Home Articles All the Data. None of the Context.

All the Data. None
of the Context.

17 Mins | Nov 13, 2026 | by Editorial Team

At a Glance

Enterprise AI context is the business knowledge that helps AI systems interpret data correctly. It includes approved metric definitions, relationships between records, operational rules and information about which sources are authoritative. Semantic layers, knowledge graphs and context-aware retrieval make this information accessible to AI applications, reducing the risk of technically correct but misleading answers.

Why AI can retrieve the right information and still misunderstand the business.

An AI assistant can write a perfectly valid SQL query, retrieve thousands of records and produce a convincing analysis. Every number in its response might be accurate.

The answer could still be wrong.

Consider a question that appears straightforward: How many active customers did we have last quarter?

The CRM might count every account marked active. Finance may recognize only customers with completed transactions. Sales could include accounts with open opportunities, while operations might exclude customers whose services are temporarily suspended.

None of these definitions is necessarily incorrect. They serve different purposes.

An experienced employee usually knows which definition applies, which report the organization considers authoritative and when an exception needs to be accounted for. An AI assistant does not acquire that knowledge simply by connecting to the company’s databases.

This is a fundamental weakness in many enterprise AI implementations. Organizations have spent years collecting information, yet much of the knowledge required to interpret it remains dispersed across applications, departments and the people who understand how the business operates.

Recent developments suggest that technology providers are beginning to address this problem more deliberately.

Why enterprise context is attracting attention

On October 16, Atlassian published an examination of context graphs in enterprise AI, describing how information about business processes is distributed across project-management systems, documents, code repositories and everyday conversations.

Consider an AI assistant investigating why a product release has been delayed. It may find a deadline in Jira, an engineering discussion in Slack and an architectural decision in Confluence.

Retrieving those records is relatively straightforward. Understanding how the engineering issue affects the release schedule, who made the relevant decision and which dependencies remain unresolved requires additional context.

Atlassian’s Teamwork Graph is designed to represent connections between people, projects, decisions and the work being performed, making that information available to AI applications.

On October 20, Neo4j announced its Context Engine, an approach to building a reusable knowledge layer for enterprise AI.

The idea is to represent important business entities, processes and relationships in a shared graph that multiple applications can reference. Individual agents would no longer need to reconstruct the same relationships independently from separate data sources.

The platform remains in private preview, so its performance across broader production environments is yet to be established.

A different approach emerged on October 26, when Oracle announced plans for a Business Semantic Layer within its Autonomous AI Lakehouse.

The proposed capability will make definitions of business metrics, calculations, hierarchies and KPIs available to AI applications. Rather than independently deciding how a metric should be calculated, an assistant could use definitions already established by the organization.

Oracle expects these capabilities to become available over the following 12 months.

These announcements approach enterprise context from different directions. Knowledge graphs help represent relationships. Semantic layers establish consistent meanings. Both address a problem that becomes more consequential when AI is expected to support business analysis and operational decisions.

What happens when AI understands the schema but misses the meaning?

On October 28, AtScale published an evaluation of Text-to-SQL accuracy using the BIRD-Interact benchmark.

The company tested seven language models against 410 business questions, comparing their performance with database schema documentation against their performance when a semantic layer was introduced.

Without the semantic layer, reported accuracy ranged from 21.5% to 36.1%.

With the semantic layer, it ranged from 44.6% to 62.2%.

The reported improvement was between 20 and 31 percentage points across the models tested.

These results deserve some qualification. AtScale conducted the evaluation itself, and the findings were not independently verified by the BIRD benchmark team. The semantic-layer results also used the best of three runs, while the baseline relied on a single run. That difference limits the strength of a direct comparison.

Even with those limitations, the study raises an important engineering question.

The models already had access to database schemas, column descriptions and supporting documentation. Yet access to that information did not consistently produce correct answers to business questions.

A schema can tell an AI model that a table contains customer records. It may not tell the model which customers count as active, how duplicate accounts should be handled or whether a particular status value has been deprecated.

Those details determine whether the resulting query reflects the business question.

