Most enterprise AI governance conversations start with LLMs. But when an AI agent returns a wrong answer, the failure usually traces back to a much simpler problem: the business logic behind the answer was never defined in a place the agent could actually use.
That’s the argument the 2026 Semantic Layer Summit was built around, and it’s what made this year’s conversations different from previous years. Teams showed what governance actually looks like when the semantic layer is doing its job: consistent metric definitions enforced at query time, access controls that travel with the answer rather than stopping at the warehouse, and full auditability of what any agent touched and why.
You can have all the right policies in place, but if every agent is reasoning over raw tables with its own interpretation of “gross margin” or “active user,” those policies have nothing to enforce against. The semantic layer is where governance becomes structural.
That’s what we set out to explore at this year’s summit, with practitioners from Papa Johns, Slickdeals, TELUS, Vodafone, Carrefour, and Blue Yonder alongside leaders from Anthropic, Accenture, Chevron, WPP, Snowflake, and Databricks.
Here are ten things I took away from the day.
1. The semantic layer has crossed from evaluation to deployment.
I opened the summit by stating that the semantic layer is no longer being evaluated as a concept. It’s being deployed as critical AI infrastructure. The practitioners who joined us from Papa Johns, Slickdeals, TELUS, Vodafone, Carrefour, and Blue Yonder were there to share what they built, what broke, and what they’d do differently.
That’s a different conversation than we were having two years ago.
2. Your AI foundation builds off of your BI foundation.
A central argument of my keynote was that while AI changes how people interact with data, it doesn’t change what enterprises require from data. They still need trusted answers, governed access, reusable business logic, and models that reflect how the business actually works.
The teams that invested in semantic consistency for BI didn’t have to start over when AI arrived. Your BI foundation is your AI foundation.
3. The problem isn’t the LLM. It’s the missing context.
Enterprise AI failures are usually framed as hallucination problems. In most cases, they’re context problems.
A schema doesn’t tell you how finance defines revenue. A table doesn’t tell you how a hierarchy should roll up. A column doesn’t tell you which time definition matters. When AI runs directly on raw data, it can generate output, but not business truth.
4. The standard for enterprise semantic layers comes down to four requirements.
Open, governed, multi-model, composable. These pillars are non-negotiables for semantic infrastructure that can support enterprise AI.
- Open means one semantic foundation serving every interface without redefining the business for each one.
- Governed means trust enforced in the execution path, not bolted on after the answer shows up.
- Multi-model means preserving the semantic depth enterprises have built over the years: dimensions, hierarchies, time intelligence, and complex calculations.
- Composable means shared business logic at the center, local extension where teams need it, and versioned change management around the whole system.
If a semantic layer can’t satisfy all four, it’s not ready for enterprise AI.
5. Customers showed what “consistent metrics” actually means at scale.
The most grounded sessions of the day came from practitioners doing the work in production.
At Slickdeals, Michael Skariah put it plainly:
“When different teams all have their own numbers, things get confusing fast. We had dashboards, spreadsheets, and SQL all telling different stories.”
– Michael Skariah, SlickDeals
His team built a single, governed layer that sits between their data platform and BI tools. Finance, product, operations, and marketing all pull from the same definitions. The result: “Those debates about whose number is right basically disappeared.”
At Papa Johns, Brian Jones described a version of the same problem playing out across franchise operations, finance, and corporate users, each working from their own definition of performance. The turning point came when they stopped asking which tool to use and started asking how to guarantee the answer was the same regardless of the tool.
“AtScale has been an absolute game changer for Papa Johns,” one of Papa Johns’ top users told Brian’s team.
– Brian Jones, Papa Johns
6. Governance has to travel with the answer, not just the data.
The warehouse may be secured. That doesn’t mean the answer path is governed, explainable, or appropriate for the user asking the question.
At the summit, we demonstrated passthrough security and governance, with governance fully embedded in the semantic objects consumed by agents. Full attribution in the event of an issue. The right answer for the right user, enforced consistently across every interface at query time.
If agentic AI is going to support real business decisions, it has to be the default, not an afterthought.
7. SML and version control are turning semantic modeling into engineering practice.
Cornell Lee from TELUS walked through something that should resonate with any team managing semantic definitions at scale. His team manages tens of thousands of raw network performance counters across four vendors, each measuring the same events differently. The only way to keep that coherent is to manage it as code.
“Instead of having metric formulas in a spreadsheet or embedded in a dashboard SQL, they are managed at scale as YAML files. This is the only place that metrics are defined. So this becomes our single source of truth.”
– Cornell Lee, TELUS
Having everything in a Git repository means version control, change tracking, and a path toward cross-team collaboration on metric definitions. That’s the kind of foundation that makes AI-driven analytics safe to rely on.
8. Semantic layers unlock AI readiness in ways teams didn’t plan for.
Several customers described arriving at AI readiness as a byproduct of doing semantics right for BI, rather than a result of an AI initiative.
Sean Francis from Vodafone described building their unified semantic layer as part of a broader cloud modernization program. When AI interfaces started becoming a business expectation, the foundation was already there:
“Although we hadn’t really considered, when we started that, how useful that was going to be when natural language integration with the data becomes available, which it is now. I think that has put us in a really, really strong position. We’re not having to rewind.”
– Sean Francis, Vodafone
That’s the argument for building open, governed semantic infrastructure: it absorbs new interfaces rather than forcing a rebuild every time one arrives.
9. Composability is about scale without drift.
Semantic drift is a byproduct of duplication: one team defines a metric, another copies it, and a third changes it slightly. Eventually, the organization is managing drift instead of managing meaning.
Composable architecture breaks that pattern. With shared business logic at the center, controlled extension at the edges, and versioned change management throughout, you can scale semantic consistency.
Brian Jones at Papa Johns described what this looks like in practice. His team started with a single “universal cube” that was fragile and hard to change. They’ve transitioned to smaller, domain-focused models where each team gets what they need without creating risk for the rest of the organization.
“Whether it be that chatbot using AtScale data or a traditional Tableau desktop or some analyst in FP&A looking at an Excel spreadsheet, they’re going to get that exact same answer every single time.”
– Brian Jones, Papa Johns
10. LLMs are a commodity. Business context is the moat.
Nobody predicted AI would be this prevalent, and nobody can predict what the next wave looks like. It’s precisely that uncertainty that makes vendor lock-in a major risk. The enterprises that standardized their semantic strategy on a single tool’s definition of what a semantic layer should be made a bet on the future and lost. They’re rebuilding right now.
Openness and freedom of choice are the only rational architecture decisions. An open, independent semantic foundation absorbs change rather than resisting it. By building a foundation that doesn’t require you to pick again tomorrow, you build an actual semantic advantage.
All sessions from the 2026 Semantic Layer Summit are available on demand. If any of the conversations above resonated, I’d encourage you to go deeper. These are the teams running semantic layers in production, with deep insights into governance, adoption, and AI readiness.
SHARE
Guide: How to Choose a Semantic Layer