AI changes our work. Not our responsibility.
Where these nine principles come from
We use AI and agents every day in real projects. Along the way we have learned where they fundamentally change work, and where human experience, control and responsibility become more important.
These nine principles grew out of that.
For us, AI first means: check first, then decide. People make the decision, and they stand behind it.
1. AI agents change how we develop software
We no longer discuss whether agents can be used at all, but where they make sense.
Our own website with more than 440 pages has been developed mostly with agents since 2026: a person describes the task, the agent implements it, a person checks the result.
For you: We bring this way of working into your projects and show transparently where agents deliver a real advantage, and where they do not.
2. For us, AI first means: check first, then decide
For every task we first ask whether an agent can do it better, faster or more thoroughly. The answer can be no.
Human in the loop therefore means more to us than correcting afterwards. People build the agents, give them context and rules, and review their work.
In our development that happens at several points: through automated tests, code review, staging and your sign-off.
Certain things we do not leave to agents alone: communicating with you, deploying to production, deleting data, issuing invoices, or publishing under our name.
For you: You can trace what was automated, which rules applied and who checked the result.
3. Our value lies increasingly in building systems in which agents can work reliably
Our team's work is shifting.
Less time goes into simply working through individual tasks. More time goes into architecture, context, rules, tests, safety boundaries, data protection and continuous improvement.
Our quality gates are one example: before every merge, 39 automated checks run. Agents benefit from this system. It is created and developed by the team.
For you: You get AI tools and, with them, people who know their limits from daily work and shape the environment accordingly.
4. Productivity gains do not belong in an AI surcharge
We continue to bill by effort and at the same hourly rates as before the broad use of agents.
When agents make a task faster, that advantage reaches you: either through less effort or through more depth of content within the same budget.
Model costs are part of our hourly rate. There is no separate line for them on the invoice.
For you: You pay for our work, experience and responsibility, not for a language model running somewhere.
5. Instructions are everywhere, experience only comes from applying them
Knowledge about AI, software development and automation is available almost everywhere today.
That is why we share our experience openly: methods, workflows, numbers, technical approaches, and also things we tried and discarded again.
You are expressly welcome to use this knowledge without us.
Our business is not access to information. Our business is applying it with expertise and responsibility.
For you: You can judge how we work before you hire us, and during a project you can follow why we make certain decisions.
6. The faster AI tools change, the more a stable foundation matters
For content, users, permissions, workflows and structured data we have relied on Drupal since 2012.
Around it we choose tools by task. Agents, models and automation platforms may change.
The foundation should remain.
For you: The core of your platform stays in place regardless of which AI tool currently makes sense.
7. Open source is our means against unnecessary dependency
We want to keep control over code, data and the moment we swap out a technology.
That is why a large part of our stack is open-source software: among others Drupal, n8n, LiteLLM, Ollama, GitLab, Redmine and Linux.
We also give something back ourselves, for example through our own Drupal modules, event sponsoring and coworking days.
For language models we decide pragmatically by quality, data protection and task. That is why we use both open models and services from Anthropic and OpenAI, directly or through providers such as Langdock.
We pseudonymise personal data on our own infrastructure before it is handed to external models.
For you: You can trace which model processes which data, and the technical architecture stays replaceable.
8. We do not use AI to replace people with machines
Our work has not disappeared. It has changed.
Those who used to mainly implement now also build and run agents, define rules, check results and develop the way we work.
We train the whole team for that.
We do not claim to be able to predict the long-term effects of AI on work. But our path today is clear: we develop this way of working with our team, not against it.
For you: We develop AI together with the people who already know the processes, the technology and the subject matter, instead of bypassing them.
9. We test ourselves before we recommend anything
Research and prototyping have always been part of our normal work.
We build automated workflows, run them in production, watch their results, change them, and also abolish them when they do not prove themselves.
Fully automated processes are often technically possible. That does not automatically make them sensible.
For you: You get recommendations from practical experience, including the cases where less automation is the better decision.
What this means for working with us
Anyone who hires arocom works with people who build agents and answer for them. In every project you see where agents worked and who checked. The invoice follows the effort, as before. The goal is growth, not a cheaper purchase of the same service: to take on new topics and to introduce AI with us in areas you do not have yet.
What we learned along the way is published openly in the blog, with numbers and with what did not work.
Talk to a human
The first conversation is free and without obligation. Axel Roth leads it, not an agent.
Does arocom replace employees with AI?
No. We currently do not use AI to replace people with machines. The work has shifted: those who used to implement now also build and run agents, define rules and check results. We train the whole team for that.
What am I paying for if agents do the work?
For work, experience and responsibility: describing the task, equipping agents with context and rules, checking the result, standing behind it. Our hourly rates are the same as before the broad use of agents. Where we bill by effort, the invoice falls with the effort. On a fixed price the benefit arrives as more depth of content for the same budget. Both models are described on the investment page.
Which AI vendors does arocom use, and where is my data?
Open models as well as services from Anthropic and OpenAI, directly or through providers such as Langdock, which is hosted in the EU. We pseudonymise personal data on our own infrastructure before it is handed to an external model. Every vendor is listed in the data processing agreement you sign with us (Art. 28 GDPR). The team is trained under Art. 4 of the EU AI Act, and we label AI-generated content under Art. 50.
Does arocom only work with Drupal?
As a CMS, yes, since 2012: for content, users, permissions, workflows and structured data. Around it we choose tools by task, a large part of them open source: n8n, LiteLLM, Ollama, GitLab, Redmine, Linux.
What does "human in the loop" mean in my project, concretely?
An agent works on its own branch. Between that and going live are automated tests, a code review by a developer, staging and your sign-off. A human always communicates with you. You can trace what was automated, which rules applied and who checked the result.
Who pays for model usage?
We do. It is part of the hourly rate and does not appear as a separate line on the invoice.
Axel Roth, Ostfildern, September 2026.
I wrote these nine principles with the team and I stand behind them. When our work changes, I change this page and note here what is different.