Your LLM Does Not Know Your Business

Technical thesis

A language model can know an extraordinary amount about language, software, industries, and the public world while knowing almost nothing about the current operational reality of a particular organization.

Model knowledge and organizational context are different assets. Treating them as interchangeable creates systems that sound informed while remaining uncertain about identity, authority, state, and intent.

REPLACEABLE CAPABILITYLanguage modelGeneral patterns · inference · generation

operates against→

DURABLE ORGANIZATIONAL CONTEXTCanonical identityEvidence + provenanceAuthorityHistorical stateRegistered actions

What a model can know

A model can encode general patterns learned during training. It can interpret text, generate plans, classify inputs, translate terminology, call tools, and reason over context supplied at runtime. Specialized models can perform narrower tasks with greater efficiency or control. External frontier models can provide capabilities that are uneconomical to reproduce locally.

Those are valuable forms of computational capability.

They do not tell the model which of two customer records refers to the same legal entity, whether a repository reflects approved intent or deployed state, which policy revision applied to a past decision, who may authorize a production change, or whether an emergency exception is still active.

These are not gaps that more parameters reliably solve. They are properties of the organization’s current operating environment.

Organizational context is governed

Organizational context includes more than facts placed in a prompt. It includes:

• canonical identities that survive application boundaries;
• relationships whose meaning and temporal validity are explicit;
• evidence with provenance and integrity;
• source authority that varies by fact and context;
• governed semantic definitions;
• current, historical, proposed, and unavailable state;
• permissions, policy, and human authority;
• registered actions and their reversibility;
• decisions, approvals, overrides, outcomes, and corrections.

This context is durable. A model invocation is temporary.

The enterprise therefore needs a Knowledge Network and operating model that exist outside the model. Different models, deterministic functions, statistical methods, humans, workflows, and applications can then operate against the same organizational understanding.

The prevailing approach

The usual response to a model’s lack of company knowledge is to provide more context. A system may place instructions in a system prompt, retrieve documents, copy database rows, summarize prior conversations, or allow the model to query tools.

These techniques are useful. They help a model perform a particular computation. They do not by themselves govern what the supplied information means.

A document saying a customer is “strategic” might represent a policy definition, a salesperson’s assertion, an obsolete plan, or an AI-generated summary. A production CRM field may be authoritative for the current sales classification but not for contract status. A ticket can establish intended work without proving that the change reached production.

If all sources become undifferentiated context, the model is asked to infer organizational authority from prose. That is a fragile assignment of responsibility.

Why prompts are not institutional memory

Prompts are execution artifacts. They assemble instructions and context for a workload. Even when they are versioned, they are not a complete history of organizational understanding.

Institutional memory must preserve what the organization believed, the evidence available at the time, the definition in force, the authority under which a decision was made, the action that followed, and the outcome that challenged or supported it.

Suppose the definition of ReadyForOnboarding changes after several failed onboarding attempts. A durable system does not overwrite version one with version two. It connects the earlier definition to the decisions made under it, the evidence used, the observed outcomes, the contradiction, the authorized revision, and the future decisions made under the new definition.

That epistemic history belongs to the organization, not to the model provider and not to a chat transcript.

The CTN and EIE position

Models are replaceable intelligence engines. The Intelligence Runtime selects an appropriate capability according to policy, capability, privacy, cost, latency, continuity, and evaluation. It assembles governed context, controls tools, captures evidence, and applies fallback and approval rules.

The model does not become the source of truth because it produced a fluent conclusion. AI interpretation remains inference. The system weighs the evidence, records conclusion confidence separately from source trust, and preserves disagreement.

This produces an important control property: enterprise intelligence survives model changes because organizational context and authority exist outside the model.

Practical implications

Before giving a model access to enterprise data, ask:

• Does the system know what each record represents?
• Are source roles and authorities explicit for the fact at issue?
• Can it distinguish observed, proposed, asserted, derived, and inferred state?
• Does it preserve time, provenance, contradictions, and unavailable information?
• Can the model recommend an action without automatically receiving authority to execute it?
• Will a later model be able to reconstruct why the organization reached this conclusion?

The correct response to “the model does not know our business” is not simply to give it more documents. Give intelligence a governed operating model of the organization.