The implication is worth examining before investing in a more capable model: some AI accuracy problems originate in how business knowledge is represented and supplied to the model, rather than in the model’s ability to reason.

Where enterprise knowledge gets lost

Most enterprise systems were built to serve specific operational needs.

A CRM manages customer relationships. An ERP records transactions. Product information may live in a separate catalogue, while support interactions are stored elsewhere.

Each application can work correctly within its own boundaries, even when it uses definitions that differ from those in another system.

A customer may have one identifier in the CRM and another in the billing system. Revenue may refer to invoiced amounts in one report and recognized revenue in another. A product might belong to different categories depending on whether the business is examining inventory, merchandising or financial performance.

Integration can move information between these systems, but moving data does not necessarily resolve differences in meaning.

Some of those differences are intentional. Finance and sales may legitimately require different measures of customer activity.

Others emerge because business rules have changed, older documentation remains available or decisions were never formally captured.

Employees familiar with these systems often compensate for the gaps. They know which numbers to question and which sources to trust.

An AI application needs that knowledge to be represented more explicitly.

Otherwise, it may select a technically valid dataset, apply a reasonable-looking calculation and return an answer that conflicts with how the organization actually measures its business.

Building the context AI needs

There is no single technology that provides complete enterprise context. Different problems require different architectural approaches.

Four capabilities are particularly relevant.

1. Semantic layers: Agreeing on what the numbers mean

Suppose a subscription business wants to measure customer churn.

One department defines churn as the number of customers who cancel. Another measures revenue lost through cancellations. A third includes customers who downgrade their subscriptions.

All three measures may be useful. They are not interchangeable.

A semantic layer allows the organization to define these measures explicitly, including their calculations, applicable conditions and underlying data sources.

When an employee asks an AI assistant about churn, the application can reference the approved definition. If the question is ambiguous, it can ask whether the user means customer churn or revenue churn.

This is particularly valuable in natural-language analytics, where users do not specify database tables, joins or calculation logic.

A semantic layer reduces the need for each AI application to reconstruct those decisions independently.

It also makes ownership possible. Business teams can review definitions, record changes and establish which measures should be used for particular purposes.

2. Knowledge graphs: Making business relationships visible

Semantic layers establish meaning. Knowledge graphs help represent relationships between entities.

Consider a manufacturer investigating a delayed customer shipment.

The relevant information may include the customer order, a production batch, a supplier delivery, an equipment failure and the revised dispatch schedule.

A document search might retrieve records mentioning each of those events. What matters is how the events connect.

A knowledge graph can represent which supplier provided a component, which production batch used it and which customer orders depend on that batch.

An AI application can then follow these relationships rather than attempting to infer every connection from unstructured text.

This becomes useful when business questions span several departments or require reasoning through dependencies.

A knowledge graph does not have to duplicate every record. It can represent important relationships while the detailed information remains in existing enterprise systems.

3. Context-aware retrieval: Finding the right evidence

Retrieval-Augmented Generation (RAG) allows language models to use information from external sources when producing an answer.

For enterprise applications, this is often a practical way to work with internal documents, policies and operational knowledge.

But a document that matches a search query is not necessarily the document that should govern a business decision.

An outdated discount policy may contain the exact terminology used in a customer’s request. A newer policy may use different language while superseding the earlier version.

A retrieval system that selects information primarily by textual similarity could return the wrong source.

Reliable retrieval therefore needs to consider more than relevance. The application may need to know who owns a document, whether it has been superseded, which business unit it applies to and whether the requesting user is authorized to see it.

This is part of the broader discipline of context engineering: deciding what information a model needs for a task, how that information should be organized and how its reliability can be assessed.

It also requires restraint. Providing an AI model with every potentially related document can introduce contradictory instructions, outdated information and unnecessary processing costs.

Useful context is selected and maintained, rather than accumulated indiscriminately.

4. Operational context: Knowing what is true right now

Historical knowledge and established definitions are only part of what an enterprise application needs.

An AI agent processing a purchase order may require current inventory availability, customer credit limits and the status of an existing transaction.

Those values can change throughout the day.

