The learnings in this blog post are based on the session, “How Mercedes turns Teamwork Graph into an AI advantage”, presented at Atlassian’s Team ’26 conference. You can check out this session and others on demand.
Foundation models have become astonishingly capable over a very short period. At the beginning of this decade, they could barely look up facts online. Today, they can produce work that would take a human engineer more than an hour, and the pace of improvement has been exponential, especially over the last 18 months.
And yet, something strange is happening.
When you ask individual users whether AI tools have helped them, more than 50% say “yes”, that they’ve actually saved time and effort. But when you zoom out and ask companies whether they’ve seen dramatic improvements to their business, 96% report that they have not. The technology is there, but unlocking the enterprise-level value has been stubbornly difficult.
Now, Mercedes-Benz’s recent experience proves the harder enterprise challenge is not whether AI can generate output; it’s whether that output can connect to the messy reality of work across tools, teams, data, and decisions.
The real bottleneck isn’t the model
Work inside organizations depends on collaboration, so when enterprise AI treats it like an individual activity, it will always fall short of its promise.
The average large enterprise runs over 367 SaaS applications, since often multiple teams use their own tools and follow their own rituals. When teams need to collaborate, people find themselves navigating unfamiliar interfaces, reconciling different planning cycles, and chasing down context scattered across a dozen systems. Goals developed in isolation can leave teams rowing in opposite directions, even when everyone has the best intentions.
“You have to use at least 11-plus tools to accomplish a task, or even more — like, I don’t know, 100-plus, 300-plus tools in the company you have to maintain, use, know, get data out of there, use it for your next task somewhere. And it’s just horrible.”
–Tobias Langjahr, Product Manager, Mercedes-Benz
With AI entering the picture, this complexity doesn’t automatically decrease. In fact, it can actually increase. How does a coding agent get the context it needs without creating churn for the rest of the team? How does it coordinate with a design agent running on a different platform? Without a shared layer of context, you end up with faster output that still requires extensive review and rework.
Without a shared layer of context, you end up with faster output that still requires extensive review and rework.
How Mercedes-Benz is applying this to defect management
Quality has always been central to how Mercedes-Benz defines its vehicles, but the key differentiator has shifted from purely hardware characteristics to the software that runs throughout modern cars. The amount of data generated by connected vehicles, test benches, and development environments is enormous, and the challenge is making that data useful in the flow of engineering work.
Like many large engineering organizations, Mercedes-Benz found that its engineers needed to use 11 or more tools to accomplish a single task, with over 300 tools in active use across the company. The time spent navigating between systems, transferring context, and remembering workflows was time not spent on what engineers are actually hired to do: build high-quality products and innovate.
“I want to have as many defects filed as possible, because we can focus on the right things. And all the other things, we don’t have to care about, because AI, the graph, Rovo is handling the rest.”
–Tobias Langjahr, Product Manager, Mercedes-Benz
The defect intake process illustrates this clearly. A year ago, when an engineer testing a vehicle detected a problem, the process depended entirely on the engineer’s diligence in documenting it. There was no automatic enrichment. Once filed, a defect manager would review the information provided and try to determine which team should own it, often leading to cycles of back-and-forth before the right group was even identified. When a defect was eventually resolved, the solution and context lived in a silo that other teams couldn’t access or learn from.
Connecting the dots across engineering systems
Mercedes-Benz recognized that if they could connect their various data sources through the Teamwork Graph, they could re-imagine the entire defect lifecycle. They brought together requirements data from IBM Doors, code repositories, release management information, physical part and asset data, and telemetry from test vehicles streaming log data to a central data hub.
With these sources connected through the Teamwork Graph, Atlassian’s data intelligence layer, the entire set of relationships between a defect, the requirement it traces to, the code that implements it, and the vehicle configuration where it was observed becomes traversable.
The Norris family of agents
With this connected data layer in place, Mercedes-Benz addressed a practical friction point: the physical act of filing a defect during vehicle testing. An engineer driving at highway speeds on the Autobahn who notices an error message on the display can’t safely stop to file a detailed ticket, so Mercedes-Benz built a suite of AI agents in Rovo (informally called the “Norris family”) to transform their defect management workflow.
