Context governance is the discipline of controlling everything that enters and leaves an AI system's working context — retrieved documents, memory, tool results, agent messages and generated outputs — by enforcing identity, purpose and sensitivity policy at the moment of use, and recording what was assembled for audit.
What is context governance?
Every AI answer is produced from a context: the set of documents, memory entries, tool results and prior messages assembled and passed to a model at the moment it responds. That context is created at runtime, exists for seconds, is different for every request, and never appears in any data catalogue. It is also where enterprise risk now concentrates. Context governance is the discipline of controlling how that context is assembled, what policy is applied to it, and what evidence is kept afterwards.
Why data governance is not enough
Traditional data governance operates on data at rest. It answers questions about ownership, quality, retention and who may open a file. Those controls remain necessary, but they assume that risk lives in the file and that access is the moment of decision. AI breaks both assumptions. A model can retrieve twelve permitted fragments and synthesise a conclusion that no one is cleared to see. An agent can carry a finding from a privileged session into an unprivileged one through memory. A tool call can move data across a boundary that no access review ever examined. In each case the access log looks clean, because no rule was broken at the file level. The exposure was created in context.
What context governance controls
Context governance applies to six control surfaces, each of which can independently create or contain exposure:
- Retrieval — which chunks may be selected, redacted, scoped or denied for this identity and this purpose, before the model sees them.
- Assembly — what may be combined in a single context window, given that combination itself creates sensitivity.
- Memory — what an agent may write, recall and carry across sessions, users and roles.
- Tools and actions — which tools may be discovered and invoked, with what arguments, and what must pause for human approval.
- Inter-agent exchange — what context one agent may pass to another, and under whose authority.
- Output — what may be returned, what must be redacted, and what must carry citations and lineage.
The five principles of context governance
- Govern meaning, not just location. Sensitivity is a property of what content means, and of what it means once combined. Policy must be attached to meaning at asset and element level. See semantic governance.
- Decide at runtime, in context. The same content may be permitted, redacted or denied depending on who is asking, why, and what else is already in the context window. Static access lists cannot express this. See dynamic policy engine.
- Bind every decision to an identity and a purpose. This includes non-human identities. An agent acting for a user must not exceed that user's entitlements, and must carry the purpose forward. See purpose-based access control.
- Minimise context. Assemble the least context sufficient for the task. Over-retrieval is the most common and least examined cause of AI exposure. See context minimisation.
- Produce evidence by construction. Every retrieval, policy decision, memory write, tool call and output should emit a structured record without anyone being asked to keep one. See AI audit trail.
How context governance is implemented
Implementation follows a consistent sequence, regardless of vendor:
- Connect and map. Inventory sources, systems, agents and models, including unsanctioned ones. Build the context graph of entities, relationships and lineage. This is the mapping step every governance framework depends on, and it is where an AI-BOM comes from.
- Classify. Generate sensitivity, semantic labels and taxonomy automatically at asset and element level, across structured and unstructured data. Manual classification does not keep pace with AI.
- Author policy. Express rules in terms of identity, role, attribute, purpose and sensitivity, and simulate their effect before enforcement.
- Enforce at runtime. Place enforcement between retrieval and generation, around memory, in front of tools, and on the output path.
- Evidence and review. Capture structured traces, map them to control frameworks, and review policy against what the traces show actually happened.
Context governance and the regulatory picture
Context governance is the operational layer beneath obligations that are otherwise difficult to evidence. GDPR purpose limitation becomes enforceable when purpose is a runtime policy input rather than a paper assertion. ISO/IEC 42001 requires an auditable management system; runtime traces are the evidence that system produces. The EU AI Act requires record-keeping, human oversight and transparency for high-risk systems, all of which are properties of how context is assembled and logged. The NIST AI RMF Map and Measure functions depend on knowing what each AI system consumes.
How Kynexa approaches context governance
Kynexa implements context governance as platform-neutral infrastructure rather than as a feature of one AI application. Context Governance is available today: enterprise SSO and identity federation for people and digital agents, pre-built connectors, automatic metadata extraction at asset and element level, and file processing for AI. The same policy engine governs retrieval, memory, tools and outputs, and emits a structured audit trail across the estate. See the platform.