The most common reason enterprise AI pilots stall is not a model problem or a technology problem. It is a definitions problem. When net sales means three different things across three finance teams, and available stock means four different things depending on who you ask, no AI model can return outputs that everyone will trust. This is the semantic layer problem, and it is more widespread than most organisations realise.

The Problem Nobody Names Directly

Ask a finance team and a commercial team in the same organisation what net sales means. In most large enterprises, you will get different answers. The finance team may define it as gross revenue after returns but before discounts. The commercial team may define it as revenue after all customer deductions. The regional team may have a third definition that excludes intercompany transactions.

None of these teams is wrong. Each definition reflects a legitimate business purpose and a historical convention that made sense in its context. The problem is not that different definitions exist. The problem is that they are not documented, not reconciled, and not visible to the AI systems that are being asked to work with the underlying data.

When an AI model queries net sales across the organisation, it has no way of knowing which of these definitions the data in any given table reflects. It returns a number. Different teams query the same model and get results that do not match. The model is blamed. The pilot stalls. The real cause, a semantic layer problem rather than a model problem, is rarely identified.

DATAVERSITY research found that 68 per cent of organisations cite data silos as their primary data concern. But the deeper issue behind many of those silos is not that the data lives in different places. It is that the same data is described and used differently by different parts of the business, and nothing in the system resolves that conflict.

3 days is the typical manual reconciliation overhead at the start of every planning cycle in multi-brand or multi-division enterprises where business metric definitions are not standardised. Before any analysis can begin, teams spend days aligning on which number is right. That overhead is not a reporting problem. It is a semantic layer problem, and it is entirely preventable.

Why This Problem Gets Worse Before AI, Not Better

In a traditional reporting environment, inconsistent metric definitions are a friction cost. Teams spend time reconciling numbers. Meetings start with a debate about which figure to use. Spreadsheets proliferate as each team maintains its own version of the data. The cost is real but diffuse, absorbed into planning cycles and management processes without ever being precisely quantified.

When AI is introduced into this environment, the problem does not diminish. It amplifies. AI models require consistent input data to produce reliable outputs. A model that is trained or queried on data where the same metric has different definitions in different source systems will produce outputs that are internally inconsistent and externally untrustworthy.

More critically, AI agents that take actions, submitting replenishment orders, generating financial forecasts, triggering commercial alerts, based on data that means different things in different contexts can cause material operational problems. An agent that sees available stock as the figure from the warehouse management system when the merchandising team means something different by the same term can generate recommendations that contradict the actual inventory position.

The impact is not just inaccurate outputs. It is the erosion of trust in AI across the organisation. Teams that receive AI outputs they cannot verify will stop relying on them. Once that trust is lost, rebuilding it requires not just fixing the underlying data but also demonstrating reliability over time. The semantic layer problem is much easier and cheaper to solve before AI is deployed than after.

What the Semantic Layer Actually Is

The semantic layer is a governed set of business definitions that sits between the raw data in operational systems and the tools, models, and agents that query it. It does not store data. It describes what the data means, and makes those descriptions accessible alongside the data itself.

A well-built semantic layer answers, for every important business metric: what is the precise definition of this term, what is the calculation that produces it, what data sources does it draw from, which business rules apply, and which exceptions or exclusions are relevant in which contexts.

When an AI agent, analytics tool, or planning system queries a metric through a semantic layer, it is not working out from raw data what the business context might be. The context is already encoded. The definition is authoritative. The calculation is consistent. Different teams querying the same metric through the same layer get the same number, derived the same way, every time.

This is what makes AI outputs trustworthy. Not model sophistication. Not data volume. Definitional consistency, enforced at the layer where the query meets the data.

The Definitions That Most Commonly Cause Problems

The semantic conflicts that most consistently block AI programmes tend to cluster around a small set of commercially important metrics. Understanding which definitions are likely to be in conflict is a useful starting point for any organisation beginning this work.

• Revenue definitions: Gross revenue, net revenue, net sales, recognised revenue, and cash received are all legitimately different figures, but the terms are used inconsistently across finance, commercial, and operational teams. In a multi-brand or multi-market business, the variations multiply further, with different regional treatments of discounts, returns, and intercompany transactions.

