Open ontology · Version 2.0
Agentic Ontology of Work
A platform-agnostic semantic model for enterprise work performed by AI agents, people, and systems. AOW defines the objectives work serves, who performs it, the rules and level of autonomy that apply, the results it produces, and how outcomes are fed back. It is published in formats that software can validate.
Purpose
Terms such as "agent," "task," "workflow," and "orchestration" are used inconsistently across vendors and teams. This makes agentic systems difficult to integrate, govern, and audit, particularly where work passes from one system to another.
AOW addresses this with a formal ontology. In addition to defining terms, it specifies the attributes of each entity and the relationships between entities. For example, an Intent serves an Objective, a Result is generated by an Action, and an Action is performed by an Actor under a Policy. Because these rules are formal, conformance can be verified by software.
Contents
AOW 2.0 defines 25 classes and 33 relationships in four layers. The model links each business Objective to the Intents, Plans, and Tasks derived from it, the Actions that agents and people take using Skills, and the Results and Outcomes produced.
Governance is part of the model: Policy, Confidence, a five-level Assurance Level scale, an independent Guardian, and Feedback recorded in Memory.
Model
Classes and relationships
The diagram shows each class by layer. Hover over or select a class with the keyboard to show its relationships. Select a class to open its definition.
Hover over or select a class to show its definition and relationships.
Layer 1
Perception
Records the work in progress and what is observed about it.
Layer 2
Cognition
Defines what should happen and the limits that apply.
Layer 3
Execution
Records who performed the work and what it changed.
Layer 4
Assurance
Governs autonomy, measures outcomes, and records feedback.
Assurance Levels
Five levels of autonomy
An Intent specifies the maximum level of autonomy permitted for its work. A Task or Skill may specify a lower level, and each Agent is cleared up to a specified level. The effective level is the lowest of these. If an Agent's confidence falls below the threshold for the effective level, the Agent must escalate to a person.
AL0
Manual
People: Performs the work.
Agents: Observes and prepares; takes no Action that changes a Work Item.
Work performed by people
AL1
Assisted
People: Decides. Approves every Action that changes a Work Item before it takes effect.
Agents: Recommends, drafts, and prepares.
Every change approved by a person
AL2
Delegated
People: Confirms when confidence is low or a Policy triggers; reviews a sample otherwise.
Agents: Decides and acts within Policy.
Default confidence threshold 0.90
AL3
Supervised
People: Reviews by exception, when the Guardian raises an alert, and in periodic audit.
Agents: Acts, and must be able to reverse or compensate its Actions.
Default confidence threshold 0.95
AL4
Autonomous
People: Sets Policy and monitors aggregate Outcomes.
Agents: Acts without routine human review; the Guardian can suspend it at any time.
Default confidence threshold 0.98
The scale is based on established levels-of-automation research, including Sheridan and Verplank (1978), Parasuraman, Sheridan, and Wickens (2000), and SAE J3016. Whitepaper, section 7
Validation
Validation checks
Every example in the repository is validated against four checks on each change.
- JSON SchemaThe document is well formed and contains no unrecognized keys.
- SHACLThe data is well formed. For example, every Result has an Action, and every relationship points to the correct type of entity.
- TraceabilityEvery Result can be traced to the Objective it served.
- OversightNo Agent acted without the human oversight required by its effective Assurance Level.
The table below shows the traceability check applied to the claims processing example. In the last row, the agent's confidence was 0.81, below the required 0.90, so the claim was escalated and decided by a claims adjuster.
| Result | Performed by | Skill | Task | Intent | Objective |
|---|---|---|---|---|---|
| Monthly benefit of $3,180 calculated. | Benefit agent | calculate_benefit | Determine the monthly benefit. | Automate eligibility determination | Faster disability claims |
| Coverage active under policy form LTD-2024 on the date of disability. | Coverage agent | check_coverage | Validate that the claimant's coverage was active on the date of disability. | Automate eligibility determination | Faster disability claims |
| Eligibility approved. | Validation agent | record_decision | Generate and record the eligibility decision. | Automate eligibility determination | Faster disability claims |
| Eligibility approved after adjuster review. | Claims adjuster 17 | record_decision | Generate and record the eligibility decision. | Automate eligibility determination | Faster disability claims |
Downloads
Downloads
All formats are generated from a single source file, aow.yaml. Terms have permanent identifiers under https://w3id.org/aow.
Whitepaper
Version 2.0 of the whitepaper, including all entity definitions, the Assurance Level scale, and the worked example.
Reference card
All classes and the Assurance Level scale on a single page.
Ontology
For ontology editors, reasoners, and knowledge graphs. Also available as JSON-LD.
JSON-LD context
Allows AOW JSON documents to be read as RDF.
Validation shapes
Validates AOW data in RDF.
JSON Schemas
Validates AOW data in JSON. One schema per class, plus a document schema.
Classes
All classes with definitions. Attributes and relationships are available as separate CSV files.
Version 1.0
The original paper, January 2026.
Related standards
Related standards
AOW reuses or aligns with existing standards where they define the same concept. Whitepaper, section 12 describes each relationship.
Related project
SOA-to-Agentic AI Terminology Mapping
Maps 28 Service-Oriented Architecture terms to their agentic equivalents. A crosswalk maps each of those terms to AOW.
Contributing
Issues and pull requests are welcome, particularly reports of real-world work that AOW cannot describe. Open an issue or see the contribution guidelines.
Citation: Garg, M. (2026). Agentic Ontology of Work (AOW), version 2.0.0. Skan.ai. https://doi.org/10.5281/zenodo.22945905