Every airline ticket you bought this year likely had its price calculated by a graph database. So did the cost of your hotel room at Marriott. 

This is according to Emil Eifrem, founder and chief executive of Neo4j, who told The Stack that graph databases, which map relationships between data points rather than storing them in rigid tables, have quietly become essential infrastructure for some of the world's largest companies. 

And as enterprises struggle to make artificial intelligence work reliably, graph databases are emerging from the back office into the spotlight. Eifrem sat down with The Stack's Nishal Ratanji to chat through how and why.

Three AI hiccups

"Where I see most organisations stumble when it comes to AI is a few things," said Emil Eifrem, founder and chief executive of Neo4j, the company that coined the term "graph database" two decades ago. 

"One is hallucinations,” Emil said. Another is explainability: “AI is a black box. You throw your data in there, you get some outcome, you have no idea why.” The third? “AI does not respect the kind of enterprise permission structure and governance structure. Graphs solve all of those three things."

The claim comes at a pivotal moment for corporate technology leaders: A 2024 survey found that 47% of enterprise AI users admitted to making at least one major business decision based on hallucinated content. 

For regulated industries, the stakes are even higher to improve how they execute: the European Union's AI Act, which took effect in August 2024, requires companies to demonstrate transparency and explainability in high-risk AI applications, with full compliance required by August 2026.

The Enterprise AI trilemma

Eifrem says conventional AI architectures are struggling to address the challenges. But “graph databases give you visibility into exactly what data the LLM is using to make its decisions. That gives you explainability and auditability, which are critical – especially in regulated industries, such as  government, financial services, and insurance,” said Eifrem.

“Our customers find this visibility valuable across the board."

Data governance is also mission-critical for enterprise AI applications.

Eifrem said: "We need to have AI systems that respect the data governance that is already in place in most big enterprise organisations. For example, in healthcare, you want to be able to say that the receptionist can get access to the patient's social security number and phone number, but not their medical records. But the doctor gets access to the full medical history..."

Industry analysts are reaching similar conclusions. A 2025 benchmark by FalkorDB, a graph database company, found that traditional vector-based retrieval systems scored effectively zero percent on complex queries involving aggregations and forecasts, while graph-based implementations achieved over 90% accuracy. 

Microsoft open-sourced its GraphRAG framework in 2024, signaling broader industry validation of the approach.

Knowledge graphs as AI's memory

The technical case for graphs centers on how they represent corporate knowledge. Unlike traditional databases that store information in tables, or vector databases that convert documents into mathematical representations, graph databases explicitly map relationships between entities, customers, products, transactions, employees, creating what the industry calls a "knowledge graph."

"In AI, there's a broad understanding now that knowledge graphs are the superior way of representing your internal company's information in a way that makes sense to AI models, in a way that makes sense to LLMs," said Eifrem. "On some level, that's the killer use case for AI in the enterprise. We all want AI to be able to operate on a company's internal data. And a knowledge graph is a superior way of doing that."

Being able to see connections provides a window into the explainability of why an AI model has made a decision. "With a graph database, you can see why an AI model chooses this data and this support article, because a particular engineer who authored the article is highly ranked and the support article is such a great match with the content," said Eifrem. "So you get that explainability of the source of the data that you hand off into the LLM."

The EU AI Act defines transparency as ensuring "AI systems are developed and used in a way that allows appropriate traceability and explainability, while making humans aware that they communicate or interact with an AI system." For enterprises subject to the regulation, the ability to trace AI decisions back to specific data sources is necessary for compliance.

The adoption barrier falls

For years, graph databases faced a practical obstacle: they required specialized skills that most enterprise developers lacked. That barrier is now collapsing, Eifrem asserts, thanks to AI itself.

"Historically, there's been two kinds of stumbling blocks which have been the hardest pieces to learn when you need to get up and going with a graph database," said Eifrem. "The first one is, how do you model your data in a graph? Everyone has been taught in school to do SQL modelling. It turns out, AI significantly helps with both of them."

Neo4j now offers automated AI tools that examine existing databases and propose graph models automatically. "You point it to your Oracle database, Snowflake, Databricks, Postgres, or MongoDB, and it will look at the data in there, and then propose a graph model," said Eifrem. "You can tweak it, and then you just hit import.” 

He adds: “It'll move the data over, and you have the graph up and running in just two clicks."

The result, according to Eifrem, is a democratization of what was once specialist technology. "We're hearing that from customers a lot. It used to be that we had this centre of excellence for graph experts, but now all developers can use Neo4j."

The technology's maturation received formal recognition in April 2024, when the International Organization for Standardization (ISO) published GQL, the Graph Query Language – marking the first new ISO database language since SQL was standardized in 1987. 

The standard, developed with participation from Neo4j, Oracle, TigerGraph, and academic institutions, is designed to give enterprises confidence that their investments will remain portable across vendors.

Business cases first, please

Despite the technical momentum, Eifrem offers a cautionary note for technology leaders tempted to adopt graph databases simply because they represent the latest approach to AI infrastructure.

"When you really want to solve a problem, it has to start with a business outcome," said Eifrem. "I need to understand my customer journey better so that I can target them and segment my customer base, for example. Or I need to understand a data element's journey throughout my organisation so that I can be compliant with all of the regulatory frameworks in my industry. Or I need to map out my IT infrastructure so I understand the blast radius when we get hit with a cyber breach."

The warning reflects hard-won experience. "If it starts with data ahead of time ('because graphs are cool, I'm just gonna start using them') I think that's an anti-pattern," said Eifrem. "It has to be that the business problem drags an application into it that solves the business problem. That application then drags the best data architecture in."

Neo4j claims customers including all 20 of the largest U.S. banks, nine of the ten largest pharmaceutical companies, and eight of the ten largest retailers. 

But Eifrem points to an unlikely proof point for the technology's power: the Panama Papers investigation of 2016, when journalists used Neo4j to trace hidden financial relationships across 1.6 terabytes of leaked documents, ultimately uncovering billions of dollars in tax fraud.

"They could never have found that without a graph view of relationships and information," said Eifrem. "A knock-on effect of that was that the day after that happened, every single IRS equivalent –all the tax authorities of the world – called us and wanted help. But not just that, all the big banks as well. Because they were like, ‘we don't want to be used for this.’"

For enterprise technology leaders weighing their AI infrastructure investments, the message is clear: the same capability that exposed hidden networks of financial fraud may now be essential to making corporate AI systems trustworthy enough for production use.

Delivered in partnership with AWS and Neo4j.

The link has been copied!