How software engineering is changing at Mercedes-Benz

Discussing the impacts of AI in the way we develop software

Introduction

I remember the first time I really used a chatbot for work. We were on a call with a colleague trying to debug some Kubernetes issues and it was such a magical moment when the AI model, within seconds of being prompted, pointed out a weird YAML misconfiguration and solved the problem we had been staring at for a couple of hours. It was clear from the start that this was a game-changer. Things were never going to be the same.

We not only realized from the get-go how powerful this technology was, but almost immediately started creating tools with it. It made us feel super-powered and more ambitious by the day and we started building stuff that we would rarely have had time for. And it was such a rush once we felt we were making our team Software Engineering and DevOps more productive with it.

One beautiful example is this tool we built together, a python-based tool that allows us to create pull requests (PRs) to merge changes into multiple projects at once based on YAML-playbooks. The changes are defined for GitHub repositories that match a certain current state (say e.g. all repositories with a Dockerfile with a deprecated image).

These changes are then applied locally, cloning the repositories to a temporary folder (tmpfs), and can be manually approved project for project or auto-approved for all projects. These approved changes are then published and finally reviewed as a PR. This tool alone provided such a productivity boost in some of our projects, specially for DevOps activities in microservice-based applications with myriad divergent repos.

I remember me and the same colleague after we had visited an amazing conference (it was WasmIO’s first edition) and after having strolled through the streets of Barcelona for a couple of hours, we ended up having a great conversation at a hotel bar with the creator of GoBot, a Golang-based robotics framework, and I still get goosebumps remembering our excitement with LLMs back then, tightly co-mingled with a gut-wrenching fear and existential dread.

Once we started on the topic, we couldn’t stop, and already started devising forecasts of what our future was shaping up to be. I think there isn’t a single software engineer or anyone that works in tech that hasn’t tried to take a peek into a (metaphorical) crystal ball and tried to get a sneak peek into the near future to glean how things might turn out. One thing I am certain of is that exciting times are coming and that changes seem to be definitely accelerating, and we cannot sleep on it.

Mercedes-Benz

Things have changed considerably at Mercedes-Benz from the time I joined one of its subsidiaries (Mercedes-Benz Tech Innovation) until now. Not only have we gone through what feels like thousands of migrations of internal services from cloud to self-managed and back to cloud again, but I have experienced the maturing of what we now call agentic programming (the mature version of vibe coding) during the 4 years that I have been working here.

I joined the company just after finishing my PhD in distributed systems and was well aware of many advances in AI but would have never thought that we were so close to coding agents that are as capable as the harnessed AI Models that we have right now.

Not only did Mercedes-Benz take a leap and started integrating AI everywhere it made sense, but the whole industry (not only automobile, but anything that is touched by tech, so pretty much all industries) seems to have jumped on it with an eagerness, a desperation that makes it all feel like a sort of surreal dream.

Mercedes-Benz Tech Innovation is becoming the AI powerhouse for Mercedes-Benz, and we have been developing and releasing AI-based solutions with a staggering tempo.

Mercedes-Benz Direct Chat

The first big AI product we have launched, back in 2023, was Mercedes-Benz Direct Chat. An internal platform that hosts different chatbots, from general ones to chatbots for generating and analyzing images or talking to internal docs. With it, artificial intelligence started becoming an integral part of everyday work at Mercedes-Benz.

It was met with rapid adoption and is a well established tool across the entire Mercedes-Benz organization.

Initially, we only supported OpenAI’s GPT models, but are increasingly opening doors for other model providers such as Google Gemini. This gives employees flexibility to choose the models that better fit their specific needs and tastes.

Mercedes-specific knowledge agents help us navigate through internal systems, making Direct Chat a central knowledge repository within Mercedes-Benz. Internal statistics report that more than 10,000 employees use Direct Chat daily as power users [1].

Responsible use of AI

Mercedes-Benz is quite concerned with not only enhancing our way of work, but doing it responsibly. Internal guidelines and platforms such as our disclosure portal, make sure that code that is deployed is compliant and holds up to our code quality and security standards. This enables us to work in an agile manner and use the full potential of LLM-assisted coding without sacrificing security or quality.

On top of that, we have also mandatory AI training and an additional AI Learning Journey, in which different topics are discussed on how to use this technology in a responsible and secure manner.

Moreover, data protection is indeed a fundamental pillar in our AI strategy. In the case of Direct Chat, for instance, employee inputs are anonymized and no personal data is used for training AI models, and services are hosted in European data centers [1].