Defect Norris handles quality assurance at intake. Analyze the Snores reaches out to central data hubs to pull log data and telemetry related to the specific vehicle and event. A third agent handles cross-vehicle-line analysis, checking whether a defect found in one vehicle model might also affect other models that share components or software. And the fourth agent, Main Norris, handles the regulatory and safety context.
Making defect filing as easy as speaking
One of the practical friction points Mercedes-Benz addressed is the physical act of filing a defect during vehicle testing. An engineer driving at highway speeds on the Autobahn who notices an error message on the display can’t safely stop to file a detailed ticket.
“If you imagine driving a car, testing something, you’re on a highway or a German Autobahn, which is both fun — how do you manage to remember what was happening if you detected something that was not working? You have to stop your car basically and note it down because otherwise you cannot remember after you’re back to the office.”
–Tobias Langjahr, Product Manager, Mercedes-Benz
To solve this, Mercedes-Benz built an in-car voice application that allows engineers to describe a defect using natural speech while driving. The app converts speech to text, associates it with the vehicle identification number, and ships it directly to Jira. Defect Norris then takes over, structuring and enriching what started as a simple spoken description into a fully formed defect record with proper metadata, component identification, and linked vehicle data.
Results and the road ahead
The impact has been significant. By measuring the lead time from when a defect is detected to when it is fixed, Mercedes-Benz has achieved a 70% reduction compared to its previous approach. They began this work with the development of the new S-Class and established a clear baseline before the new process went live.
The philosophy driving this is straightforward: the easier you make it to file defects, the more defects get filed, and that turns out to be a good thing when AI and the graph can automatically handle the heavy lifting of enrichment, duplicate detection, routing, and prioritization. Engineers can focus on the defects that truly require human expertise and innovation while the system handles everything else.
Looking ahead, Mercedes-Benz envisions extending this approach well beyond defect management. The broader ambition is a kind of digital twin for engineering work, where AI, powered by the Teamwork Graph, handles 80-90% of routine effort, freeing people to focus on delivering genuine value rather than wandering between tools searching for the right context or the right person to talk to.
And while functional safety regulations appropriately limit how much autonomy AI can have in safety-critical automotive systems today, connected data, rich context, and intelligent automation will continue to compress the distance between identifying a problem and resolving it.
What AI actually needs to deliver enterprise value
Mercedes-Benz’s story is a powerful example of what AI can do at an enterprise, organization-wide level instead of just at an individual productivity level. And for AI to work effectively at that organizational level, it requires three things.
- The richer the context you can provide, the better the result. If you ask an AI tool to help you prepare for a meeting with a specific client, it needs to understand how decisions have been made for similar customers in the past, the current state of the relationship, and your team’s goals.
- AI becomes truly valuable when it doesn’t just answer questions but participates in workflows. Generating a weekly status update, creating a page, notifying stakeholders, or triaging an incoming ticket are all examples of actions that free people to spend their time on higher-value work.
- When AI is taking actions on your behalf, you need to be confident that it honors your organization’s permissions and privacy policies. If a user doesn’t have clearance to see a particular Jira ticket, the AI shouldn’t know it exists either.
Co-building with the graph
This is why we built the Teamwork Graph. While Teamwork Graph captures the most important use cases across industries, almost every organization has its own unique tools and workflows. A home loan company might need to bring data from a specialized testing platform. An aerospace company might want to connect its project management tool. A social media company might need to surface data from a custom sales system.
To handle this, the Atlassian platform supports building with the graph directly. It currently offers more than 100 out-of-the-box connectors, and organizations can build their own custom connectors to bring data from proprietary or legacy systems into the graph. Once data is ingested, it becomes searchable and actionable across Atlassian products, and it can power custom agents and applications built on top of the graph.
Experience the power of the Teamwork Graph for yourself and see your graph preview.
