Context - The Missing Link Between Enterprise Data and Useful AI

Context - The Missing Link Between Enterprise Data and Useful AI

Oliver Boult

Oliver Boult

16 November 2026 - 11 min read

AIData
Context - The Missing Link Between Enterprise Data and Useful AI

Enterprise AI has made rapid progress on two fronts. Models have become markedly more capable, and organisations have invested heavily in connecting them to the data held in their warehouses, lakehouses and operational systems. Even so, many technology leaders are finding that an AI assistant with access to accurate data can still return answers that are inconsistent, wrong or unusable in a business setting. Asked for last quarter’s revenue, it may produce a figure that matches neither the finance pack nor the sales dashboard. It has no way of knowing whether the business means booked or recognised revenue, gross or net of returns, or whether a recently acquired subsidiary should be included. 

The missing ingredient is meaning. Gartner now describes the capability that supplies it as a context layer, and in October 2026 it named AI context platforms among its top strategic technology trends for 2027. 

This article explains why access to data is insufficient on its own, what a context layer contains, why it matters more as AI moves from answering questions to taking actions, where organisations already hold the raw material, and how to build a context layer around the use cases that matter most. 

At a glance 

  • AI models connected to accurate enterprise data can still give wrong answers when they lack the business meaning behind that data. 
  • A context layer holds that meaning in a governed, machine-readable form, covering definitions, relationships, lineage, usage policies and process knowledge. 
  • The need grows as AI agents move from answering questions to taking actions, where errors carry direct operational and regulatory consequences. 
  • Most large organisations already hold much of the raw material, spread across BI models, catalogues, glossaries and documentation. 
  • Building around a few priority use cases, with business ownership and measured accuracy, delivers value early and avoids a large up-front modelling programme. 

Access to data is only half the picture 

Enterprise data records facts. Tables capture transactions, customers, products and events, and schemas describe how those records are structured. Neither tells a reader what the data means in the language of the business. That meaning lives elsewhere: in agreed definitions, in business rules and their exceptions, in knowing which system holds the authoritative version of a record, and in the experience of people who have worked with the data for years. 

Humans fill these gaps without noticing. An analyst with five years in a finance team knows that the sales system records orders before credit checks, that one region reports in a different currency, and that a particular product code was retired and later reused. A model reading the same tables has none of that background. It makes reasonable assumptions, and reasonable assumptions about an unfamiliar organisation are frequently wrong. 

Inconsistency of this kind is deeply rooted in large organisations. When the National Audit Office examined data use across UK government, it found more than 20 ways of identifying individuals and businesses across 10 departments and agencies, with no standard format for recording basic details such as name, address and date of birth. Many private-sector organisations carry similar inconsistencies, accumulated through acquisitions, system replacements and departmental autonomy. 

Analysts now treat this as a central obstacle to enterprise AI. Gartner predicted in 2025 that through 2026 organisations will abandon 60% of AI projects unsupported by AI-ready data, and reported that 63% of organisations either lack data management suited to AI or are unsure whether theirs is ready. In May 2026, they then warned that a lack of semantics causes inaccurate AI agents and wasted spending. It advised data and analytics leaders to establish a context layer as a core component of their infrastructure, because schema-based data models alone lack the business context and meaning that agentic AI requires. 

Inside a context layer 

A context layer is the governed, machine-readable body of knowledge that sits between enterprise data and the AI systems that use it. It captures what an experienced employee would know about the data and makes that knowledge available consistently to every AI tool that needs it. Presenting the 2027 trends, Gartner’s Gene Alvarez compared a context layer to a new employee who learns how the business works over time and grows stronger as it is given better information. 

Implementations vary, and the components below illustrate what most context layers bring together. 

Business definitions and metrics: 

The agreed meaning of terms such as customer, active account, churn or revenue, including how each is calculated and which variations exist for different purposes. 

Relationships between entities: 

How customers, contracts, products, sites and suppliers connect to one another. Ontologies and knowledge graphs are common ways of expressing these relationships so that an AI system can reason across them. 

Metadata and lineage: 

Where each dataset came from, how it has been transformed, how recently it was refreshed and what is known about its quality. This allows an AI system to judge how far a figure can be relied on and to explain its sources. 

Policy and usage context: 

Classifications, retention rules and the purposes for which data may be used, so that an AI system understands the boundaries attached to the information it retrieves. 

Process knowledge: 

Business rules, exceptions and operating procedures that currently live in documents, wikis and the heads of experienced staff. 

This isn’t arguably new. Data teams have built glossaries, catalogues and semantic models for years, however what has changed is the consumer. Previously, those assets were designed for people who could read a definition, ask a colleague and apply judgement. A context layer that has to serve AI systems that apply whatever they are given literally and at scale, needs to be consistent, current and structured for machines to query. 

Why agents raise the stakes 

Business intelligence has always had a person in the loop. A dashboard showing an unexpected spike prompts questions before anyone acts, and an experienced reader often spots a definitional mismatch instinctively. Much of the cost of poor context has therefore been absorbed quietly by analysts reconciling figures before they reach a decision-maker. 

