Context is not an afterthought.

Enterprise AI is failing at the architecture layer

MIT found that 95% of enterprise GenAI pilots return nothing measurable.1 The models are not the constraint. The data underneath them is. One use case works in a pilot. The same architecture buckles when it has to carry a hundred AI agents in production.

1MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025,” July 2025. Based on 52 executive interviews, surveys of 153 leaders, and analysis of 300 public AI deployments.

95%

of enterprise GenAI pilots return nothing measurable.

And they were. Someone ran the report, saw the number was wrong, and fixed it before anything happened. That person was the contextual data layer, which is why the systems underneath could stay fragmented. Agents remove the assumption.

Reads

A person opens a handful of records. An AI agent traverses thousands of entities in one reasoning step.

Conflicting data

A person knows which record is current. An AI agent picks one and acts.

Query pattern

A person queries occasionally. An AI agent queries continuously, in loops, across concurrent sessions.

Explainability

A person reconstructs their reasoning afterward. An AI agent can only do that if the sources were machine readable to begin with.

Nobody is left in the middle to catch the difference.

Arango holds graph, vector, document, key-value, and full-text search in one data platform, with one query language and no synchronization layer between separate specialized stores.

That decision was the founding idea for Arango. In 2014, connected data meant running three databases, because no single system handled graphs, documents, and key-value pairs well. Most teams ran all three and wrote pipelines to reconcile them. We built one instead.

One question. Two paths.

Which open disputes involve this customer, and which policy governs them?

One request

The traversal, the similarity ranking, and filtering resolve in the same transactional foundation, and the result carries its own lineage.

The difference compounds with every agent loop:
more round trips, more context assembled, more tokens spent carrying it.

Arango began in Cologne in 2014. The question was not how to store enterprise data, which was already well solved. It was how to represent what connects it.

Storage kept getting cheaper. Understanding kept getting harder.

That architecture went into production where relationships were the whole problem: fraud detection, cybersecurity, digital twins, supply chain, network and asset management, investigations. More than 200 organizations run it today, including HPE, PSI, Zscaler, and the U.S. Air Force.

Nobody called it a contextual data layer then. It is what it was, and it is what agentic AI now depends on. Twelve years of commits, releases, and documentation are public, so the architecture can be read in the order we built it.

AutoGraph

Automatically builds and maintains the graph from documents and unstructured data sources, so nobody hand-models a knowledge graph for every project.

Arango AutoRAG

Automatically adapts the retrieval strategy to each question across relationships, semantics, and documents, and returns lineage with every result.

LLM API MCP DB MLOps VPC

ArangoDB

The data foundation. Graph, document, vector, key-value, and search in one graph-native platform.

Retrieval tells you where the information is. It doesn’t tell you how it fits together, or whether acting on it is safe.

Six things have to be true before an agent can act:

One customer, not five near-duplicates.

Across customers, contracts, assets, policies.

What is true now, and what was true at the moment it acted.

Every result traces to a source.

The right retrieval path for the question, with evidence attached.

Relationships, documents, semantics, and operational data resolved together.

And all of it has to stay continuously updated and governed at the platform layer, not in the application above it. Twelve years of production taught us that list. Not the last two.

The Contextual Data Layer for Enterprise AI

6 architectural requirements for building agentic-AI-ready systems

What Arango is Not

Snowflake, BigQuery, and Databricks handle analytical workloads well. Quarterly aggregation is not what this platform is for.

Dashboards stay where they are.

Arango runs alongside existing data infrastructure and brings the context it needs from it.

Unlike vendors that solve only one piece of the AI context problem—data storage, knowledge graphs, metadata, or retrieval—Arango unifies the entire contextual data lifecycle on one native platform.

What we own is the connected, governed, current business context an agent reasons over, sitting between the enterprise systems and the model.

If your AI workload is one model against one clean table, you do not need us.

  • Articul8
  • Cloudera
  • Cloud Imperium
  • Cycode
  • Dun and Bradstreet
  • Deloitte
  • Emerson
  • ESRI
  • HPE
  • IC Manage
  • Johnson Controls
  • French Defense Ministry
  • Johns Hopkins
  • Linx
  • Kaseware
  • London Stock Exchange
  • NIH
  • Mercedes Benz
  • Orange
  • NVIDIA
  • Stockbit
  • US Air Force
  • Telstra
  • US Federal Government

A capable team can build a working AI use case in a few weeks. The architecture that ships the first one quickly is rarely the one that survives the hundredth, and early speed is usually borrowed against later scale.

Your advantage is your proprietary data. Your customers, your fraud patterns, your supply chain, whatever your team knows that nobody else does. It is almost never knowledge graph modeling, index synchronization, retrieval tuning, or entity resolution, and those are general problems a platform should solve once.

Bring us your hardest question.