Semantic layers for AI: why trust depends on governed meaning

At a glance

What this is: NHI Mgmt Group's analysis, drawing on Collibra's blog post: Meaning, not just access, determines whether LLMs and agents produce trustworthy outputs, making a governed semantic layer matter for AI.

Why it matters: IAM, NHI and data governance teams need to treat meaning as a governed control surface because AI systems can be authorised to see data but still act on the wrong interpretation.

By the numbers:

• In an independent KU Leuven test cited by Collibra, the same model answered correctly 92% of the time with governed context in the loop and 62% without it.
• The failure rate dropped from 38.5% to 7.7% when governed context was added, according to Collibra's cited KU Leuven test.

Context

A semantic layer for AI is a governed layer of business meaning that sits between raw data and the AI that uses it. It defines terms, certified metrics, relationships and rules so a model or agent reasons from the organisation's intended meaning rather than inventing its own.

For identity and access teams, the key issue is not whether AI can reach data but whether it can interpret that data under governed rules. Once agents start composing steps and taking actions, meaning becomes a control surface, not just a modelling convenience.

That shifts the governance question from data availability to decision validity. In AI systems, unauthorised meaning can be just as damaging as unauthorised access, because a wrong but plausible answer is often harder to detect than a denied one.

Key questions

Q: How should security teams govern AI systems that rely on business definitions at runtime?

A: Security teams should require certified definitions, approved metrics and policy metadata to be delivered together at runtime. The AI should not be allowed to invent business meaning from raw tables or prompts. Governance should cover lineage, ownership, freshness and allowed use, so meaning and permission stay aligned when the system acts.

Q: Why do LLMs still get the wrong answer when they have access to the right data?

A: Because access to data does not guarantee access to the correct meaning. An LLM can retrieve a table, join related records and still apply the wrong business definition if no governed semantic layer tells it which metric or term is authoritative. The failure is interpretive, not just technical.

Q: What breaks when semantic definitions are inconsistent across AI tools?

A: Consistent access may still produce inconsistent decisions. If one tool treats a metric one way and another tool uses a different definition, the organisation gets contradictory answers, broken trust in reporting and uneven agent behaviour. Governance fails when meaning is not centralised and certified.

Q: Why do AI systems need semantic context layers before production use?

A: Because retrieval alone does not explain business meaning. A semantic layer tells the system what a field, record, or relationship represents, which reduces misinterpretation and policy drift. Without it, the model may be technically accurate but operationally wrong, especially when decisions affect regulated data or business workflows.

Technical breakdown

Governed meaning versus raw data semantics

A semantic layer normalises business terms so AI does not infer meaning from column names, table joins or prompt context alone. It maps source data to certified concepts such as customer, revenue or churn, then exposes those concepts in a machine-readable form. Without that layer, LLMs and agents rely on pattern matching and probabilistic guesswork, which works poorly when the organisation's definitions differ from public-language defaults. The technical point is that the model is not broken; the meaning substrate is missing or inconsistent.

Practical implication: define and certify business terms before letting AI reason over them.

Why context layers matter more once agents take action

A context layer extends semantics with lineage, quality, ownership and policy so the AI knows not only what a term means but whether it is trustworthy and allowed for the task. That distinction matters more for agents than for dashboards because agents can act, chain steps and reuse outputs across workflows. A governed context layer therefore becomes an execution guardrail, not just a reporting aid. It reduces the chance that an agent will assemble a plausible but invalid answer and then operationalise it.

Practical implication: attach lineage and policy to the meaning layer before agents can call downstream tools.

Open standards and runtime delivery for AI grounding

The article's architecture depends on meaning being delivered at runtime, not copied into one application or frozen in a static catalogue. That is why open standards matter: they let LLMs, retrieval pipelines and agents resolve the right definition, metric or relationship when they need it. The technical benefit is consistency across use cases such as RAG and text-to-SQL, where the same word can map to different governed concepts in different business contexts. Runtime delivery is what turns semantics into operational control.

Practical implication: expose governed meaning through standards that AI systems can query in real time.

