Viewpoints

Seven vendors now sell agents a semantic layer. Who signs off what the data means?

Richard White · · 6 min read

In six weeks, Microsoft, Dell, Oracle, SAP, Zoho, NVIDIA and Snowflake all announced a layer of business definitions for AI agents to use. The industry now agrees agents need approved meaning before they answer. What most announcements leave out is who approves it, and whether the approval travels when your data sits on more than one platform.

What happened

Between late September and early November, seven data vendors moved on the same layer. Microsoft added agent-led ontology creation from Power BI semantic models and OneLake data to Fabric IQ. Dell launched a unified semantic layer and knowledge graph in its AI Data Platform. NVIDIA released Auto Ontology as open source under the Apache 2.0 license; it reads database metadata and proposes a starting set of business definitions for agents to use. Zoho added semantic and ontology layers to Zoho Analytics.

On 26 October, Oracle announced a Business Semantic Layer for its Autonomous AI Lakehouse, built from catalog annotations, database annotations and AI-suggested analytic views. Oracle says it will be available within the next 12 months. On 27 October at TechEd in Berlin, SAP announced Ontology Manager, which it describes as an AI-native way to build ontologies across SAP and other data; SAP did not give an availability date. And on 3 November, Snowflake announced Business Ontology in private preview, with one approved definition per business term, an accountable owner, imported candidates marked as drafts for the owner to approve and an audit trail of changes.

Oracle’s and SAP’s announcements don’t describe who approves a definition before agents use it. Snowflake’s does.

My view

This is the right layer, and I’m glad the market has found it. An agent that answers from the right data with the wrong meaning is more dangerous than one that fails, because the answer looks finished. Approved definitions served before the model answers are the fix. Demo 1: same question, two answers shows the difference in under a minute.

But a semantic layer is software, and meaning is a decision. Three questions decide whether the layer helps or just moves the ambiguity somewhere harder to see.

Who approves?

AI-suggested definitions are a fine starting point. They are not approved definitions. Someone in the business has to own each term, decide between competing versions and sign off before an agent may use it. Where an announcement is silent on this, buyers should ask, and should test whether an unapproved draft can reach an agent.

Does approval travel?

Few firms of any size run on one data platform. If approved meaning lives only inside one vendor’s layer, every term gets signed off again for each platform, and the copies drift. Ask whether definitions can be exported, in what format, and with their owner and approval history attached.

Is meaning kept apart from logic?

A definition says what “active customer” means. A metric calculation says how to count it from a given table. They change for different reasons and need different owners. Layers that merge the two make every schema change look like a change in meaning. Ontology or Semantic Layer? sets out how to keep them apart.

Why it matters

For the industry. The tools have arrived before the ownership. The firms that get value from these layers will be the ones that treat each definition as an owned, approved, versioned asset, the way they already treat a certified data product, and that can show which definition an agent used for a given answer.

For banks

A bank already spends real effort reconciling what “exposure,” “customer” and “default” mean across risk, finance and regulatory reporting. An agent that quietly picks one definition, or blends two, can undo that work in a single answer. Banks should require that any agent answering a risk or finance question uses only approved terms, and that the term and its version are recorded with the answer so a reviewer can reconstruct it.

For insurers

“Claim,” “open claim” and “loss ratio” mean different things to claims, actuarial and finance teams, often for good reasons. The NAIC model bulletin on insurers’ use of AI, adopted in December 2023 and taken up by a growing number of states, expects insurers to govern AI systems and data, including what they acquire from third parties. A vendor’s semantic layer is exactly that kind of third-party component. The insurer still owns the meaning, and should be able to show who approved each definition an agent relied on.

For private equity and portfolio companies

This is where portability pays most. One portfolio company runs on Snowflake, another on Fabric, a third on spreadsheets. The value creation plan, the lender pack and the board deck all depend on definitions such as recurring revenue, net revenue retention and adjusted EBITDA being the same everywhere. A sponsor that approves those definitions once, in a portable form, and requires every company’s agents and dashboards to use them, gets comparable numbers across the portfolio and a faster, cleaner diligence story at exit. Locking them into each company’s vendor layer means approving them again for every company, every platform and every buyer.

How it connects to our work. Ontology Studio is built for exactly this step: drafting business definitions, recording an owner and a reviewer for each, keeping drafts labeled as drafts, and exporting them in open formats so the same approved meaning can feed any platform’s layer. Data Product Certification covers the ownership and sign-off model that definitions should borrow from certified data products. I looked at Snowflake’s version in more detail in Snowflake puts business meaning in the catalog. Before switching on any vendor’s semantic layer for agents, I’d check:

  • Named owners. Every term an agent can use has one accountable owner.
  • No drafts in production. An AI-suggested or imported definition can’t reach an agent until it is approved.
  • Portable approval. Definitions export with their owner, version and approval history.
  • Traceable answers. Each agent answer records the definitions and versions it used.

This is part of a short series on putting agents into production safely, written for banks, insurers and private equity portfolios. If you want the next one, the sign-up is below.

Sources

DataAI agentsGovernance

Linked sources are the record. Check them for the full reporting.

Read next

Viewpoint · · 4 min read

Snowflake puts business meaning in the catalog. That’s the right layer.

Snowflake announced a Business Ontology that records approved, owned definitions of business terms and serves only those to agents, alongside a gateway that applies existing access policies to agent tool calls. Approved meaning before the model answers is the right pattern. Two questions remain for buyers: who owns each definition, and whether an agent should inherit a person’s full access.

DataOntology

Viewpoint · · 6 min read

One prompt, every agent: blast radius is the agent question that matters now

Researchers showed that one prompt to one exposed agent on Amazon Bedrock AgentCore could reach every agent in the same account and region, because they all leaned on a broad default role. As vendors ship thousands of prebuilt agents and Microsoft heads into Ignite calling agent adoption “a trust decision,” the question for banks, insurers and private equity portfolios is the same: if this agent is turned, what else does it reach?

SecurityAI agents

Viewpoint · · 10 min read

AI finds the flaws faster than you can fix them. Fixing is now the job.

McKinsey argues that frontier AI has moved the security bottleneck from finding software flaws to fixing them. I agree, and the lesson is older than AI: like data defects, software flaws only get fixed when someone owns them and has a budget. That gets hard when the business, not IT, owns the code, and it is the same consequence-and-authority logic the OCC just gave its examiners.

SecurityAI agents

All viewpoints Discuss this with Richard