Ontological Primitives
Entities
Every AI-business configuration is a composition of these entities and the interfaces that connect them.
Primitives
Business
An organization with boundaries, processes, inputs, outputs, and external interfaces. The container.
boundedhas interfaceshas inputs / outputs
Human
A node that reasons, decides, acts, and bears accountability. Can interface with any other entity and the external world.
adaptiveaccountablegoal-directed
Model
The foundational capability. General-purpose reasoning, broad knowledge, and adaptability. Invoked on demand — the engine that powers everything above it.
general-purposeon-demandfoundational
External
The world outside the business boundary — customers, suppliers, regulators, markets, partners. The set of all actors and systems the business interfaces with through its boundary touchpoints.
heterogeneousuncontrolledinterfacing
Composed
Agent
A model situated within the business — given domain knowledge, tools, persistent state, and a role. Not a primitive: an Agent is what a Model becomes through the act of situation. Its degree of autonomy is not inherent — it is granted by humans based on where the agent is positioned and what the role requires.
situatedstatefuldomain-awarerole-bound
Interface Taxonomy
Types of Interfaces
Interfaces are first-class ontological objects. The distribution of interface types determines the class.
H ↔ H
Human–Human
Delegation, collaboration, handoffs, reporting. The default wiring.
H ↔ A
Human–Agent
Direction, delegation, review, approval, escalation.
A ↔ A
Agent–Agent
Orchestration and handoff without human mediation.
H ↔ Ext
Human–External
Sales, support, negotiation. Human at the boundary of the firm.
A ↔ Ext
Agent–External
Automated transactions, conversations, API exchange. Agent at the boundary.
H → Biz
Governance
Oversight, constraint, audit. Humans setting boundaries on the business and its agents.
The Six Classes
From Traditional to AI-Autonomous
Each class is a distinct structural relationship. The diagrams show the essential pattern — the entities, the interfaces, and their position relative to the business boundary.
Class 0
Traditional
"No AI in the loop"
An all-human network. People hold every role, every interface, every boundary touchpoint. This is the structure AI enters.
Remove AI → Nothing changes.
Class 1
AI-Assisted
"AI on the side"
The human network is unchanged. Individuals invoke AI products — ChatGPT, Claude, Copilot — on demand. Powerful, but external to the network. No interfaces change.
Remove AI → People slow down, nothing breaks.
Class 2
AI-Augmented
"AI in the team"
The model is situated — given a role, domain knowledge, tools, and persistent state. It becomes an agent: a participant in the network. The human directs it. The H↔A interface emerges. Humans may grant the agent some autonomy, but the structure doesn't require it.
Remove AI → Capacity drops, humans can cover.
Class 3
AI-Extended
"AI at some doors"
The agent takes over some boundary interfaces. Because it now faces the outside world without a human mediating, humans must grant it autonomy — the position requires it. The human retains other boundaries and remains the principal.
Remove AI → Some channels go dark, core survives.
Class 4
AI-Operated
"AI runs the shop"
The agent holds all boundary interfaces and runs operations. Autonomy now extends inward — the agent directs workflows. The human is still inside the machine: managing, configuring, handling exceptions. The knowledge of how to operate the business remains with the humans.
Remove AI → Operations halt. Humans retain the knowledge — they can rebuild.
Class 5
AI-Autonomous
"AI is the business"
The agent network and the business are coextensive. The agent network is self-coordinating. The human is outside the machine: setting strategy, defining constraints, auditing outcomes — but not managing or configuring. The operational knowledge lives in the agents, not the humans.
Remove AI → The entity ceases to exist. The knowledge of how to operate it is gone.
Operated vs Autonomous — The Critical Distinction
AI-Operated: Humans are inside the machine. They manage agents, handle exceptions, reconfigure workflows. Remove the AI and operations halt — but the humans retain the knowledge of how the business operates. They could rebuild it.
AI-Autonomous: Humans are outside the machine. They set strategy and audit outcomes. The agent network is self-coordinating. Remove the AI and the knowledge of how to operate the business is gone — no one knows how to do the work manually.
The test: "Do the humans in this company have the knowledge to operate it without the agents?"
Operated → Yes. Operations halt, but humans know the work. They can rebuild. Autonomous → No. The operational knowledge lives in the agent network.
In Conversation
"We're Traditional — no AI in the loop. Everything runs on people, process, and spreadsheets."
"Most enterprises are AI-Assisted — same org, same process, people just use copilots on the side."
"We became AI-Augmented when we gave agents their own tasks on the sprint board."
"We're AI-Extended — agents handle support and procurement, but sales is still all human."
"That startup is AI-Operated — agents run everything, a few domain experts manage them."
"A true AI-Autonomous company doesn't need an AI strategy. It needs a governance strategy."
Key Distinctions
Ideas That Underpin the Framework
Situation — The Key Transformation
Situation is the act of placing a Model within a business — giving it a role, persistent context, domain knowledge, and tools. It is the ontological operation that produces an Agent. An Agent is not a different kind of AI — it is a Model that has been situated.
This is not a promotion or an upgrade. The model's capabilities don't change. What changes is its relationship to the organization: it gains a position in the network, a defined scope, and an identity within the business. It goes from external utility to embedded participant.
The boundary between Class 1 and Class 2 is this act of situation. Before it, AI is a tool people use on the side. After it, AI is a node in the organizational graph.
Autonomy is Granted, Not Inherent
An agent doesn't start autonomous. In Class 2, it's primarily directed — the human tells it what to do. Some agents may be given autonomy here (e.g. background tasks, scheduled routines), but the structure doesn't require it. As the business places agents in more exposed positions, humans must grant them greater autonomy — an agent at a boundary interface (Class 3) needs to act independently because there's no human mediating that interface. The architecture tells you where autonomy is needed; humans decide to grant it.
Class 2 — Agent is directed. Autonomy may be granted, but isn't required.
Class 3 — Agent is given autonomy at the boundary. The position requires independent action.
Class 4 — Agent is granted broad autonomy. Directs workflows, humans handle exceptions.
Class 5 — Agent network is designed to self-coordinate. Humans govern, not manage.
A Note on Scheduled Autonomy
There is a category of agents that exhibit autonomy through scheduling — given a heartbeat, waking at intervals to poll for conditions, execute routines, or perform maintenance without human prompting. This is a real form of autonomous behavior.
However, this framework deliberately excludes polling or schedule-based autonomy. The classes here describe relational autonomy — autonomy that emerges from an agent's position in the organizational network — not temporal autonomy driven by a cron job or heartbeat. A scheduled agent can run inside any class from 2 upward. It doesn't change the class.
The progression tracks three shifts: what the AI is — from foundational engine, to situated participant, to operational core; where it sits — from outside, to inside, to coextensive with the boundary; and what is required — autonomy is not spontaneous, it's granted to meet the demands of position. Place an agent at a boundary and it needs to act independently. Make it the core and it needs to self-coordinate. The architecture tells you where autonomy is needed; humans choose to grant it.
Prepared by Alex Collet & Akhil Aryanv9 · February 2026