rule for who sets them. - Open question: has the team built a machine that knows when to stop trusting its own past judgments? ## Did we build an ontology at all? - Merging variant names under one canonical entity resembles part of **authority control**, where an authority record holds the established name, its variants, and the source and date that established it. The design has the first two; the third (provenance and time) is missing. - Standards on the roadmap come from the same tradition: **SKOS** (concepts, hierarchies, labels; its preferred vs alternate label resembles the merge logic), **BARTOC** (a registry of thesauri, ontologies and classifications catalogers have used for years), **Dublin Core** (creator and date are among its original fifteen elements). Provenance and time were not new requirements. - The harder criticism: an ontology in the field's sense carries axioms and constraints a reasoner can check and that fail when violated. **OWL** exists to express them, **SHACL** to validate against them. The output has neither: entities, metrics, properties and relationships, documented and traceable, with nothing enforcing a single rule. - The honest description may be smaller: a careful account of tables, joins and the metrics computed on top. Useful, and a lesser claim than "ontology". If agents are meant to reason over it, the gap matters a great deal. - Outside reviewers: some called it a reasonable starting point; one ontologist said flatly it should not be called an ontology, and the author did not disagree. - Underneath: do people want an ontology because they have something to do with it, or because everybody decided they should have one? ## What should bottom-up inherit? The bottom-up commitment was not the error: no inherited ontology would say how a particular deal desk computes win rate. The error was letting the answer to "where does meaning come from" settle "how is it represented". The hard part is how much form to inherit before the inherited model starts deciding what the system is allowed to find. The most common question from experienced practitioners, when something should be an entity versus a metric or property, is the inheritance problem in miniature: adopting an industry model that treats win rate as a metric on an opportunity decides something about a business that may treat it as an object with an owner and revision history. Principle from the closing diagram: discover meaning locally; do not reinvent how meaning is represented. Local evidence supplies meaning, definitions, variants and usage; established form supplies time, authority, provenance and constraints. The author asks ontology practitioners where this distinction breaks down. ## About the author Emily Winks is a data governance expert at Atlan who came to data through library and information science. Over 18 years she has worked as a librarian, information architect and data product architect, including at the New York Public Library and WeWork, and now works at the intersection of data governance, enterprise data management and context for AI. [LinkedIn](https://www.linkedin.com/in/emily-winks/). ## Related reads - [Four Arguments Ontologists Never Finished](https://atlan.com/context-and-chaos/issue/four-arguments-ontologists-never-finished/) (May 2026) - [Ontologies, Context Graphs, and Semantic Layers: What AI Actually Needs in 2026](https://atlan.com/context-and-chaos/issue/ontologies-context-graphs-and-semantic-layers-what-ai-needs-in-2026/) (January 2026) - [Conceptual Modeling Is the Context Engineering Nobody Is Doing](https://atlan.com/context-and-chaos/issue/conceptual-modeling-is-the-context-engineering-nobody-is-doing/) (April 2026) Context & Chaos is Atlan's community newsletter on context engineering, governance, architecture and discovery. Browse all articles at https://atlan.com/context-and-chaos/ or pitch a piece at https://atlan.com/context-and-chaos/contribute/.