AI agents change that arrangement. As organisations deploy agents that draft customer communications, recommend pricing, triage cases or prepare regulatory returns, conclusions drawn from data increasingly flow straight into actions. A misread definition that once produced a confusing chart can now produce an incorrect refund, a misdirected letter or a submission built on the wrong figures, and the error repeats at machine speed until someone notices. Retrieval-based approaches, which fetch relevant documents and records at the moment a question is asked, reduce some of this risk by grounding answers in enterprise sources. They still depend on the system retrieving the right material and interpreting it correctly, which is exactly what a context layer supports. 

Without a dedicated context layer, organisations can face severe operational risks, including overprivileged agents, compliance failures, low trust and excessive token consumption.  

Token consumption is interesting and often overlooked – agents lacking good context compensate by retrieving and processing far more information than a task requires, which drives up running costs. Gartner predicts that organisations prioritising semantics in AI-ready data will improve agentic AI accuracy by up to 80% and reduce costs by up to 60%. 

Regulatory expectations point in the same direction. Gartner’s May 2026 analysis expects regulators to demand greater semantic transparency and boards to treat semantic governance as a strategic risk. For UK organisations in regulated sectors, showing which definition an AI system used, and where its figures came from, is likely to become part of demonstrating responsible use. 

The raw material most organisations already hold 

The scale of the task can look daunting, but large organisations rarely start from nothing. Most already hold much of the raw material for a context layer in semantic models built for BI tools, data catalogues and business glossaries created for governance programmes, master data that identifies authoritative customer, product and supplier records, data classification and protection policies, and a wealth of process documentation. 

The difficulty lies in the condition of this material. Definitions in the BI layer may contradict those in the glossary. Catalogues are often partially populated and drift out of date once the original project ends. Process knowledge sits in documents written for human readers and full of implicit assumptions. Different departments maintain their own versions of key terms, each correct for its own purpose. 

The hardest material to capture is tacit knowledge. Experienced staff know which figures to treat with caution in the last week of a quarter, which customer accounts run under special arrangements and which reports quietly exclude a business unit. Very little of this is written down, and much of it only surfaces when someone is asked to explain an anomaly. Structured conversations with subject matter experts, often prompted by the mistakes an early AI pilot makes, are one of the most effective ways of drawing it out. 

Building a context layer is therefore largely an exercise in reconciliation and governance. The organisation has to decide which definitions are authoritative, resolve conflicts between them, assign owners and keep everything current. That work depends on the disciplines covered in our earlier article, Why Data Governance Is the Foundation of Trustworthy Data at Scale, and organisations with mature governance will find a context layer considerably easier to build. 

Building context around priority use cases 

Attempts to model an entire enterprise before delivering any value is not the best choice. A more reliable route builds the context layer around a small number of priority AI use cases and extends it as the portfolio grows, mirroring the use-case-led approach set out in Enterprise Data Foundations: The Determinant of AI at Scale [link]. 

Starting from a use case clarifies scope. For example, a customer service assistant needs to understand customers, products, orders, entitlements and complaint categories, and a finance assistant needs revenue, cost, margin and the organisational hierarchy. Identifying the entities, metrics and rules each use case depends on gives the team a defined body of context to reconcile, and later use cases which can reuse.  

Each definition requires an accountable owner who can arbitrate when departments disagree, with data teams facilitating those decisions. Treating definitions as managed assets also helps, so versioning them so changes are deliberate, testing them so a metric that drifts outside expected bounds is flagged, and recording which version an AI system used when it produced an answer. 

Architecture is also a key consideration. Exposing context through standard interfaces, such as APIs and the Model Context Protocol that many AI tools now support, allows the same definitions to serve several assistants and agents and reduces dependence on any single AI product. Platform capabilities can accelerate the work, and our article Fabric IQ and Ontology: Giving Data a Shared Business Meaning looks at one example within Microsoft Fabric. 

Context also needs maintaining as the business changes. Whether that’s reorganisations, acquisitions, new products or regulatory updates, these all alter definitions and relationships, so the context layer needs a routine process for keeping it current, owned in the same way as any other business-critical data. 

Finally, measure the effect. A set of test questions with agreed correct answers, tracked before and after context is added, shows leaders how much accuracy has improved and where gaps remain. Evidence of that kind also strengthens the case for extending the context layer to the next use case. 

From capable models to useful AI 

Return to the leadership team and its conflicting revenue figures. A more powerful model would have made little difference, because the missing ingredient was knowledge about the organisation itself. With an agreed definition of revenue, clear lineage for the figures and an owner accountable for both, the assistant could have given the answer finance recognises and explained how it was derived. 

AI models will continue to improve, and their usefulness inside any organisation will still depend on how well they understand that organisation. The context layer is where that understanding is captured, governed and shared, and for technology leaders it is fast becoming as fundamental to enterprise AI as the data platform itself.

Ebook Available

How to maximise the performance of your existing systems

Free download
Oliver Boult

Oliver is Lead Data Consultant at Audacia, with over 5 years' experience delivering data platform and reporting solutions for clients across a variety of sectors including local government, emergency services, housing and transportation. He specialises in Azure, Purview, Fabric and Power BI, helping organisations improve the scalability, governance, and performance of their data estates.