On October 29, Lenovo published an Enterprise Context Layer reference architecture describing how agents can receive task-relevant information while the underlying enterprise systems retain authority over their data and operations.

The architecture addresses the provenance and freshness of information, along with controlled access to enterprise systems.

An agent should be able to distinguish between a current inventory value retrieved from an operational application and a historical figure found in an older report.

It must also operate within the permissions associated with the task.

Lenovo’s publication describes a proposed architecture, rather than independently measured production improvements. Still, it illustrates an important design principle: the context supplied to AI must reflect the state of the business at the time the task is performed.

The less visible work behind context-aware AI

The architectural discussion can make enterprise context sound like a problem that can be solved by introducing another platform.

In practice, much of the work involves organizing information, establishing ownership and maintaining business definitions.

One of the harder problems is knowledge that has never been formally documented. Business teams often operate with conventions, exceptions and definitions that experienced employees understand but enterprise systems do not capture.

Making this knowledge available to AI requires a process for eliciting it from stakeholders, resolving conflicting interpretations and maintaining an approved version as business practices change. A semantic layer or knowledge graph can help organize and apply that knowledge, but neither can independently establish the correct business rules when those rules have never been captured or validated.

What this looks like in practice

A Nineleaps engagement for a leading apparel and activewear brand illustrates this challenge.

The company had sales data distributed across four channels, with inconsistent units and reporting logic. Important KPI definitions and business rules remained dependent on institutional knowledge held by employees, making it difficult to obtain consistent answers across the organization.

Addressing the problem required work on the data and its interpretation.

Nineleaps unified and reconciled the sales information into a governed analytics foundation comprising more than 20 analytics-ready tables. Business knowledge was codified into a semantic layer, with approximately 100 business questions modeled to support the information needs of different users.

The resulting solution enabled employees to ask questions in plain English through a role-controlled GenAI system operating within India’s data boundary.

The significance of this approach lies in the work behind the conversational interface. Consolidating the data was necessary, but the application also needed a consistent understanding of the business definitions and reporting logic that employees relied upon.

It demonstrates why developing enterprise AI often requires substantial data and knowledge engineering before the model can produce useful answers.

Explore the Nineleaps case study: Empowering Every Level With Instant Access to Actionable Intelligence

Other organizations are addressing related challenges through their knowledge-management practices.

Atlassian’s October 16 article described how HR software company Personio consolidated organizational knowledge into Confluence.

Knowledge-base pages are assigned active owners, helping ensure that information remains current and available to employees and AI tools.

Atlassian reported a 60% increase in knowledge-base usage.

That is a measure of adoption, not evidence of improved AI accuracy. The more useful observation is that the organization established a process for keeping knowledge discoverable and maintained.

Another example appeared in Oracle’s October 26 announcement.

Energy services company Liberty Energy is using Autonomous AI Lakehouse to bring together financial, operational, supply-chain and workforce data across a multicloud environment.

According to Liberty Energy, the architecture gives decision-makers visibility into field revenue, inventory usage, costs and profitability.

The information comes from multiple technologies, including Amazon S3, Microsoft SQL Server, MongoDB and PostgreSQL.

Oracle did not publish an independently validated improvement in AI-agent accuracy for this implementation. It is better understood as an example of the data integration work that can support more dependable analytics and AI applications.

Together, these examples highlight an important dependency.

Before an AI system can consistently use institutional knowledge, an organization needs to determine where that knowledge lives, who maintains it and which information should take precedence when sources disagree.

A retail example: When the correct calculation tells an incomplete story

Imagine a retailer asking its AI analytics assistant:

Why did sales decline in the western region last month?

A basic implementation might retrieve monthly sales totals and calculate the percentage difference.

That answers how much sales changed. It does not establish why.

A context-aware system would begin by resolving the business terms in the question.

Does sales mean gross sales or net revenue? Which stores belong to the western region? Are returns attributed to the purchase month or the month in which they were processed?

Once the relevant definitions are established, the application can retrieve information from the appropriate systems.

It might examine inventory availability, promotional activity, returns and transaction patterns. Relationships between products, stores and suppliers could help identify which operational events coincide with the decline.