Breaches seen in the wild

• Moltbook AI agent keys breach: Moltbook breach exposed 1.5M AI agent keys.
• DeepSeek database exposure 2025: An unauthenticated DeepSeek ClickHouse database exposed over a million log lines with plaintext chat history and API keys in 2025.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Governed meaning is now part of identity governance for AI systems: access control decides what an AI can touch, but semantic control decides what it is allowed to conclude from what it touches. Once LLMs and agents operate on business terms, certified meaning becomes a governance primitive alongside identity, privilege and policy. Practitioners should treat semantic governance as an extension of AI access control, not a separate analytics concern.

Meaning without governance creates a new form of authorisation drift: a system can be correctly authenticated to data and still be operationally wrong if it interprets business terms outside their certified definition. That is a distinct failure mode for AI governance because the output looks valid while the underlying decision is not. The practical conclusion is that governance has to move upstream from retrieval to meaning itself.

Semantic drift is the AI equivalent of stale entitlement state: if definitions, metrics and relationships are not governed centrally, agents will act on versions of truth that no longer match the business. That creates avoidable inconsistency across dashboards, workflows and autonomous actions. Practitioners should expect semantic governance to become a baseline requirement for trustworthy AI operations.

Open, governed meaning reduces intelligence-layer lock-in: when one platform owns the definitions, the organisation loses portability of trust as well as portability of data. That is an architectural risk because the meaning layer becomes part of the control plane, not a convenience feature. The field should treat neutral semantics as a governance requirement, not a product preference.

Certified meaning is the named concept that AI security has been missing: the decisive question is no longer only who can access data, but which meaning the AI is authorised to use. That concept matters because the same data can produce different actions depending on the interpretation applied. Practitioners should anchor AI governance on governed meaning rather than assume access controls alone solve trust.

From our research library:

• AI-related credential leaks surged 81.5% year-over-year in 2025, with the surrounding AI infrastructure leaking 5x faster than core LLM providers, according to the State of Secrets Sprawl 2026.
• Read next: Agentic AI Identity Guide

What this signals

Certified meaning is becoming a governance boundary: AI systems can only be trusted when they are grounded in definitions, metrics and relationships that are owned and current. That shifts control from static data access to runtime meaning delivery, which is a broader problem than traditional BI ever had to solve.

The practical risk for security and data teams is not simply a wrong answer. It is an agent acting on a plausible interpretation that no one can trace back to a certified business rule, which is why governed context needs to sit alongside identity and authorisation controls.

Governance programmes that stop at access policy will miss the real failure mode in AI adoption. According to the State of Secrets Sprawl 2026, AI-related credential leaks surged 81.5% year-over-year in 2025, and surrounding AI infrastructure leaked 5x faster than core LLM providers, which shows how quickly adjacent controls become the weak link.

For practitioners

• Define certified business terms first Inventory the terms that drive AI decisions, then assign owned, approved definitions for each one so models do not improvise meaning from raw data labels.
• Bind lineage and quality to semantic delivery Expose the source, freshness and quality signals alongside each governed metric so AI can distinguish trusted context from merely available context.
• Separate analytics semantics from runtime AI governance Use the semantic model to define meaning, then add policy and approval rules for where that meaning may be used by LLMs and agents.
• Test AI outputs against certified definitions Compare model answers to the organisation's approved metric and term definitions before allowing those outputs into reporting, workflows or agent actions.

Key takeaways

• AI systems can be properly connected to data and still fail if they are not grounded in governed business meaning.
• The article's key evidence is the KU Leuven test, where governed context raised accuracy from 62% to 92%.
• Practitioners need to treat semantics, lineage and policy as part of AI governance, not as optional metadata.

Key terms

• Action Layer: The action layer is the point where an identity moves from asking for access to doing something with that access. For AI agents, this layer matters because tool use can happen faster than human review, and the meaningful risk appears when actions are chained across systems.
• Context Layer: A context layer is the connective information that turns isolated data findings into actionable security intelligence. It correlates ownership, access, policy, movement, and regulatory exposure so teams can decide what to remediate, review, or restrict without manually stitching together separate tool outputs.
• Certified Metric: A certified metric is the approved calculation the organisation recognises as authoritative for a business measure such as revenue or churn. For AI systems, certification matters because the same label can hide multiple calculations, and only one should drive automated reasoning or action.
• Governed Meaning: Governed meaning is business interpretation that is owned, versioned and exposed through policy, rather than left to the model to infer. It is the control point that keeps AI reasoning aligned to organisational definitions, especially where action follows the answer.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are building or maturing an identity security programme, it is worth exploring.