AI is changing how software is built – and how companies buy it. This category bundles everything on AI-assisted development, agentic AI and the governance questions behind them: from the EU AI Act via spec-driven development to the honest question of what AI acceleration really costs and delivers. Written for CIOs and IT decision-makers who want substance over hype – with numbers, sources and clear limits.
Mistral Large 4 has been in public preview since October 6, 2026; the weights are due at the end of October. Via the API, sovereign use is not possible yet: Mistral may train on your inputs, US corporations sit in its supply chain, and a C5 attestation is missing. Test now; process sensitive data only once the weights are out.
Matthias Radscheit
41 percent of German companies use AI, and a third report significantly higher costs than planned. The difference is a roadmap: I show you four phases, an open budget calculation (16,000 to 35,000 euros for a real pilot), five measurable abort criteria, and funding programs that can cut your costs in half.
Matthias Radscheit
Self-hosting a 70B model currently costs 1,197.30 € net per month plus 599 € setup at Hetzner. The same model class is available from German data centers via API from 0.65 € per million tokens: the break-even sits at around 1.84 billion tokens per month. I show you the calculation — and when your own server pays off anyway.
Matthias Radscheit
In 2026, fine-tuning pays off in three cases: fixed tone and format, extraction at scale, and a small specialist model for latency and unit costs. After OpenAI's wind-down, that practically means LoRA on open-weight models. The training run often costs just the $4 minimum fee — the expensive parts are data preparation and evaluation. For pure company knowledge, you need RAG instead.
Matthias Radscheit
Your company knowledge belongs in a database, not in model weights: RAG looks knowledge up at runtime, stays current, cites its sources, and remains deletable for GDPR purposes. Fine-tuning trains behavior, not knowledge — and OpenAI is even winding down its fine-tuning by January 2027. We show you the pgvector stack, the real costs, and the three limits where RAG fails.
Matthias Radscheit
Workflow automation rarely fails because of the tool. It fails on three questions comparison articles leave out: Where do your data run, who is even allowed to operate n8n, and what does operating it really cost? I show you the license trap in agency hosting, an honest cost calculation with dated prices, and the line beyond which low-code is no longer enough.
Matthias Radscheit
You almost never need your own AI model: training GPT-4 alone cost around 78.4 million USD in compute. What you actually want is delivered by one of four stages: structured prompting, RAG over your data, or LoRA fine-tuning on open-weight models, each of which you can optionally run on EU infrastructure. This guide walks you to the right stage with four guiding questions, backed by sourced numbers instead of gut feeling.
Matthias Radscheit
Your shop search only finds what is in the index: no search engine compensates for missing attributes, empty descriptions, and inconsistent categories. I show you how to enrich facet attributes, synonyms, and short descriptions with an LLM, what that costs (a calculation path instead of a flat rate), and why invented values are worse than empty ones.
Matthias Radscheit
Semantic search finds results by meaning instead of by character string. The embedding fees are cent amounts; the real costs sit in engineering and upkeep. For websites you build it hybrid with Meilisearch, for RAG you use pgvector. And often you don't need it at all: measure your zero-result rate first, then decide.
Matthias Radscheit
On August 2, 2026, the EU AI Act's high-risk obligations take effect. What that means for companies that commission or operate software with AI features – risk classes, obligations, checklist.
Matthias Radscheit
Agentic AI is this year's Gartner vocabulary – but what separates an agent from a chatbot, where does real value emerge for mid-sized companies, and why are 40% of projects canceled? A sober assessment.
Matthias Radscheit
SAP, TYPO3, Salesforce and ageing ERP systems aren't going anywhere. They still have to talk to modern frontends. How Next.js works as an integration layer and Backend for Frontend to decouple legacy step by step, what holds the pattern up, and where it tips over.
Matthias Radscheit
GDPR-compliant AI is possible, but never as a blanket yes. Separate data flow from data origin and every use case sorts into one of four categories. Three pieces of homework decide whether the system stays legally operable.
Matthias Radscheit
Book a no-obligation call, send us an email, or use the form – we'd love to hear from you.