Click a name to inspect it. Click the triangle to fold a branch. Click the square to add or remove a class from your view — downloads contain only what is ticked.
This is a skeleton class model for autonomous and semi-autonomous agents transacting in commerce settings. It covers who acts, under whose authority, over which offers, with what evidence and under which constraints. Catalogue modelling is left to schema.org and GS1 rather than reproduced here.
It is a draft at version 0.1, presented for community review at the Ecommerce AI Agent workshop, Zurich, 7 September 2026.
Each class sits in one of three warrant tiers. The tier matters more than the class name, because two of these tiers rest on published documents and the third rests on nothing but a drafting session. Click any class in the tree to see which tier it is in and what supports it.
Terms were elicited from public sources by ESL, with stakeholder involvement. No competency questions were written first and no source corpus was sampled, so class coverage reflects the sources listed above rather than a survey of the field.
AP2 splits authorisation into three mandates by transaction stage. This model collapses that into a single Mandate class, closest to AP2's Intent Mandate, and the Cart and Payment mandates have no class here. Anyone crosswalking from AP2 should expect that gap.
The specifications cited have thin deployment behind them. ChatGPT's Instant Checkout, ACP's flagship implementation, was retired in March 2026 after roughly a dozen merchants shipped against it. Citing a specification is not evidence of a settled practice.
EU consumer, payment-services and AI Act obligations are unmapped, though JurisdictionConstraint gestures at them. Without competency questions there is also no test for whether any tier-3 class earns its place.
Most published vocabularies state what they contain and stay quiet about where it came from. The result is that a reader cannot tell a term backed by a W3C Recommendation from one somebody invented on a Thursday, and both get reused with equal confidence. Separating warrant from content is the correction, and the count above is what it produces: a number you can act on rather than a claim to trust.
The same audit runs on any vocabulary, a data dictionary, a taxonomy, a set of reporting definitions, a schema inherited from a vendor. What changes is the domain. What stays is the question of which terms have external warrant, which are local inventions, and whether anyone has written that down.
This work serves as the basis for executive training in knowledge modelling and ontology development, run clinic style: participants bring a vocabulary from their own organisation and audit it the way this one has been audited. Engagements cover ontology development, knowledge audit, and decision quality in AI-driven enterprises where the modelling has been left to vendors.
Enquiries to contentwire@gmail.com.