Your AI model is temporary. Your organisational understanding should not be.
Current business context makes an agent useful now. Durable, model-independent organisational understanding keeps the business operable when models, people and providers change.
DataInbox
Architecture & Research
ARC Prize recently published an analysis of how OpenAI's GPT-6 Astra approached ARC-AGI-3. The benchmark places an AI system in unfamiliar interactive environments and tests whether it can explore, understand the situation, form goals and act successfully.
One observation matters far beyond the benchmark. Astra turned unfamiliar environments into compact symbolic world models. It described objects, coordinates, state, rules and possible actions. When the harness allowed it, the model also created parsers, state trackers and planning tools to make that understanding operational.
ARC Prize does not make an enterprise architecture claim. The following is our inference: useful agents need more than access to documents or a large context window. They need a coherent representation of the world in which they are expected to act.
For a business, that world model is its organisational understanding.
Current business context is necessary, but not sufficient
An agent may know the latest customer message, active order, open case or consent state. That is current business context. It answers an essential question:
What is true now?
But knowing what is true does not yet tell the agent what it means or what the organisation permits. It must also understand:
- What a customer, order, case, obligation or business event means.
- Which sources and relationships are authoritative.
- Which policies, purposes, permissions and exceptions apply.
- Which people, systems or agents may perform an action.
- Which outcome is required and what evidence must remain available.
That is organisational understanding. It answers a different question:
How may this organisation interpret and act on what is true now?
AI needs both. Context without organisational meaning produces plausible guesses. Documentation without current state produces informed but outdated answers.
Most documentation stops where the operation begins
Companies already document policies, responsibilities, data definitions, processes and architecture. The problem is not the complete absence of documentation. The problem is that documentation usually stops at the edge of the runtime.
A policy describes what should happen. Application code determines what actually happens. A process diagram shows the intended route. Integrations and workflow configurations implement another route. A data definition explains a field. Each application may interpret that field differently.
The gap grows every time the business changes faster than its software estate. AI makes that gap more consequential because an agent can now interpret information and take action at machine speed.
If the organisational model exists only in documents, people must continually translate it into prompts, code, workflows and access rules. If it exists only inside one AI provider's context or memory, the organisation has made its own understanding dependent on a temporary model.
Documentation should become infrastructure
Documentation becomes infrastructure when the same controlled meaning that explains the operation also participates in the operation.
It should be possible to use organisational definitions to:
- Validate an incoming business message.
- Resolve which context belongs to a situation.
- Determine which rule or policy applies.
- Limit the data and capabilities available to an agent.
- Define acceptable actions and result contracts.
- Preserve the evidence needed to explain an outcome.
This does not mean every paragraph becomes executable code. It means durable knowledge has an explicit role, owner, scope and lifecycle, while the operational parts become machine-interpretable conditions that can be inspected before and after execution.
The explanation and the execution no longer drift as two unrelated worlds.
A model-independent operating architecture
A durable agentic business needs three distinct layers.
1. Durable organisational knowledge
Definitions, policies, responsibilities, decisions and unresolved questions must remain owned by the organisation. They should not disappear when a consultant leaves, a model is replaced or a prompt history is cleared.
Canonrail provides a way to govern this durable project and organisational knowledge. It makes the role, owner, lifecycle, scope and open attention of documentation explicit. It does not run source-system operations or grant runtime permission.
2. Current governed business state
DataInbox receives business messages and maintains the current state, history, objectives, commitments, permissions and outcomes needed for a specific operating context. Deterministic rules establish what is permitted before intelligence is invoked.
This gives people and agents a governed projection of the situation instead of unrestricted access to every source or repeated reconstruction from documents and transcripts.
3. Replaceable intelligence
Approved AI models and agents can interpret questions, propose changes and perform authorised work. They use the organisation's meaning and current state, but they do not own either.
The model can improve, change provider or disappear. The business does not have to rebuild its memory, authority and operating logic around the replacement.
The model replacement test
There is a practical way to assess whether organisational understanding is truly independent.
If you replace the AI model tomorrow:
- Can the new model discover the same business entities and relationships?
- Can it inspect the same current state without rebuilding memory from conversation history?
- Can it see which policies and permissions constrain every action?
- Can people inspect the same definitions without reading provider-specific prompts?
- Can the organisation explain which approved version and authority produced an outcome?
If those answers disappear with the model, the organisation has rented intelligence but has not built durable understanding.
The organisation is the world the agent must understand
The most important lesson from Astra is not that a model can play unfamiliar games. It is that capable action follows from a compact representation of objects, state, rules, tools and goals.
A business is not a closed game. Its environment includes people, regulations, incomplete information, changing responsibilities and consequences in the real world. Its organisational model therefore cannot remain an improvised prompt or a hidden layer of application code.
Current business context makes an agent useful for the situation in front of it. Documentation as infrastructure makes that understanding durable, inspectable and transferable. DataInbox connects both to permitted action.
The AI model is temporary. The organisation's understanding should survive it.