Semantic Layer vs. AI Control Plane: Why Enterprise AI Needs Governed Context

Quick Answer 
AI control planes govern how agents act. A governed semantic layer governs what they act on. Strategy Mosaic is the universal, governed semantic layer enterprise AI needs to produce trustworthy answers at agent scale, not just auditable ones. 

The semantic layer conversation has arrived. Dave Mariani put it well ahead of the AtScale summit: the biggest hurdle for enterprise AI in 2026 isn't the model. It's the meaning. 

Strategy Software has been building toward this exact problem for the last decade. The diagnosis is right. But the conversation about how to solve it is heading somewhere that worries me. 

In the last 90 days, Databricks, Snowflake, Salesforce, ServiceNow, and SAP have all announced some version of the same thing: an AI control plane. The pitch is real. As agents move into production, enterprises need to govern what they do. Audit logs, rollback capabilities, visibility into agent actions. Nobody disagrees with the problem. 

But here's what a control plane can't do. It can't tell your agent what "revenue" means. It can't prevent a coverage limit from being calculated at several times the correct value because of a join error that looks fine on the surface. It can't stop a catastrophe payout from being attributed to the wrong policyholder. A control plane governs how agents act. It has no opinion on what agents act on. 

Those are different problems. Only one of them is actually stopping enterprise AI from being trustworthy at scale. 

In internal testing, Strategy compared AI-generated analytics against a raw SQL baseline on insurance claims data. The SQL approach got most questions right. The ones it got wrong weren't obviously wrong. They returned plausible, confident numbers. A coverage limit came back at several times the correct value. Nobody would have flagged it as an error. 

This is the failure mode nobody's talking about clearly enough. The risk isn't an agent that says "I don't know." The risk is an agent that files the wrong number with complete confidence. At dashboard scale, a human analyst catches that eventually. At agent scale, when the agent is calculating, deciding, and acting without a human in the loop, it doesn't. 

A semantic layer that holds your metric definitions is useful. A governed semantic layer is what actually closes the gap. That means definitions your business has approved, exposed through an auditable interface, enforced above every platform your data touches. 

That's what Strategy Software's Mosaic is. Not a new category. A semantic layer built with 25 years of battle-tested experience, extended for the age of AI agents. The same definitions that power your dashboards power your agents through MCP. One governance policy set applies whether the consumer is Power BI, Claude, or a custom workflow you built last week. 

The difference between a semantic layer and a governed context layer isn't a product distinction. It's an architectural commitment. Either your metric definitions are the source of truth for every system that touches your data, or they're one input among many. 

AtScale's framing, semantic layer as the context layer for enterprise AI, is right. It's worth knowing that Snowflake is an investor in AtScale. That doesn't make the argument wrong. It does raise a question worth asking any semantic layer vendor: whose platform are you optimizing for? 

Infrastructure follows the money. A semantic layer with a major cloud platform as a strategic investor will eventually (and inevitably) optimize for that platform's gravity. When you're building a multi-year architecture, you aren't just buying a tool. You're buying their incentive structure. 

Five criteria separate production-grade semantic layers from pilot-grade ones: 

1. Multi-cloud portability. The semantic layer runs above your data platforms, not inside one of them. Definitions move with you across AWS, Azure, GCP, and on-prem.
2. Tool-agnostic governance enforcement. The same governance policy applies whether the consumer is Power BI, Tableau, Claude, ChatGPT, or a custom agent.
3. Demonstrated production AI agent deployments. Real customers running agents against the semantic layer at enterprise scale, not pilot demos.
4. Real-time governance observability. When an agent queries the layer, you can see what it asked, what it received, and which policy decisions applied — live, not in a next-day audit log.
5. Alignment with open standards. MCP and OSI compatibility so the layer doesn't become its own lock-in.

My colleague Aidan Reilly wrote a detailed breakdown of all five and how to evaluate against them: Semantic Layer for Enterprise AI in 2026: What Production Use Requires. Worth reading if you're in active evaluation mode. 

The diagnosis everyone's landing on — meaning is the hurdle — is correct. The question is whether you're solving for context or for governance of context. Only the second one holds up when agents are making decisions your board will have to stand behind.