Semantic intelligence is becoming the enforcement layer for enterprise AI agents

The same idea from three different directions

Within 72 hours this week, three enterprise vendors from different categories made announcements with a striking amount of overlap.

Rubrik rolled out what it calls the Semantic AI Governance Engine (SAGE) — a domain-specific small language model built to interpret policies by intent rather than enforce them through keyword matching. In internal benchmarks, Rubrik's custom SLM processed agent interactions 5x faster than GPT-5.2 and caught policy violations at a higher accuracy rate (StorageNewsletter). The framing is explicit: SAGE exists because legacy deterministic systems "cannot comprehend natural language nor adapt to dynamic and unforeseen actions taken by agents."

M-Files deepened its Microsoft 365 integration around what it describes as a context-first architecture — organizing content based on what information is rather than where it's stored. The practical difference: agents reasoning over customers, contracts, processes, and components rather than raw file blobs. "AI is only as strong as the information behind it," said Ryan Barry, VP strategic operations at M-Files. "Without the right context, even the most advanced models struggle to deliver accurate, actionable insights" (FinTech Global).

Glean published platform updates putting its permissions-aware knowledge graph at the center of everything — Agentic Engine 2, governance tooling via Glean Protect Plus, and a 15+ LLM hub designed to avoid hyperscaler lock-in. The company doubled ARR to $200 million in roughly nine months and hit a $7.2 billion valuation (Futurum Group). The knowledge graph isn't just search infrastructure anymore. It's the control plane for what the agent does next.

None of these companies compete directly. Rubrik is security operations. M-Files is document management. Glean is enterprise search turned agent platform. That they're converging on the same architectural argument this week is worth taking seriously.

Why static rules and generic retrieval stopped working

The governance approach that enterprises used for earlier automation — deterministic rules, keyword blocklists, role-based access lists — was built for predictable workflows. An AI agent operating in natural language is neither predictable nor deterministic.

Consider a policy that says: "Do not give financial advice." A keyword filter catches mentions of specific terms. It misses an agent that, in response to a question about Q3 runway, lays out a cash management recommendation without using any flagged words. Rubrik's SAGE is designed exactly for this gap — interpreting what a policy means, not just what it says.

Generic retrieval has a parallel problem. Most enterprise RAG deployments return documents. What agents actually need is business meaning: not "here is the contract PDF," but "this is a Master Service Agreement with Vendor A, which governs data handling, is still active, and can be seen by the procurement team but not external parties." The difference between those two is a semantic model of your organization's content.

M-Files calls this reasoning over business entities rather than file blobs. Glean calls it a permissions-aware knowledge graph. The terminology differs. The requirement is the same.

The four jobs a semantic layer does for enterprise agents

The category is still taking shape, but the functional requirements are becoming consistent across vendors. A semantic layer for enterprise agents has to do four things:

Map business entities and relationships. Before an agent can do anything useful, it needs to know that "customer" means a Salesforce record with certain attributes, that contracts have parties and statuses, that "the finance team" is a specific group with specific access rights. This mapping has to be built and maintained — it doesn't emerge from raw document retrieval.

Preserve permissions and provenance. Not every user should see every document, and when an agent surfaces information, the audit trail needs to record what it accessed and why. Permission-aware answers, in M-Files' framing, means the agent's outputs respect the same access controls as the underlying documents.

Interpret policy intent. This is Rubrik's territory. A policy is a natural language statement about desired behavior. The semantic layer has to translate that statement into runtime logic that generalizes across situations the policy writer didn't anticipate. Static rules handle the cases they were written for. Semantic interpretation handles the rest.

Guide or constrain action at runtime. This is the operational step that distinguishes semantic governance from semantic search. The agent doesn't just retrieve better answers — it receives active guidance or constraints about what it's allowed to do. SAGE frames this as "active, semantic enforcement" rather than monitoring after the fact.

What enterprise buyers are actually asking for now

There's a shift happening in how buyers evaluate agent platforms, and it's visible in how these vendors are positioning their announcements.

A year ago, enterprise AI conversations were about capability: what can this agent do, how accurate is it, which model is underneath it. Those questions haven't disappeared. But the questions arriving in procurement now are different: how do we know what the agent accessed, how do we prove it operated within policy, and how do we unwind a mistake when it happens?

Rubrik's SAGE includes an "Integrated Remediation" feature that triggers Agent Rewind when something goes wrong — instantly undoing destructive actions. Glean's Protect Plus is a paid governance SKU built explicitly around trust and auditability. These aren't afterthoughts. They're the features being bought.

Enterprise buyers are moving toward systems that are permission-aware, policy-compliant, and auditable by default. The semantic layer is how vendors are answering that requirement.

Semantic intelligence still depends on governed knowledge underneath

Here's the part that doesn't make it into most vendor announcements.

A semantic layer can organize meaning, map relationships, and interpret policy intent. What it can't do is fix the source material. If the documents underneath are stale, contradictory, or weakly attributed, a semantic layer will structure those problems more efficiently. Better semantics on top of bad documentation produces more elegant mistakes — not fewer ones.

This matters practically. An agent reasoning over business entities still has to retrieve facts from actual documents. If your contract management system has three versions of the same MSA, and two of them contradict each other on data handling terms, a semantically sophisticated agent will confidently reason over the wrong one. The semantic interpretation was correct. The source was not.

The implication for enterprise teams building agent infrastructure: semantic intelligence and knowledge governance aren't alternatives. They're dependencies. You can have an excellent semantic layer and still expose agents to contradictory, outdated, or unattributed source knowledge. The control plane handles agent behavior; the knowledge layer has to handle document accuracy.

Platforms like Mojar AI are built specifically for the knowledge side of this problem — scanning document sets for contradictions, attributing every answer to a specific source, and managing the ongoing accuracy of a knowledge base as it changes. The argument isn't that semantic governance is wrong. It's that it works at the meaning layer, and the meaning layer sits on top of documents that still need to be right.

What to watch

Expect the term "semantic governance" to become standard vendor language within the next quarter or two. The announcements this week are still fragmented across categories — security, document management, enterprise search. When a major platform consolidates them into a single pitch, the category will feel official.

The more telling signal is AtScale's move into the Databricks MCP marketplace. Semantic layers have lived inside BI and analytics tools for years. When they start appearing in agent infrastructure and MCP registries, that's the moment enterprise deployments acquire something closer to a governed meaning layer. That shift is already starting.

The market is moving from "agents need context" to "context must be structured, permissioned, semantically understood, and policy-enforceable." The vendors announcing this week are early. They won't be the last. The market is moving from "agents need context" to "context must be structured, permissioned, semantically understood, and policy-enforceable." The vendors announcing this week are early. They won't be the last.