Enterprise AI business context is the missing layer that semantic models alone cannot provide.

Over the past few weeks, this article sparked far more conversations than I expected. I had the privilege of discussing these ideas with customers, architects, researchers, and many people in the Data & AI community whose thinking I deeply respect. Special thanks to Mario De Felipe, Frank Gundlich, Alexander Zeier, Carsten Bange, Joseph O’Leary, Néhno Van Eemeren, Sebastian Dannehl, and of course Tara. While the conclusions are my own, your ideas and our conversations have significantly shaped my thinking.

When I first wrote this article, I used the term semantic layer in the broad way it is commonly used across our industry. The more I reflected on it, the more I realized that we are using one term to describe several fundamentally different concepts. Rather than leaving the original article untouched, I decided to update it. I believe thought leadership should evolve as our understanding evolves. If we expect AI systems to learn continuously, we should hold ourselves to the same standard.

++++ Updated blog ++++

As said this article sparked more conversations than I expected, with CDOs, enterprise architects, data leads, and SAP practitioners. And one observation keeps surfacing:

As an industry, we’ve correctly identified the problem. But we’re still struggling with the language.

Today, almost everything is called a semantic layer: knowledge graphs, metadata catalogs, taxonomies, business glossaries, ontologies, or data catalogs. Everyone and their mother seems to have one.

I’m saying this jokingly but that’s actually good news. It shows the industry has realized that enterprise AI needs more than data to reason correctly in an enterprise context. But I also believe we’ve overused semantics as a catch-all term for everything that sits between data and AI. That’s blurring an important distinction: 
Ontologies, metadata, and semantic models are the mechanisms through which context is organized, represented, and made accessible to AI.

They are not the Business Context itself.

Let me explain why.

What Semantics Can and Cannot Do

Ontologies define entities and relationships. Semantic models make data machine-readable and understandable. Metadata catalogs tell you what data exists and where to find it. These are truly valuable. I am not arguing against them.

But none of them can tell AI how your business actually operates.

They cannot encode your revenue recognition policy. They cannot capture the pricing logic your finance team negotiated three years ago. They cannot represent the approval thresholds that vary by entity, country, and transaction type. They cannot document the exceptions your SAP configuration was built around — or the ones that were never documented at all.

That is Business Context. And it is something categorically different from a semantic layer.

A Simple Example

Ask any AI to calculate net margin. It knows the formula and will confidently explain:

Net profit divided by revenue.

Textbook correct. Operationally useless. You need the net margin for your order 123 or your profit center XYZ. And that requires answers to questions no ontology contains:

• How do you recognize revenue—at delivery, invoice, or payment?
• Which costs are classified as cost of goods sold versus overhead?
• Are intercompany transactions included or excluded?
• How are rebates, discounts, and write-offs handled?

Every one of these questions has a different answer depending on your organization and sometimes even more than one.

AI doesn’t need more data to answer that. It needs to understand your business.

And right now, for most organizations, that understanding or context simply doesn’t exist in a form AI can access.

This valuable context lives in many sources across your organization:

• revenue recognition policy buried in SharePoint
• SAP configuration created and fine-tuned over decades
• exceptions documented in spreadsheets
• experience of your finance team
• sometimes only in the head of Helga

The Enterprise AI Iceberg

I find it useful to think about this as an iceberg:

AI can only reason about the context it can see.

Above the waterline is everything AI can easily see:

• Tables, fields, and data structures
• Industry terminology and standard definitions
• Publicly available knowledge, research, and best practices
• Generic business concepts and financial formulas.

Below the waterline is Business Context:

• Ways of working
• Institutional knowledge
• Historical decisions
• Company-specific rules
• Exceptions
• Business logic that exists only in Helga’s head

Humans rarely notice this distinction. We naturally reconstruct missing context from experience, conversations, and years of working inside the business. But as long as Business Context remains below the waterline, AI models have little choice but to infer what is missing. Sometimes it guesses correctly. Sometimes it hallucinates. Sometimes it makes incorrect assumptions. And sometimes it burns valuable tokens trying to reason about information that simply isn’t there.

This is not a model problem. It is not a data problem. It is a context visibility problem.

Context is king.

From Semantics to Business Context

Many organizations are now building a stack that looks something like this:

Data → Ontology → Semantics → Business Context

Each layer adds meaning:

• Ontologies define entities and relationships.
• Semantic models make data understandable.
• Business Context explains how the business actually operates and makes decisions.

This is why concepts such as federated reasoning become important. Not because AI needs access to more data, but because it needs to reason across fragmented pieces of Business Context that are distributed across systems, policies, configurations, and people.

Consumer AI works because it draws on broadly shared knowledge. Enterprise AI must operate within Business Context that is unique to every organization. That is a fundamentally different problem, and it requires a fundamentally different approach.

A Small Change in Vocabulary

Personally, I’ve started using the term Business Context instead of referring to every sophisticated implementation as a semantic layer.

Not because semantics are unimportant—they are essential. However, the goal is enabling AI to understand how a business works to support reliable decisions with repeatable results.

To me, that is Business Context.

This Raises a Question

If AI can only reason about the context it can see, and if most Business Context currently sits below the waterline, then the next question is inevitable:

How do we make Business Context explicit?

And more importantly:

How do we ensure AI makes decisions that are consistent, governed, and aligned with how the business actually operates?

That is where I’ll continue in the next article.

In the meantime, I’m curious:

How visible is your iceberg today?

How much of the Business Context that governs your operations actually exists in a form AI can access — versus living in configurations, documents, and people?

I’d genuinely like to hear where you are with this. Feel free to reach out directly or share your experience in the comments.