The Semantic Layer’s New Job

Eighteen months ago, Blue Yonder’s analytics engineering team made a decision most data organizations haven’t made yet: they stopped being a BI team.

They saw what was coming. AI agents don’t just need access to data. They need to understand what the data means. In the not-too-distant future, the semantic layer’s most important consumer is no longer a person. It’s AI.

And yet, enterprise AI is struggling. According to S&P Global Market Intelligence, the share of enterprises scrapping most of their AI initiatives jumped from 17% in 2024 to 42% in 2025. An MIT study found that only 5% of enterprise AI pilots achieve meaningful revenue impact. 

The bottleneck isn’t the LLM. It’s context. 

AI Can’t Tolerate Inconsistency

Blue Yonder is the supply chain platform behind a meaningful slice of the world’s trucks, warehouses, and store shelves. The questions that drive decisions aren’t abstract: Is this shipment on time? What’s available in this warehouse right now? Where does margin go after freight allowances, returns, and regional negotiating differences?

When Jeremy Arendt, Senior Engineering Director of Analytics at Blue Yonder, joined the company in mid-2024, his team was managing roughly 1,000 dashboards and 800 data sets. Metric logic was scattered across Databricks, Power BI, Power Query, Excel, and SharePoint. There was no single authoritative source for core business definitions.

In a dashboard-centric world, that fragmentation was manageable. Human analysts could work around inconsistency, knowing which numbers to trust and which to double-check. AI doesn’t have that fallback. An AI agent querying fragmented logic doesn’t flag uncertainty. It picks a definition and runs. The output looks confident, but it might be wrong. 

The logic at Blue Yonder was “scattered and duplicated,” said Jeremy. In the AI era, that’s an infrastructure problem.

Why Blue Yonder Stopped Counting Dashboards

Blue Yonder’s answer was to rebuild from the ground up, with AI in mind. The team stopped measuring success by the number of dashboards delivered and began measuring it by the number of semantic models deployed, the number of semantic objects in production, and infrastructure adoption across the business. 

Today, that team operates 10 semantic models covering 600 deployed semantic objects. The end users of those models are business users using AI, not just BI analysts using a visual tool.

They rebuilt 800-plus tables into a governed dimensional model. Business logic moved out of individual tools and into shared infrastructure. Governance was embedded directly into metric definitions, with business sign-off required on each metric and calculation, version tracking built in, and lineage available by default.

To make this concrete: a supply chain semantic model might define a Shipment fact table joined to dimensions for Carrier, Origin Warehouse, Destination, Time, and Product. The metric “on-time delivery rate” is defined once, with business sign-off documented and version-tracked in the model itself. 

The goal was a governed definition, accessible across every BI tool and AI system, with no one forced into a single consumption layer.

“It’s not about laying AI on top of BI dashboards. It’s about what becomes possible when our infrastructure is semantic first.”

– Jeremy Arendt, Senior Engineering Director of Analytics at Blue Yonder

Why Context Has Become the Critical Variable

For the better part of a decade, the enterprise data industry focused on solving access through data lakes and cloud warehouses. Now, most large organizations can access their data, but they still lack a shared understanding. 

Without context, the questions that matter most to the business (what counts as revenue, how churn is defined, which customers are active)  are answered inconsistently across tools and systems. A data catalog tells you what a metric is supposed to mean. It doesn’t guarantee that every tool calculates it the same way.

The semantic layer is the mechanism that closes this gap, defining the metric once and ensuring it computes identically wherever it is consumed.

90 Seconds vs. 30 Hours

Blue Yonder tested what happens when AI is connected directly to a governed semantic layer via an MCP server (Model Context Protocol, the standard that enables AI tools to query business logic programmatically).

The baseline: a financial deep-dive analytics that previously required roughly 30 hours of work over four days, spanning multiple people across leadership and analytics.

The AI workflow completed the same analysis in about 90 seconds.

The outputs and financial figures closely matched the manually produced version. “Even if it wasn’t ready to present that very minute, it got us 80, 90, 95% of the way there,” said Jeremy. “It was a fantastic starting point.”

A result like that is only possible when the logic underneath is governed. When an AI tool queries the semantic layer through MCP, it’s working from logic that has already been approved and governed, which is what makes the output trustworthy enough to act on. Governance is built into the infrastructure, not applied on top of it.

The Deliverable Is No Longer a Dashboard

The lesson from Blue Yonder’s transformation is that the highest-performing data organizations may increasingly look like platform teams. The deliverable is no longer a dashboard or a dataset. It’s governed by business context: definitions, logic, relationships, and access controls that serve as the foundation for every downstream analytical and AI workflow.

With that foundation in place, the semantic layer’s most important consumer shifts from a person to an AI. The data teams that build it now spend less time debugging AI outputs and more time acting on them.

“When we can have that information, the context, the logic, as an infrastructure component, what can that enable from an AI perspective, from an analytics and reporting perspective?”

That’s the question every data leader needs to be asking: Is our semantic infrastructure ready to be the foundation AI builds on?

What Semantic-First Infrastructure Makes Possible

At AtScale, we have spent years building the infrastructure to make this possible: an open, portable semantic layer that works across cloud platforms and connects to AI tools via open standards such as MCP and Semantic Modeling Language (SML). Blue Yonder’s trajectory is a concrete illustration of the architecture that makes enterprise AI trustworthy at scale.

The organizations building governed semantic infrastructure now will have a compounding advantage. Every AI capability they add will work from shared, trusted business logic. Every decision will move faster.

The semantic layer has a new job. The teams that treat it as infrastructure, not a reporting tool, will be ready when the AI comes to query it.

Read Jeremy’s writing about semantic layers for O’Reilly, including “The Best Risk Mitigation Strategy in Data? A Single Source of Truth.”