• Inventory definitions: Available stock, on-hand stock, available-to-promise, net available, and in-transit stock describe related but distinct inventory positions. Merchandising, supply chain, and finance teams often use these terms interchangeably while meaning different things. An AI model working with inventory data in an organisation where these definitions are unresolved will produce demand and replenishment outputs that cannot be trusted by any of the three functions.

• Customer and order definitions: When does an order become completed? Is a customer active if they have purchased in the last 12 months, or the last 6, or the last 3? What constitutes a return versus a cancellation? These definitions drive customer analytics, commercial performance measurement, and operational reporting. Inconsistency here produces AI outputs about customer behaviour that different parts of the commercial organisation will dispute.

• Product definitions: In a multi-brand or multi-category business, product master data is frequently inconsistent. The same physical item may exist under different identifiers in different systems, with different attribute definitions, different unit-of-measure conventions, and different category hierarchies. AI models working on assortment, pricing, or demand planning in this environment cannot produce reliable outputs until the product master data is resolved.

How to Approach the Semantic Layer Build

Building a semantic layer is a business process exercise as much as a technical one. The technical work of building a governed data model is straightforward. The harder work is reaching agreement on the definitions themselves, and that requires bringing the right people from the right business functions into the same conversation.

1. Start with the decisions AI will be supporting. Rather than attempting to define every metric in the business, identify the specific decisions the AI programme is intended to inform, and work backwards to the metrics those decisions depend on. A demand planning AI needs available-to-promise stock, net sales by SKU, and weeks of cover. Start there, not with a comprehensive data dictionary.

2. Assemble the right stakeholders. Each metric needs an owner from the business function it primarily serves, a technical owner who understands how it is currently calculated in source systems, and a governance representative who can document the agreed definition and enforce its use. All three are necessary.

3. Document the conflicts before resolving them. Before attempting to standardise a definition, document how it is currently used across all relevant teams. This surfaces the legitimate reasons for variation, which often reflect genuine business differences, and allows the standardisation conversation to begin from a factual rather than an assumed baseline.

4. Build the semantic layer incrementally. Start with the 20 to 30 metrics that matter most for the first wave of AI use cases. Get those definitions agreed, documented, and encoded in the semantic layer. Expand to additional metrics as subsequent use cases are scoped. Attempting to build a comprehensive semantic layer before any AI goes into production is both slower and less accurate than building incrementally in response to real use cases.

42% of enterprise AI initiatives abandoned in 2025 cited data quality and consistency as the primary reason for failure, according to S&P Global research. The majority of those failures were not caused by poor underlying data. They were caused by data that was inconsistently defined across sources, making it impossible for AI models to produce outputs that different teams could agree on and trust.

The Business Outcome of Getting This Right

Organisations that invest in the semantic layer before deploying AI report a consistent set of outcomes that go well beyond improved AI accuracy.

Planning cycles that previously required days of manual reconciliation to align numbers across teams become single-source queries. Management meetings that started with a debate about whose figures were correct begin with agreement on a shared set of numbers and spend their time on decisions rather than data validation. AI outputs that were previously disputed because different teams expected different results become trusted reference points.

The operational shift is significant. Teams that were spending a meaningful proportion of their planning time on data reconciliation redirect that time to analysis and decision-making. AI models that were blocked because their inputs contradicted each other go live. Forecasting accuracy improves not because the models changed but because the data feeding them became consistent.

The semantic layer is not the most visible part of an AI programme. It does not produce demos that impress a board presentation. But it is the part that determines whether everything else the programme builds will be trusted and used, or disputed and abandoned.

About VE3

VE3 is a UK-based enterprise AI, data, and digital transformation consultancy and Microsoft Solutions Partner. We work with enterprise clients to design and build the semantic layers and data foundations that make AI outputs trustworthy at scale. Our approach starts with the decisions the business needs to make, works back to the metrics those decisions depend on, and builds a governed, queryable definition set that any tool, model, or agent can rely on.