Why Data Modeling Matters More in the Age of AI

The age of AI is here, and the technology seems to advance every week.

Organizations are understandably eager to take advantage of it. We want AI to analyze our data, answer questions, identify trends, build visualizations, explain what changed, and eventually take action on our behalf. But there is an important difference between this technology revolution and many of the ones that came before it.

For decades, we built software that humans operated. We clicked buttons. We selected filters. We wrote SQL. We built dashboards. We interpreted the results.

Now we are increasingly asking software to do some of that interpretation for us.

AI Can Understand Data. But Does It Understand Your Business?

Imagine asking an AI assistant:

"How are sales doing?"

For a human analyst who has worked at your company for five years, that question contains an enormous amount of unstated context.

What does sales mean?

Does it mean:

• Bookings?

• Orders?

• Invoiced revenue?

• Recognized revenue?

Should cancelled orders be included? Are we looking at gross sales or net sales? What currency should we use? Does "this year" mean calendar year or fiscal year? Which date determines when a sale occurred? Should internal transactions be excluded? And what does doing well mean anyway? Compared with last year? Budget? Forecast? Target?

The database probably doesn't contain the answers to all of those questions.

Your Database Contains Data. Your Organization Contains Meaning.

Companies have accumulated decades of business knowledge that doesn't necessarily exist in their databases. It exists in meetings, spreadsheets, documentations, calculations, and even in arguments between Finance and Sales about what "revenue" means. It exists in the head of the employee who has been with the company for 17 years and knows why one particular product category has to be handled differently.  Humans have traditionally supplied that missing context.

When an analyst sees `ORDER_STATUS = 'C'`, they may know that it means cancelled and shouldn't be included in a particular calculation. When an executive asks about "customers," an analyst may know that the executive means active customers who purchased within the last 12 months.

The database doesn't necessarily know any of that and an LLM certainly can't be expected to magically know it.

This Is Why the Semantic Layer Matters

If we want AI to reason about our businesses, we need to give it more than tables and columns. We need to give it semantics. A well-designed semantic layer describes not only where the data lives, but what the data means.

That includes things like:

• relationships between business entities

• definitions of metrics and KPIs

• calculations and business rules

• appropriate dimensions for analysis

• field descriptions

• synonyms and business terminology

• default aggregations

• time relationships

• security and governance rules

• departmental definitions and perspectives

In other words, the semantic layer starts to capture some of the institutional knowledge that humans have historically carried around in their heads.This is extraordinarily valuable to AI.

The Better the Context, the Better the AI

Large language models are remarkably good at reasoning when they are given the right context.  Without context, they have to infer and that is where the problems come in. Inference is exactly what we don't want when we're asking questions about important business metrics.

If a field is called `REV_AMT`, an AI system can make an educated guess about what it represents. If the semantic layer tells it:

Net Revenue — recognized revenue after discounts and returns, excluding intercompany transactions, using the invoice date for period attribution

there is much less to guess.

The same principle applies to relationships. Giving an AI access to 200 database tables doesn't necessarily make it smarter. It may simply give it 200 opportunities to choose the wrong join. A semantic model that describes customers, orders, products, targets and organizational structures (and how those concepts relate to one another) gives the AI a map of the business rather than simply handing it the keys to the warehouse.

There Is More Than One Version of the Truth

Another subtle problem is that businesses don't always have one universal definition for everything. Finance may define revenue differently from Sales. Operations may define an active customer differently from Marketing. Neither definition is necessarily wrong. They exist because those departments are answering different questions.

Historically, experienced analysts understood these nuances and translated between them. If AI is going to augment those analysts, those nuances need to become part of the context available to the AI. The goal of a semantic layer isn't necessarily to force every department into one definition. Sometimes the more important job is to make those definitions explicit.

Instead of AI guessing which version of "revenue" somebody means, the semantic layer can tell it:

Sales Revenue is used by the Sales organization for pipeline and quota reporting. Recognized Revenue is the Finance definition used for financial reporting.

That is business knowledge.

AI Actually Makes Data Modeling More Important

There has been an interesting narrative around generative AI that perhaps we won't need traditional analytics infrastructure anymore. Why build models, dashboards, and semantic layers? Why not just point an AI at the database and ask questions?

