Notifications
Clear all
Topic starter
25/08/2026 6:10 pm
TL;DR: AI agents can produce syntactically correct but materially wrong answers when they query raw warehouse tables without a semantic layer, because business meaning, edge cases, and metric definitions are missing from the schema, according to Island. The operational lesson is that AI-ready analytics depends on versioned, testable business logic, not just model capability.
NHIMG editorial — based on content published by Island: Scaling Meaning, The Case for a Semantic Layer in the AI Era
Questions worth separating out
Q: How should security teams govern AI agents that can access enterprise systems?
A: Security teams should govern AI agents as non-human identities with explicit ownership, scoped privileges, and continuous monitoring.
Q: Why do AI systems make weak data governance more dangerous?
A: Because they remove the natural limits that used to slow discovery.
Q: What breaks when business metrics are defined only in SQL files?
A: What breaks is consistency.
Practitioner guidance
- Inventory your highest-risk metrics first Start with the small set of business definitions that drive board reporting, automation, or customer-facing decisions.
- Require testable definitions before agent access Do not let AI agents query raw warehouse tables for critical use cases until the metric definitions are versioned, reviewed, and covered by reconciliation tests.
- Separate definition ownership from validation ownership Assign data teams to maintain the semantic definitions and business stakeholders to validate whether those definitions match operational reality.
What's in the full article
Island's full blog post covers the operational detail this post intentionally leaves for the source:
- The worked examples showing how raw warehouse queries can silently return wrong answers despite valid SQL.
- The specific semantic definitions used for active users, event categories, and audit-log meaning.
- The Island experiment notes showing how defined behaviour changed what AI agents could infer from browser events.
- The examples of how the team validated agent outputs against expected answers and real usage patterns.
👉 Read Island's analysis of why AI agents need a semantic layer for warehouse data →
Semantic layers for AI agents: what governance teams are missing?
Explore further
25/08/2026 6:11 pm
Semantic drift is now an AI governance problem, not just a reporting problem. Island’s argument shows that bad definitions can survive every technical test and still produce business damage when AI agents consume them at scale. The issue is not model quality alone. It is the absence of a governed meaning layer that can be reviewed, validated, and reused across systems. For identity and governance practitioners, this is the same class of problem that appears when access policy is technically valid but operationally misaligned. Practitioner conclusion: treat semantics as part of control design.
A question worth separating out:
Q: How do organisations know a semantic layer is actually working?
A: A semantic layer is working when downstream tools, analysts, and AI agents all resolve the same metric to the same tested definition, and reconciliation checks confirm the results match business validation. If teams still debate the number after the layer exists, the governance model is incomplete rather than solved.
👉 Read our full editorial: AI agents need a semantic layer before they can query data reliably
