← Back to the journal

DataFocus · Journal

Business Ontology for AI Agents: What It Is, What It Contains, When Not to Build One

A business ontology models one business area's objects, relationships, rules, actions and permissions in one model people and agents share. Agents need it to close three gaps a schema cannot: context, boundaries and an approval path. Plus a rule for when not to build one.

A business ontology is a maintained model of one business area: the objects it works with, how those objects relate across the systems that hold them, the rules and metric definitions currently in force, the actions that may be taken on them, and who may see and do what. An AI agent needs one because a database schema answers only "where is the data." It does not say which of three "Customer A" records is the same company, what "delivered" means in this plant, whether this user may see cost prices, or how a correction gets made and approved. Those are the three things a model lacks inside a company: context, boundaries, and a way to act with a person in control. The ontology supplies all three from one model that people and agents share, which is why it belongs under the agent, not inside its prompt.

Two meanings of "ontology", and the one that matters for agents

The formal meaning is the W3C's. The OWL 2 Primer, a W3C Recommendation dated 11 December 2012, describes OWL 2 as "an ontology language for the Semantic Web with formally defined meaning": classes, properties and axioms, built for inference.

The operational meaning comes from platforms. Palantir's Foundry documentation (Ontology core concepts, as documented in November 2026) calls the Ontology "a categorization of the world" and defines an object type as "the schema definition of a real-world entity or event." Applications, and now agents, run on the model. DataFocus uses the word this way. The "OT" in DataFocus OT stands for Ontology, and its product brief defines one as "A model of your business, covering its objects and how they relate, the rules and definitions you work by, the actions that can be taken, and who may see and do what. People and AI work on the same model."

For agents the operational sense is the useful one, and the sentence to hold onto is the last one. The agent is not handed a copy of the model or a prompt that paraphrases the data dictionary. It works on the same object, rule and permission a person sees on a control panel.

What the model contains

The DataFocus brief scopes an ontology to one business area and uses its built-in purchase-approval ontology to show five parts. The brief's caption puts that graph at 147 nodes and 228 relationships over nine tables.

PartWhat it holdsIn the built-in purchase-approval ontologyQuestion it answers for an agent
Objects and relationshipsTables from different systems, how they relate, what each relationship means, shown as an ontology graphPurchase requests, approval events, process variants, bottlenecks, compliance findings, improvement tasks, audit recordsWhich table is the customer, and how does it join to the order?
Rules and knowledgePolicies, rules and metric definitions in a document center, each linked to the data it governsCompliance rules, SLA and bottleneck rules, event-log definitions, the write-back and audit procedureWhich rule is in force, and against which data is it evaluated?
IndicatorsMeasures delivered to named people on a scheduleScheduled indicator deliveryWhat is being watched, and by whom?
ActionsExecutions (scripts that calculate, check and write back) and business actions on control panelsRe-run mining, create a corrective task, request evidence, close a deviation, reopen itWhat may I do here, by name?
PermissionsWho may see and do whatExisting users, roles and resource permissions; agents granted one ontology at a timeWhat may I see, and on whose authority?

Two things are absent. There is no new database: the ontology sits over tables that already exist and attaches meaning, rules and controls to them. And there is no separate "AI layer": every ontology has its own agent, answering from its data and documents with evidence. Every entry in the last column is a question an agent on raw schemas answers by guessing; the next section shows what that costs.

What a schema cannot tell an agent

The DataFocus brief frames the problem in one line: "What AI lacks in the enterprise is not intelligence." Each of the three gaps has a precise shape and independent evidence.

Context. A schema names a column cust_id. It does not say whether that is the CRM's account_id, or which version of a policy applies this quarter. That knowledge lives in documents the query engine never reads. Benchmarks measure the cost. Spider 2.0, 632 real-world enterprise text-to-SQL workflow problems over databases often containing more than 1,000 columns, reported that an o1-preview code agent framework solved 21.3% of tasks, against 91.2% on the older academic Spider 1.0 and 73.0% on BIRD (Lei et al., arXiv, 2024, revised 2025). The gap has outlasted a model generation: BEAVER, built from private enterprise warehouses (9,128 question-SQL pairs, 812 tables), reports state-of-the-art agentic frameworks using GPT-5.2 at 10.8% accuracy, rising to 30.1% when the agent is given subtask hints (arXiv, 2024, revised May 2026). Models that saturate academic schemas fail on enterprise ones; hints about business structure are what move the number.

Boundaries and action. A model has no native notion of your roles or data classification; it sees whatever the connection string can see. The OWASP Top 10 for LLM Applications 2025 names the consequence twice: Sensitive Information Disclosure (LLM02) and Excessive Agency (LLM06). OWASP traces Excessive Agency to three root causes, excessive functionality, excessive permissions and excessive autonomy, and lists human approval of high-impact actions among the mitigations. Those root causes map onto three parts of the table above: named actions bound functionality, permissions that travel with the model bound permissions, and an approval step bounds autonomy.

A worked example: "Customer A" across three systems

Consider a hypothetical mid-size distributor (illustrative; nothing here was run against a live system). Its CRM holds account A-10492, "Acme Industrial Ltd". Its ERP holds customer 300871, "ACME INDUSTRIAL". A signed framework agreement, stored as a PDF, names "Acme Industrial Limited". A manager asks an agent: "What is Acme's open order value?"