Adoption through the organization and GenAI Nexus

I was an early adopter from pretty much all stages, from generating code with OpenAI’s ChatGPT, trying out the first offline LLMs that could run on my older M1 MacBook and have seen an explosive adoption from colleagues (not only developers, but also requirement engineers, product owners, managers and pretty much everyone else that works in the company or in the Mercedes-Benz group).

Even friends that work in other sectors, that were a bit more conservative or reserved with AI adoption, such as banking, are getting fully on board the AI train.

Not only are employees able to interact with AI models through our Direct Chat, but we are also able to programmatically consume different language models via the Mercedes-Benz Model Garden through our GenAI Nexus platform, a centralized entry point for interacting with any AI model throughout the organization. This allows us not only to avoid duplication of efforts, but also allows for better tracking with respects to usage and costs.

Having multiple models with different capabilities at our disposal, makes it easier to optimize our token budget.

From my experience, I would say that most projects tend to over-provision for early stages (PoC, MVP, etc.) and it is always an important step to eventually factor in costs and keep them in check before going into production.

For this, we need to question ourselves, for instance, whether we really need a model with more than 500 billion parameters for doing agentic search in an environment where no hard reasoning is needed and the data is not that complex. The beauty of the model garden, is that it fosters innovation by allowing any employee to easily subscribe to multiple models and to start building a proof of concept without any complicated setup.

FOMO

AI has eaten the world. Is is pretty much everywhere and I think it is sometimes overwhelming to try and stay on top off everything AI-related (more like impossible).

I remember the first months of the AI mania and how I mourned over the loss of the trending repos in GitHub (loss is probably the wrong word, but loss of repos that were not somehow related to LLMs).

Generating code has become so cheap and we see this reflected on the sheer amount of tools out there. I have lost count of how many different models and harnesses I have experimented with. From voice models to coding models, that keep getting better and better as time progresses.

And I am still playing around with many different tools. But the sheer scale of things and the absurd amount of code being generated right now, makes it impossible for you to stay on top of everything. Recently, I was quite positively surprised at the performance of the latest medium-sized open-weight models, when running them on my MacBook M4 Max.

As a result of it, I have been experimenting with using some of them in as a code reviewer (haven’t used them much for generating code as of now). In addition, my latest experiments involve using non-autoregressive decision engines combined with LLMs trying to block prompt injection attempts in some of our in-app AI assistants.

In fact, fear of missing out (FOMO) is quite real in this AI-driven space and the potential for burn-out as dangerous as it ever was. So lately, I am trying to focus more on my mental health and have dialed back on trying to hop on every train and try every new AI-based tool, unless it might be really useful for work.

Social Isolation

Another aspect of the new reality we are living in is that this new world can be also extremely isolating. AI hasn’t necessarily created walls between developers, but it certainly has created incentives for colleagues working on different projects (or maybe on the same project even) to rarely talk to each other, just ask Claude 5, or GPT 5.6 or Kimi K3 or whatever model is at hand. I am quite confident in saying that AI-based workflows have helped us solve issues quicker, but it has also made it harder to circulate information within the team.

Yes, we still get a lot of human contact (virtual or in-person), specially being one of the technology powerhouses for Mercedes-Benz, we are involved in myriads projects and these need to be developed, discussed, refined, maintained and eventually given up (the lifecycle of things in general), and this involves meetings of many different sorts, alignment meetings, sprint reviews and workshops, deep dives, support calls, etc.

In special, hackathons have proved themselves invaluable in getting people together and not only developing great ideas but also helping foster a stronger human connection. Also, as of late, I have loved doing AI-assisted pair debugging sessions with one colleague or another. I feel it is a very organic way of working, as if we’re collaboratively piloting a spaceship.

Humans still hold the keys

So human to human communication and deciding what gets built hasn’t been outsourced yet to LLMs (at least not entirely). Sure, many co-workers, including managers are using LLMs to create project proposals, presentations, documenting stuff, and this has its own side-effects.

One of them is the deluge of data and documents that gets generated or composed each month and it makes nearly impossible to read through everything and keep up to date. Meetings are transcribed, page-long READMEs are generated, Confluence pages are filled with plans and architecture decision records, that in the end are rarely read back-to-back except maybe by AI models.

It is such a wealth of information that sometimes overwhelms more than informs. It is also quite challenging to tame this beast and I think we still don’t have a proper solution for this, but we are working on it.