Suppose several high-volume products were unavailable during the period.

The AI could present the affected products, the duration of the shortages and the corresponding sales patterns. It should also acknowledge that a correlation between shortages and declining sales does not, by itself, prove causation.

Further analysis might be required to determine how much of the decline can reasonably be attributed to inventory availability.

This example is illustrative, but it shows where context improves the quality of an answer.

The model still performs analysis. The surrounding information architecture helps establish which questions to ask, which evidence is relevant and how confidently the findings can be interpreted.

Does every enterprise need a knowledge graph?

No.

The interest in context engineering should not lead organizations to introduce complexity without a clear purpose.

A document assistant working with a small, well-maintained policy repository may perform adequately using conventional retrieval and metadata filters.

An analytics assistant working across several financial databases may benefit more from a governed semantic layer.

An application investigating dependencies between suppliers, production processes and customer orders might justify graph-based retrieval.

These approaches can complement one another, but they are not all necessary for every application.

The appropriate choice depends on the problem the existing system is failing to solve.

Is it retrieving outdated information? Confusing different business definitions? Missing relationships between records? Using information the requesting employee should not be able to access?

Each failure points toward a different intervention.

A knowledge graph may improve relationship discovery while doing little to correct an inaccurate financial calculation. A semantic layer can standardize metrics without resolving the challenge of retrieving a current operational record.

And introducing a shared context layer creates responsibilities of its own.

If an incorrect definition is made available to several AI applications, the mistake can spread across multiple workflows. Reusable business context needs ownership, versioning, access controls and regular validation.

The value comes from maintaining a dependable source of knowledge, not simply from connecting more systems.

Start with a question the business already struggles to answer

Organizations do not need to resolve every data-quality problem before improving their AI applications.

A more practical starting point is a business question that employees already answer inconsistently.

Return to the question of active customers.

Ask the relevant teams to agree on the definitions they use and when each definition applies. Identify the authoritative data sources, document important exceptions and establish who is responsible for maintaining that information.

Then assemble a small evaluation set of representative questions with answers reviewed by the appropriate business owners.

Test how the AI application performs before introducing additional context.

Does it choose the correct metric? Can it identify the source of its answer? Does it recognize when the question is ambiguous?

Once the necessary definitions and relationships are available through a semantic layer, retrieval system or other appropriate mechanism, repeat the evaluation.

The comparison can reveal whether the changes improve accuracy, consistency and traceability.

It also helps teams understand whether the original problem arose from model limitations or from missing business knowledge.

This is a more defensible basis for investment than assuming that either a larger model or a new context platform will improve every workflow.

The knowledge already exists. The challenge is making it usable.

October’s developments reveal a common concern among enterprise technology providers.

Atlassian is connecting organizational knowledge through its Teamwork Graph. Neo4j is developing infrastructure for reusable business relationships. Oracle is introducing semantic capabilities intended to provide consistent definitions across analytical and AI workloads.

AtScale’s evaluation offers preliminary evidence that this kind of structured business context can improve natural-language analytics, although its methodology limits how broadly the findings can be applied.

The technologies differ, and their results will depend on implementation. They nevertheless address a problem that many organizations already recognize: critical business knowledge is dispersed across systems that were built for different purposes.

For enterprise AI teams, the immediate task is to determine which definitions, relationships and operational facts an application needs to perform reliably.

That requires collaboration between data engineers and the people responsible for the underlying business processes.

It may also require resolving disagreements that existing reporting systems have allowed to persist.

If three departments cannot agree on what counts as an active customer, connecting an AI model to all three databases will not settle the question.

The organization must first decide which definition applies, then make that decision available to the systems expected to use it.


About The AI Delivery Layer

The AI Delivery Layer is a monthly newsletter by Nineleaps, exploring the technologies, architectures and engineering decisions shaping enterprise AI.

Each edition examines emerging developments and their practical implications for organizations building and operating intelligent systems.

A question for enterprise leaders: If three teams in your organization were asked to define the same business metric, would they arrive at the same answer?

Share your perspective in the comments and subscribe to The AI Delivery Layer for future editions.