On raw schemas, the agent must decide on its own that three records are one company, guess the join key, then guess which of six order statuses count as "open." Guess well and the number is right; guess badly and it is confidently wrong, and nothing in the answer tells the manager which happened.

On an ontology, both decisions were made once, by the people who own them:

  1. The data owner confirmed that A-10492, 300871 and the contract's legal name are one Customer object, linked by those keys.
  2. The order-to-cash owner defined "open order" as a specific status set, linked to the orders table. In DataFocus a definition set once is used by search analysis, dashboards, control panels, the ontology agent and external agents, so the agent's figure matches the dashboard's.
  3. A fourth record with a similar name stays separate. DataFocus's homepage demo states the inverse rule: until a customer identifier mapping across CRM, ERP and contracts is confirmed, agents do not treat two similar names as one company.
  4. The answer comes with its working. In DataFocus OT every query returns the SQL that ran, the tables it used and a data fingerprint, so a reviewer can rerun it rather than take it on trust.

Published evidence says this layer helps. A 2023 data.world benchmark on an enterprise insurance schema found GPT-4 with zero-shot prompts reached 16% accuracy querying the SQL database directly and 54% when the same questions were posed over a knowledge graph built from an ontology and mappings (Sequeda, Allemang and Jacob, arXiv, 2023). Its limits: one schema, a GPT-4-era model, 2023. It shows direction, not a number to expect. The mechanism is what carries over: the model stopped guessing joins and definitions because the layer had already fixed them.

Rules and actions belong in the same model as the data

A definition in a wiki never reaches the query. If "on-time delivery" is defined in a policy document and computed three ways in three dashboards, an agent will pick one, and you will not know which. DataFocus links each policy, rule and metric definition to the data it governs: in the purchase-approval example, the SLA rule that defines a bottleneck is attached to the approval-event data, so when the agent explains why a request is flagged, the rule and the data come from one model.

Rules also carry versions. In the DataFocus homepage demo on simulated data, a support policy changes from a 24-hour to a 4-hour response time, effective midnight on 25 September. Until the owner approves the change, agents keep answering with v1; after approval, new queries use v2, and answers given under v1 stay on record, unrewritten. Effective time and approval time are different facts; an agent that conflates them misreports history.

Actions follow the same principle, and OWASP's first root cause explains why. A database connection is excessive functionality by construction: it permits every write the credentials allow. The purchase-approval ontology instead exposes five named business actions, each with a defined meaning and effect, and together they are the complete list of what can be done. That is a smaller and more honest surface than "the agent will write the right UPDATE."

Autonomy is bounded last. When AI in DataFocus OT wants to change something, it can only submit a draft, which takes effect after a person approves it. The change path runs draft, diff, validation, approval, commit and read-back, and the approval binds a specific plan and base version: if either changes, approval must be given again. In the homepage's synthetic-data demo, a coding agent proposes a draft of four object types, three relationships and one definition it cannot settle, stops to ask, reports "Source writes 0 · access unchanged," and cannot approve its own change. This is OWASP's human-in-the-loop mitigation as a property of the model rather than a line in a prompt.

Permissions and receipts: the intersection rule

This is where "people and AI work on the same model" becomes a security property. In DataFocus OT, ontology permissions reuse the organization's existing users, roles and resource permissions, and agents are granted access one ontology at a time. An agent gets only what the authorizing user may see. Requests outside its grant do not reveal whether the data exists. Withdrawing access takes effect at once.

The effective scope is an intersection: the authorizing user's permissions, narrowed to the ontologies the agent was granted. A purchasing agent granted the purchase-approval ontology cannot reach HR data even if the user who authorized it could. The homepage demo shows the refusal shape: "No order O-1003 was found in your authorized scope." Not the content, not a count, not whether O-1003 exists. Identity sets the visible scope; the wording of the question does not.

Receipts close the loop. Every call records sources, definitions and returned scope. Every change records before and after, the diff and who approved it, and a version change forces a new review. Every query has a receipt, every change has an approval record, and every agent action is recorded in an audit log. If you work to the NIST AI Risk Management Framework, voluntary and organized around Govern, Map, Measure and Manage, these records are the raw material for Measure and Manage. You cannot review agent behavior you did not record.

When not to build one

An ontology is a commitment: someone owns the objects, confirms the definitions and approves changes, and that cost is not worth paying everywhere. It records decisions rather than making them; if finance and sales disagree on "revenue," modelling it first encodes the argument. A model nobody maintains goes stale, and a stale definition is worse than none, because agents trust it and cite it. Analysts exploring new data need freedom, not governed objects.

Three tests, all of which should be yes before you start:

TestAskIf no
ActionWill agents change anything here, or only read?A semantic layer or curated views may suffice
SpanDoes the area cross more than one system of record?The schema plus a few definitions may be enough
OwnerCan you name the person who confirms definitions and approves changes?Do not model it yet; it will rot

Purchase approval, order-to-cash and production exceptions usually pass all three; an ad-hoc marketing analysis usually fails the first and third. The DataFocus brief's suggested first step follows the same logic: pick one business question with data, policy documents and a clear owner, build that one ontology, get business users working on it, then authorize an agent, then move to the next area.

Next step

The OT product brief describes the five parts in full, with the purchase-approval ontology and the SQL receipt behind each answer. Both editions are free; the download and editions page explains how OT Flash and OT Heavy differ and links the installers and release notes.