Context switch and brainstorming

I feel that the ongoing trend is for engineers to try and tackle more projects at the same time. This most definitely gives you a sense of always being busy, but it can be on the other hand extremely exhausting, not only due to decision-fatigue (you have a high rate of decisions per hour) but also due to context switches. I think we need to find a healthy sweet spot between productivity gains and sustainable output and should not pressure ourselves into always doing more and more. Sometimes the best course of action is to slow down, stop and think.

Maybe even go full-on old-school and scribble down your thoughts on a piece of paper. Coming from research, I love thinking about problems and how to solve them and I don’t like AI to introduce too much bias into my thoughts too soon into the design process.

So often enough I will just brainstorm by myself with Obsidian, write down some ideas and only then will I grab an LLM to do some further research and brainstorming. I feel it has helped me feel a deeper sense of ownership for the systems I have built lately and not feel like I am over-delegating my thinking process to any AI model.

Manual coding, a thing of the past, for some

I don’t think there is a single developer in my team (regardless of seniority) that has manually coded much in the past 10 months, as in the olden times of 2024-2025. We are all to some extent assisted by AI models and besides manual human-based code validation and maybe some small (really small) code tweaks, we can fully rely on LLMs for code generation for most tasks. Yes, they still need steering, help with architecture and getting the right context, help building meaningful UX in our systems, help following internal Mercedes-Benz guidelines, help with accessibility issues and readability, sure. Even so, the heavy-lifting of code writing is all theirs. And it is astounding how good the current agents are at it.

I am quite sure though that there are teams in our company that still need or prefer manual coding, in special those more deeply involved with 3D modeling and simulation, areas in which most LLMs still struggle due to their current sub-par capabilities w.r.t. spatial and physical understanding.

They are already quite good at 2D perception, but 3D is still very tricky for them. There are certainly ongoing attempts of improving this by adding e.g. video material to training datasets, but I think there are still a lot to improve upon. To be honest, even 2D spatial understanding is harder than it looks. Have you ever tried building auto-layout functionality for a diagram modeller with an LLM? I have, and it hurt. Let’s just say, the agents need extremely specific guiding and even so, it is no trivial task to accomplish.

One real side-effect of automated code generation is that the volume of code has certainly increased. PRs got much bigger with time, and even though I have initially tried staying on top of it with manual reviews, nowadays reviews for big PRs are mostly AI-driven. We still have shorter PRs, but these are mostly for small changes, configuration or DevOps-related.

Still with help of tools such as SonarQube and Black Duck we are able to make sure quality and security standards are respected. In addition, through AI-assistance we can surface any additional issues these tools might have missed (e.g. maybe some internal conventions weren’t respected).

Test generation is now super cheap and I have seen a very positive increase in using automated tests throughout different projects being developed by our team. It is not that we haven’t written tests before, but they just weren’t as exhaustive and consistently maintained as they are right now.

I just can’t say that I miss struggling with Spring Boot bean configurations and Testcontainers setups that would kind of work to only break with minimal setup changes and sometimes eat up a whole afternoon of trying to fix some test suite. This seems now to be a thing of the past, since current models are already pretty capable in creating and maintaining a proper test suite. I usually define a set of must-test components and corner cases as well as a scope (e.g. do only end-to-end integration tests for backend and only unit test for critical paths) and then let the model run with it.

It has worked really nice so far, and it is definitely a powerful tool for keeping applications stable (no one enjoys their API breaking intermittently between deployments).

Bye bye sunk cost fallacy

It was never easier to create different variants of your ideas. To experiment and create throwaway features or apps even. It has been an exhilarating (and also humbling) experience to be as ambitious as I can be and still be able to achieve working versions of these ideas in mere weeks or even days. Even if some of these ideas do not pan out and are ever used in productive systems, we are still able to gain insights from them.

One good example: 18 months ago, I supervised a bachelor thesis and we have evaluated the performance of open-weight offline models for creating pull request descriptions. It was a really nice project to not only play around with the current capabilities of offline models, but also to work out on how to operationalize the usage of these models in productive settings.

The models needed to be medium to small sized so that they would run on all MacBooks used for development in our team (which have varying capabilities) and they needed to be smart enough to handle each project’s contexts and distill code changes into proper high-quality descriptions (looking just at just an isolated commit will usually not provide enough context for meaningful descriptions).

We got pretty interesting results out of the project, but the PR summarizer got old pretty fast and never saw wide adoption, since soon enough we had powerful harnesses with access to GitHub and that could use whole codebases to create well-described PRs.

And even though the project was not a proper success, I refuse to see it as a failure, since it taught us many things related to operationalizing LLMs.

