Denodo has updated its data platform with expanded knowledge graph, semantic metric, and connectivity functions intended to give AI agents a more reliable account of what enterprise information means, where it came from, and how it may be used.

Platform 9.5 extends the Denodo Data Marketplace through a “360 graph” representing relationships between data, applications, notebooks, data pipelines, business glossaries, governance controls, contracts, sharing agreements, and AI skills. It also introduces metric views for defining business measures once and reusing them across analytics, data products, and AI applications.

Those capabilities address a problem that predates the current enthusiasm for autonomous agents. Large organisations frequently calculate apparently standard measures such as revenue, active customer, margin, or order value differently across departments and tools. An AI system connected to several versions may produce a confident answer without recognising that the definitions conflict.

Metric views allow formulas, filters, dimensions, relationships, and documentation to be stored inside a semantic layer. A dashboard, analyst, application, or AI agent can then draw on the same governed definition instead of recreating the measure separately.

Denodo has also extended its development assistant with conversational support for exploring metadata, generating its query language, refining logic, and troubleshooting data views. New or expanded connections include Databricks, Azure AI Search, vector search, Delta tables, and Iceberg environments.

The company’s Platform 9.5 overview describes these functions as “active context”: information supplied with business meaning, policy, lineage, relationships, and current operational relevance rather than retrieved as an isolated value.

Agents increase the cost of ambiguity

Semantic layers, data virtualisation, catalogues, knowledge graphs, business glossaries, and governed metrics are established parts of enterprise data management. Agentic AI has not created the need for context, but it raises the consequence of unresolved definitions because software may act on an answer rather than wait for a person to interpret it.

An error in an analytics dashboard may mislead a meeting. An agent using the same error could change a price, contact a customer, approve a workflow, move a case, or trigger another system. Reliable context therefore includes more than a shared definition: it must indicate freshness, authority, jurisdiction, confidence, ownership, and whether human approval is required before action.

Knowledge graphs can represent those relationships, although they are only as dependable as the metadata maintained around them. Many catalogue programmes become stale when ownership is unclear and operational teams do not update descriptions after a process changes.

Connecting more assets may provide a richer view while exposing how much of the estate lacks dependable documentation. An organisation cannot buy its way out of unclear data ownership by placing another interface over the top.

Metric governance creates an organisational problem as well as a technical one. Departments sometimes use different calculations for legitimate reasons. Finance may recognise revenue under accounting rules, sales may track bookings, and product teams may examine usage based value.

Imposing one definition can conceal those distinctions unless the semantic model records context and intended use. Governed consistency should prevent accidental disagreement without erasing measures that answer genuinely different questions.

Denodo’s architecture may appeal to organisations whose information is spread across operational systems, warehouses, lakehouses, search indexes, and cloud services. Logical access can reduce the need to copy every dataset before it becomes available, although live queries introduce performance and availability dependencies on the original systems.

An agent relying on operational data needs instructions for what to do when a source is slow, unavailable, stale, or only partly updated. Proceeding with incomplete context may be acceptable for a low risk draft but unacceptable for a payment, customer decision, or regulatory report.

Permissions become more complicated when an agent combines information. A user may be allowed to view two datasets separately while being prohibited from inferring a sensitive relationship between them. Retrieval controls, output filtering, prompt handling, and audit trails have to operate across the full chain rather than stopping at the source database.

Air Europa, which tested the graph capabilities in beta, says it uses Denodo as a central semantic and governed access layer. That provides a customer example rather than independent proof of performance, return, or implementation effort. Prospective users will have to test the release against their data quality, workloads, existing tools, and internal ownership.

Platform 9.5 reinforces a recognised direction in enterprise architecture rather than announcing that context has suddenly become important. The change lies in what consumes that context: when software can execute a task, missing lineage and ambiguous definitions become operational risks, while semantic work that once appeared bureaucratic forms part of the system controlling what an agent is allowed to do.