2025 wasn’t just another year of AI hype. It was the year enterprise data shortcomings became visible at scale.
Multiple industry studies show that most AI initiatives stall before production or fail to scale. MIT’s Project NANDA found that roughly 95% of generative AI pilots show no measurable P&L impact, with only about 5% driving material revenue acceleration. Gartner forecasts that more than 40% of agentic AI projects will be abandoned by 2027 due to unclear outcomes and poor integration.
The pattern is clear: AI projects are failing because the data foundation isn’t ready.
That conclusion is increasingly shared across the industry. In his “Data 2026 Outlook,” analyst Tony Baer argues that the next phase of AI adoption will be defined not by models or tooling, but by semantics; specifically, who controls the definitions, context, and relationships AI systems rely on to reason.
He describes the emergence of “semantic spheres of influence,” where ambiguity is no longer tolerable, and meaning must be made explicit. What enterprises experienced in 2025 confirms this framing: AI didn’t introduce new data problems; it exposed the ones organizations had been living with for years.
As generative AI moved from proof of concept to production, long-standing issues around definitions, governance, and metric consistency surfaced in ways that couldn’t be ignored. The tolerance businesses had for semantic drift, including conflicting definitions, inconsistent metrics, and undocumented logic, ran out the moment AI agents started consuming that data.
What changed wasn’t the technology. What changed was the cost of ambiguity.
Shift 1: AI Exposed Semantic Drift
Humans have always tolerated a certain amount of ambiguity in data definitions. If two teams calculated revenue slightly differently, they reconciled the differences in meetings or footnotes. Dashboards could be annotated. Reports could include caveats.
AI doesn’t work that way.
Consider a common scenario that played out across enterprises in 2025:
- Finance reports “Revenue” as $10.2M in Power BI
- Marketing shows “Revenue” as $10.4M in Tableau
- An AI copilot surfaces “Revenue” as $9.8M in Slack
Which number do you bring to the board meeting?
The problem isn’t just one of conflicting totals. It runs deeper. Finance includes only booked revenue after returns processing. Marketing counts gross transaction value to measure campaign effectiveness. The AI agent, lacking semantic context, calculates whatever pattern it finds in the raw transaction tables, sometimes including test data, sometimes excluding international sales, depending on how the question is phrased.
This kind of drift existed before AI, but humans could navigate it through institutional knowledge and reconciliation meetings. When an LLM is asked a business question, it needs deterministic answers. If one system defines “active customer” as anyone with a purchase in the last 90 days, and another defines it as anyone with an open account regardless of activity, the AI will return conflicting results depending on which data it accesses.
For enterprises, the stakes are too high to tolerate errors. Business analysts demand accurate reports and dashboards because someone gets fired if the numbers are wrong. That same standard now applies to AI-driven analysis. If users can’t trust AI results, they won’t adopt the technology.
The challenge is threefold:
- LLMs are inherently probabilistic.
- Business data is stored in bespoke, complex schemas.
- Every organization has its own way of calculating core metrics, logic that generally trained LLMs know nothing about.
In 2025, organizations realized that AI couldn’t paper over semantic drift. It amplified it.
Shift 2: Governance Moved from Policy to Infrastructure
For years, data governance lived in documentation. Policies were written. Data dictionaries were maintained. Training sessions were held.
Then AI arrived, and documentation proved insufficient.
Take gross margin as an example. In most organizations, gross margin is not a simple calculation. It involves complex logic around cost allocation, returns processing, promotional discounts, and shipping adjustments. The specific formula varies by industry, business unit, and sometimes product line.
An AI agent tasked with “calculating gross margin by region” can’t be expected to read a governance policy and apply the correct formula. It needs logic embedded in the infrastructure itself. Defined once, enforced consistently, and made accessible through a semantic layer.
Without this, you get fragmentation:
- Excel users create manual calculations that drift from official metrics
- Power BI developers rebuild the same logic that exists in other dashboards
- AI agents hallucinate calculations based on raw data structure, not business context
- Data science teams struggle to find trusted, reliable data to power their models
This is where semantic layers became critical. A semantic layer translates physical databases into an enterprise-specific business context. It documents how metrics are calculated, ensures that queries are translated consistently, and automatically enforces access controls. When gross margin is defined in the semantic layer using expressions for time-based calculations, conditional logic, and nested aggregations, both Tableau and AI agents return the same result.
The shift in 2025 was recognizing that governance isn’t a compliance exercise. It’s the technical infrastructure. Without it, AI cannot be trusted, and AI-powered analytics systems will never be widely adopted.
Shift 3: Open Semantics Moved from Ideal to Necessity
Platform-native semantic capabilities made progress in 2025. But they also hit their limits.
Consider a beauty retailer that used different definitions across Finance, Marketing, and Store Operations. Before implementing a universal semantic layer, their teams used different systems: ERP, CRM, and POS, each with its own definitions of sales, revenue, and margin. Meetings were dominated by reconciliation rather than strategy.
When your semantic logic is tied to a single tool, you create dependencies and lock in. Your metrics are only as portable as the tooling vendor allows. Your AI agents can only consume semantics from systems that support proprietary formats. And when you need to federate data across platforms or support multiple analytics tools, which most enterprises do, you’re forced into manual reconciliation or duplicate modeling.
The heterogeneous environment reality is unavoidable:
- Data spans multiple clouds and platforms
- Teams use diverse BI tools based on their specific needs
- AI systems must access consistent business logic regardless of where they run
- Vendor strategies change, but business definitions must remain portable
Open semantics solves this problem. Standards such as the Model Context Protocol (MCP) enable semantic models to be consumed by any AI agent or analytics tool, regardless of platform. YAML-based modeling languages, such as Semantic Modeling Language (SML), provide structure and version control for semantic definitions, making them portable, auditable, and scalable.
By adopting a composable semantic layer, that same retailer created a single source of truth across all platforms. Power BI, Tableau, and ERP users now access identical definitions, each composed from the same governed metric and dimension objects. Updates propagate seamlessly across platforms via Git version control, so “Revenue” is defined once and reused everywhere, from Excel to Power BI to GenAI copilots.
In 2025, portability became a must-have. As organizations deployed AI across multiple clouds, analytics platforms, and operational systems, they needed semantic logic that could travel with the questions, not logic locked inside individual tools.
What This Redefinition Means
The semantic layer is no longer a BI convenience. It’s enterprise infrastructure.
Organizations that treated it as an optional tool in 2025 found themselves unable to scale AI responsibly. Those that invested in semantic layers, particularly open, governed, universal ones, were able to deploy faster, with greater stakeholder buy-in and measurably better outcomes.
This creates new expectations for how analytics and AI architectures are designed. Semantic layers must support both dashboards and AI agents using the same governed logic. They need to be platform-neutral, not platform-dependent. And they need to enforce governance automatically, not rely on documentation that AI can’t interpret.
The bar is now higher. And it’s not going back down.
Looking Ahead
2025 established the foundation. 2026 will be about execution.
The organizations that recognized these shifts early will accelerate. Those still treating semantic layers as optional will face mounting technical debt, trust failures, and delayed AI programs.
The cost of a trust failure at enterprise scale isn’t just a bad output. It’s a paused program, a reputational hit, and years of progress undone.
To see how leading enterprises are responding to these changes, download the 2026 State of the Semantic Layer report.
SHARE
2026 State of the Semantic Layer