Challenges for (young) developers

I think the current work environment has become both easier and harder for younger engineers to navigate through. On one hand, it was never so easy as it is right now to start contributing to a project, especially if it is a well-established boring tech stack (LLMs have definitely gotten better at coding in general but they still struggle a bit with novel or more recent technologies, since the training data is obviously underrepresented for those technologies).

It is also quite easy, particularly for less-experienced developers to relinquish the whole thinking process to LLMs. Nonetheless, generating code blindingly is far from proper ownership and I would say there still is a hard need for understanding computer science fundamentals, for understanding how your tech stack works (at least on a basic level) and which moving parts compose your infrastructure (e.g. the basic of GitOps, Kubernetes, the tool you are using for continuous deployment, etc.).

Letting an agent crunch by itself with no supervision, on the other hand, is not only an opportunity lost (you rob yourself the chance of learning how things work), but also dangerous, i.e. you never know when a model is going to hallucinate and decide to wipe out your production database, so you have to stay on top of what it is doing and also never accept all its decisions at face value.

If this is already a hard thing to handle for experienced software engineers, imagine then how challenging it must be for developers fresh out of college.

I had a random button appear in one application once, and when asked, the model told me with a total lack of candor that it was needed for some obscure data validation (it did not make much sense) and when further pressed, it confessed that clicking on the button would wipe out all data from the database. Imagine my reaction.

AI initiatives — human in/on the loop

Our business stream Car Engineering - Future Engineering & Validation in involved in many initiatives that are currently trying to modernize and optimize workflows within car engineering departments throughout Mercedes-Benz. And what better tool to enhance engineers than AI. One of our core tenets when integrating artificial intelligence into working processes within Mercedes-Benz, is that we want to boost human capabilities while keeping the human part relevant. We should automate processes, but also keep humans accountable and fully in control.

Human have to own and supervise these processes and always be able to validate and if needed reverse or stop any triggered actions in order to guarantee that processes run to completion in a correct and safe manner. It is also important to make sure we support engineers and allow them to outsource repetitive dull work while helping them still feel empowered and not feel like AI babysitters.

There is a wealth of domains from within our organization, where AI can make a difference, some of them are: validating compliance against regulations and internal requirements, enriching knowledge graphs with internal Mercedes-Benz knowledge, automating Computer-Aided Design (CAD)-based workflows, making search in different domains (such as finding proper test technology offers within the organization) more efficient.

And we have been so busy, but it is always very rewarding to receive so much positive feedback and excitement from different departments about our initiatives, as we have in the past few months.

Parting words

AI has brought so many changes to the way we think about and produce software and it has not stopped there, it is also revolutionizing other engineering-related areas such as physics and mathematics. As of September 4th Anthropic’s Claude has composed the first ever fully computer-checked proof of Fermat’s Last Theorem in only… 11 days [2]. Wild stuff.

A couple days later (four days to be exact), OpenAI says a fleet of their strongest model at the time (still not public) has solved the Navier-Stokes problem, one of the seven Millennium Prize Problems, with 10,000 AI agents in 88 hours [3]. Mind-blowing.

I think, the tendency is for agents to become better and better at solving many different types of scientific and computational problems at scale and there is certainly no dearth of problems that can still be solved by throwing a humongous amount of thinking compute on it. I hope we will eventually be able to solve issues such as climate change, cure cancer, who knows, one can dream.

It has been a wild journey arriving at this point in history. It is incredible to think that my grandmother (over a 100 years old) has lived through so many pivotal moments in history, and I think the development of thinking machines is one of the most fundamental moments in our history.

How we handle this moment, will define the future of not only our economy, but also all aspects that define our way of living and the world our kids are going to inherit.

So I think it is of utmost importance for us to keep a holistic view of this technology, and do not forget that we live in a world with finite resources, which means we also need to focus on sustainability and treat it as a fundamental component of AI-development and not as an afterthought.

References

  1. Mercedes-Benz Group. Mercedes-Benz Direct Chat. https://group.mercedes-benz.com/technology/digitalisation/artificial-intelligence/mercedes-benz-direct-chat.html, October 2025. Accessed 23 September 2026. ↩
  2. Anthropic. Formalizing Fermat's Last Theorem. https://www.anthropic.com/research/formalizing-fermats-last-theorem, September 2026. Published 4 September 2026. Accessed 23 September 2026. ↩
  3. OpenAI. On the Navier–Stokes Millennium Prize Problem. https://openai.com/index/navier-stokes-solution/, September 2026. Published 8 September 2026. Accessed 23 September 2026. ↩