Notifications
Clear all
Topic starter
23/06/2026 6:59 pm
TL;DR: LLMs can summarize and guide, but they still operate on representations of systems rather than the governed systems themselves, according to Collibra. That gap becomes operational risk when AI agents choose data or trigger workflows without trusted context, because plausible outputs can still be the wrong operational decision.
NHIMG editorial — based on content published by Collibra: AI needs context, why LLMs still do not understand real-world systems
Questions worth separating out
Q: How should teams govern AI systems that choose data in real workflows?
A: Teams should govern AI data choice the same way they govern privileged access: by defining authoritative sources, exposing certification and lineage metadata, and enforcing policy checks before execution.
Q: Why do AI assistants create risk even when access is technically approved?
A: Because access approval does not guarantee contextual correctness.
Q: What breaks when metadata is not available at decision time?
A: Decisioning becomes pattern matching without governance.
Practitioner guidance
- Classify authoritative context sources Identify which datasets, workflow states, and policy records are authoritative for each use case, then mark the rest as non-authoritative even if they remain technically accessible.
- Expose governance metadata to AI decision paths Publish lineage, certification, ownership, and policy constraints in forms that AI tools and agents can query before selecting data or triggering actions.
- Add contextual checks before workflow execution Require AI systems to validate dataset freshness, policy applicability, and approval status before they can invoke downstream processes.
What's in the full article
Collibra's full blog post covers the operational detail this post intentionally leaves for the source:
- How Collibra maps governance metadata to data, models, and processes inside AI workflows.
- The article's own explanation of how the Model Context Protocol and Collibra MCP server expose trusted context to AI systems.
- A stepwise view of how certification, ownership, and policy constraints can be embedded into operational decision paths.
- The vendor's final framing of controlled system intelligence and why context drift becomes the next governance problem.
👉 Read Collibra's analysis of why LLMs still need governed context →
LLMs in real workflows: where context-aware governance fails?
Explore further
25/06/2026 2:45 am
Governed context is now an identity control problem, not a documentation problem. The article is correct to separate language fluency from operational correctness, but the deeper lesson is that enterprise AI fails when it cannot identify which data, policy, or workflow state is authoritative. That is not just an information architecture issue. It is a governance boundary problem that affects human users, NHIs, and AI systems alike. Practitioners should treat context trust as part of access control, not as a downstream convenience.
A few things that frame the scale:
- 80% of organisations report their AI agents have already performed actions beyond their intended scope, including accessing unauthorised systems (39%), inappropriately sharing sensitive data (31%), and revealing access credentials (23%), according to AI Agents: The New Attack Surface report.
- Only 52% of companies can track and audit the data their AI agents access, leaving 48% with a complete blind spot for compliance and breach investigation, according to SailPoint research.
A question worth separating out:
Q: How do security teams keep AI from acting on stale context?
A: They should require freshness, certification, and policy checks before the system can select data or trigger workflows. Stale context is a governance issue, not just a data quality issue, because it can drive compliant-looking actions that are operationally wrong. The practical control is to validate context at the point of use, not only at publish time.
👉 Read our full editorial: AI needs governed context, not just better language outputs
