// AI in software development
Using AI in software development the right way.
Most development teams bought the licences a while ago. Hardly anyone knows whether Claude Code, Copilot or Cursor actually saves them time. I help teams in Germany and the EU find out, and set things up so data protection, the works council and liability are settled before the rollout, not after.
Daily
Claude Code in client projects with Shopware and Laravel, for over a year and a half
~20%
time saved in my own work. Not 10x, and I'll tell you how uncertain that number is
5 years
of QA and test automation (2015–2020) before I built software myself
EN / DE
consulting, training and documentation in English or German
Why gut feeling misleads you here
In METR's 2025 randomised study, experienced developers took 19 percent longer with AI. Afterwards they believed they had been 20 percent faster. The 2026 update looks friendlier, but the confidence intervals still include zero.
DORA 2025 describes AI as an amplifier of whatever a team already has. Rob Bowley went through the seven team types in the report: only two get more throughput without breaking production more often.
For me it is around 20 percent, mostly in unfamiliar code, and I give some of it back in review. In my experience the model rarely decides whether a coding agent pays off. Review rules, tests and the choice of tasks you hand it do.
How it works
- 1
Assessment
1–2 daysI look at your repository, pipeline and current AI use and talk to two or three people on the team. We pick typical tasks and estimate up front how long they take without an agent. That is your baseline. You get a written assessment of where AI pays off for you and where it does not.
- 2
Team rollout
alongside your work, a few weeksProject configuration for the agent (CLAUDE.md, rules, your own skills instead of marketplace plugins), review rules for AI diffs, and tests as a hard gate before merge. Plus a brief for your data protection officer and works council, and a training session you can file as AI-literacy evidence under Art. 4 of the EU AI Act.
- 3
Measure and adjust
after 4–6 weeksWe time the same tasks again, with a clock instead of a feeling. Whatever does not help gets dropped. If you want, I stay on as a reviewer for AI-generated changes.
What matters in Germany
The technology is usually the easy part. In German companies, rollouts fail more often on questions nobody asked in time.
GDPR and trade secrets
Every prompt sends code to a provider, usually in the US. The classes themselves are rarely the problem; what travels with them is: customer data in fixtures, logs, .env files. I work out which plans do not train on your data, where you need a data processing agreement, and which files the agent must never see.
Works council
Coding assistants log who did what and when. Under German law (§ 87(1) no. 6 BetrVG) that almost always makes them subject to works council co-determination. Find out after the rollout and you start over. I prepare the technical facts so a works agreement can build on them.
EU AI Act
Since February 2025, companies deploying AI systems must ensure sufficient AI literacy among their staff (Art. 4). For coding assistants that mostly means: train people and document it. The high-risk obligations normally do not apply here.
Liability
Code from the agent is your code. The new EU Product Liability Directive, due to be transposed by December 2026, explicitly treats software as a product. Review and tests become a liability question too.
More on product liability →I'm a developer, not a lawyer. I provide the technical groundwork. The legal assessment is up to your data protection officer and legal counsel.
Who this is for
Small and mid-sized development teams at agencies, mid-market companies and SaaS businesses. Especially if you work with PHP, Shopware or Laravel, because I know the code the agent writes there from my own work.
What you won't get from me
Slides about AI transformation, or a promise of 10x. I use Claude Code every day, but I recommend what fits your stack and your data-protection constraints. And if the assessment shows a coding agent doesn't pay off for your team right now, that is exactly what the report will say.
FAQ
Can our source code go to Anthropic, OpenAI or GitHub at all?
It depends on the plan and on what is in the code. Business and enterprise plans from the major providers usually rule out training on your data by contract and offer a data processing agreement. What matters is what lives in the repository: credentials, personal test data and customer data should be removed first or blocked from the agent. Your data protection officer signs off; I give them the technical facts.
Does the works council have to agree?
If you have one: very likely yes. Under established German Federal Labour Court case law, it is enough that a system is objectively capable of monitoring behaviour or performance. Coding assistants produce usage data that can be tied to individuals. Involve the works council early; it saves months.
What does the EU AI Act require from us?
For coding assistants, mainly Art. 4: since 2 February 2025, companies deploying AI systems must ensure sufficient AI literacy among the staff who use them. In practice that means training and documentation. High-risk obligations normally do not apply to software development tools.
Who is liable for code an AI wrote?
You are, just as for code your team writes by hand. The new EU Product Liability Directive must be transposed into national law by December 2026 and explicitly treats software as a product. A traceable review and test process for AI-generated changes is therefore risk management too.
Which tools do you recommend?
Whatever fits your stack, your data-protection constraints and your budget. I use Claude Code daily and know it best. Where I'd need to get up to speed on another tool first, I'll tell you.
What does it cost?
The assessment is offered at a fixed price. I quote the rollout after the intro call, once I know your team size and starting point.
Do you work remotely or on site?
Remotely across Germany and the EU. On-site sessions, for example workshops or training, by arrangement.
Talk first?
Tell me how big your team is, which tools you use and what you hope to get out of them. You'll get a first assessment before anyone talks about prices.