How Daimler Trucks North America built a living knowledge graph of its business

A graph-based representation of everything happening on its internal network has enabled business users at Daimler Trucks North America (DTNA) to extract not just data, but knowledge.

According to AI Strategy and Architect Conor O'Shea, business leaders at the Portland, Oregon-headquartered company can now ask the company's new 'DTNA Chat' chatbot questions like, 'Analyze our recent warranty claims, correlate them with all available datasets, and tell me why they've increased.'

The underlying principle: AI and graphs surface the inner truth of the business by revealing how everything is related, creating a model that can reason and act effectively.

Key to that capability, O'Shea adds, is the power to represent DTNA activities over time and across all its data and processes.

Specific benefits of using Neo4j as the graph foundation include the ability to ask the system how customers are likely to respond to any proposed component design upgrade, how the change affects production, and what downstream effects might occur.

Being able to ask useful questions like this and get answers back also means his users don't have to work across multiple systems - or, as he jokes, 'climb The Wall of Awful.'

Rich legacy

O'Shea works for the North American subsidiary of the world's largest maker of commercial vehicles, Germany's €55bn ($65bn) Daimler Truck AG.

It distributes a range of medium- to heavy trucks under brand names including Freightliner and Western Star, as well as the Thomas Built school bus line.

The organization was originally Portland, Oregon-based trucking company Consolidated Freightways, whose roots stretch back to the 1930s. Freightliner was acquired by Daimler in 1981, with the name switch to Daimler Trucks North America occurring in 2008.

To make things more complicated, Daimler Truck (including DTNA) spun off from Daimler AG (now Mercedes-Benz Group) in late 2021.

The divestiture proved to be an important catalyst for adopting graph technology, says O'Shea.

He had been experimenting with the free version of Neo4j simply to explore better ways of representing the company's systems:

I had been introduced to Neo4j's graph technology while working on an unrelated challenge, and saw it could help us discover relationships between datasets that didn't appear connected.

Nice, but no pressing use case presented itself at that time. However, 2021 meant separating the company from a single entity into two fully independent organizations.

Nearly all the applications and architecture were integrated across both divisions, so splitting them cleanly without accidentally breaking critical systems on either side was going to be a major challenge.

Documentation existed, O'Shea notes, but documentation "ages" quickly - after a month or two, its accuracy drops dramatically.

In his view, to understand what an application is actually doing, the business needs a reliable picture of how it behaves today, not how someone thought it behaved six months ago.

Building such a picture was initially expected to be the job of an external consultancy - estimated to cost an easy seven figures.

Such a picture would also be historic to at least some extent, and filtered through non-DTNA eyes - so it could end up not being a true snapshot anyway.

He says:

If they missed even one obscure dependency - say, a weekly SMTP batch that turns out to be critical for the supply chain - the cutover could cripple operations.

To avoid both issues, O'Shea pitched the idea to his bosses of building a knowledge graph of the real interactions and relationships across all DTNA systems.

Out of nowhere, LLMs had also appeared - suggesting to him the possibility of automating all the discovery work.

This proved highly useful, he says. Not only was the consulting fee avoided, but the organization also learned there were many things in its environment it simply hadn't been aware of.

Channeling his inner Donald Rumsfeld, for O'Shea this was a process of discovering:

We simply didn't know what we didn't know.

A new way of looking at business operations in real-time

To build this AI-fronted IT infrastructure and application stack knowledge graph, O'Shea built an ETL pipeline that ingested all DTNA network traffic.

He used network monitoring tool ExtraHop to continuously gather network flow records.

Every five seconds, the pipeline pulls in around 2,400 flow records on a 24x7 basis, which are then fed into Neo4j's cloud-native graph database-as-a-service, Aura, running in AWS.

This process turns all that data into a graph of the current live DTNA system, which is then interrogated through a natural language chat interface using Claude.

Completing the picture is use of Model Context Protocol, which allows the firm to present the LLM with data from not just the network but multiple business data sources and applications.

The result, O'Shea claims, is an accurate view of the actual state of the business in real-time that users can access without ever having to learn database query languages.

And it turned out there really were 'unknown unknowns' that could have complicated the creation of an independent company:

We surfaced undocumented SMTP batch jobs tied to unnamed applications, SSH automation scripts triggering proxy backups to servers still managed by the company we were leaving, and numerous other hidden dependencies.

Each one of those could have been a potential failure point in the separation process, but thanks to the graph, each was identified, categorized, and assigned to the proper teams before the final cutover. We had finally achieved a complete understanding of our environment.

Could he have done all this without a graph? For O'Shea, meaning is not in the things themselves - it's in the relationships between them. So, no:

Traditional relational tables obscure relationships; graphs surface them directly. This lets us see not only what our business does, but how everything is connected, and therefore how changes ripple through the organization.

Next steps

Perhaps unsurprisingly, O'Shea sees the next step for graph, LLM and knowledge at his company to be agent-shaped.

This will come, he hopes, once the model can identify problems, propose fixes, and even apply them.

He says:

Instead of developers manually patching a script, the model might ask, 'Would you like me to fix this for you?'

The future we're building toward is agent orchestration integrated with real business data, enabling systems that not only describe what's happening but also maintain and improve themselves.

Summing up his experiences, O'Shea concludes:

By itself, unless it's connected through time and causality, data is meaningless. By combining graph technology with large language models, we turn live enterprise activity into knowledge so that the organization can understand what it's actually doing right now, make faster decisions, and solve problems in minutes instead of weeks.

AI is what allows us to democratize complex systems so people in Finance, HR, or Operations can simply ask questions and get real answers, instead of relying on specialists, tickets, or outdated documentation.

Ultimately, technology only matters if everyday employees can use it effectively from day one.