The reason is that generating SQL is not the hard part. Understanding what the SQL should mean is the hard part.

AI is becoming extremely good at translating natural language into technical instructions, but no model, regardless of how capable it becomes, can reliably infer business rules that were never provided to it. If anything, AI increases the value of good data modeling. The more autonomy we give AI, the more important it becomes that the AI operates from a governed and well-defined representation of the business.

There Is Good News: AI Can Help Build the Context

At first, this may sound like an enormous documentation project.

Do we now have to manually describe every field, relationship, calculation and business rule before we can use AI? Provide AI with all the policies that govern and frame our business?

Fortunately, no. AI can help us build the very context that AI needs. It can examine schemas, analyze existing calculations, suggest field descriptions, identify likely relationships, summarize documentation, and help uncover inconsistent definitions. Interestingly, it can also ‘interview’ subject-matter experts and turn their explanations into structured metadata. This allows the AI to uncover ‘missing pieces’ and get a more fulsome picture of the business.

The human role becomes less about manually documenting thousands of fields and more about discovering, defining and governing business meaning. AI can accelerate that process by surfacing patterns, inconsistencies and possible definitions, but people still have to determine what those definitions should mean and how they should be applied.

AI Can Help Build the Context. It Can’t Own the Decisions.

This doesn’t mean that building a semantic layer becomes an automated exercise.

AI can examine a schema and identify that CUSTOMER_ID appears in six tables. It can suggest relationships between those tables. It can read existing calculations and documentation and propose definitions. What it cannot know, without help, is whether those relationships reflect how the business should actually be analyzed. That distinction matters.

A technically valid join can still produce the wrong answer. A calculation that has existed in a dashboard for ten years can still represent an outdated business rule. Two systems can contain similarly named customer fields while representing very different concepts. Documentation can also be incomplete, contradictory or simply wrong This is where data expertise and business expertise become important.

Building useful context for AI requires understanding the data architecture, but it also requires asking the right questions of the people who understand the business. Which systems are authoritative? At what level of detail should these datasets be related? Why does Finance calculate this metric differently from Sales? Which definition should an AI use in a particular situation? What information should different users be allowed to see?

AI can accelerate the discovery process enormously. It can inspect, document, compare and suggest. But someone still needs to recognize the ambiguity, facilitate those conversations, design the model, test the assumptions and help the organization decide what should become trusted business context.

In many ways, this changes the role of the consultant rather than eliminating it.

Instead of spending weeks manually documenting schemas and writing repetitive calculations, we can use AI to accelerate that work and spend more time on the difficult part: understanding the business, identifying where meaning is missing or contradictory, designing the semantic model, and validating that the answers it produces are actually correct.

The goal isn’t to have consultants do work that AI could do. The goal is to combine AI’s ability to analyze and generate with human experience in data, analytics and business processes. That combination is what turns a collection of metadata into a representation of the business that people, and increasingly AI agents, can trust.

Better Context Also Means More Efficient AI

There is another practical advantage to all of this. Good semantic context doesn't just improve accuracy. It improves efficiency.

If an AI agent has to repeatedly inspect schemas, examine tables, test joins and reconstruct business logic every time someone asks a question, it requires more work and more context. That means more latency, more computation and more tokens.

A semantic layer gives the AI a head start. Instead of repeatedly rediscovering how the business works, the AI can operate from a curated representation of that knowledge. Better context means less guessing. Less guessing means less work. And less work can mean faster, more reliable and less expensive AI.

The Semantic Layer Is Becoming the Interface Between Your Business and AI

For years, we thought about semantic models primarily as an analytics tool. They made dashboards easier to build. They standardized calculations. They helped users explore data without understanding the underlying database. Those things still matter, but AI gives the semantic layer a much larger role. It becomes a way of explaining your organization to a machine.

Your databases tell AI what happened. Your semantic layer helps explain what it means. And the richer that context becomes, the more capable AI becomes at helping people understand the business.

The organizations that get the most value from AI won't necessarily be the ones with the largest models or the most data. They may simply be the organizations that are best at explaining their business to the AI.

And that brings us right back to something we've been talking about for years:

Good data modeling matters.

Perhaps now more than ever.