Articles

What drives us.

Our own pieces on AI engineering, compliance and digital visibility - and outside recommendations we think are worth reading. Each with a short commentary: what it says, what caught our attention and what we take from it.

Article · 6 min

The reality-check sprint: from idea to decision in a few weeks

A workshop, a clickable prototype or a first measured result - and an honest recommendation on what is worth it. This is how we get started.

Article · 4 min

Why good AI products start with a rejected idea

The first idea is rarely the best - with AI even less so. Why we deliberately drop variants early and what that has to do with good product design.

Recommended5 studiesEvidence review · agent-sii.com · 2026„Evidence review: AI Act and TISAX - what applies when“Deadlines, fines and audit catalogues from the primary sources of the European Commission and the ENX Association, as of November 2026.→ Deadlines are not a matter of opinion - planning starts from the original sources.Our take →
Article · 5 min

How this website thinks along - and what it deliberately doesn't do

Questions instead of forms, a magazine instead of one page for all, a picture of your visit that stays in your browser: how our website uses AI to respond to you and why it shows you what it assumes about you.

Article · 8 min

Time and materials, fixed price or sprints? How we bill software projects

Complete requirements upfront are rare in software development. Which billing models exist, when each one fits - and where the savings from AI end up.

Recommended8 studiesEvidence review · agent-sii.com · 2026„Evidence review: how far along are companies with AI?“20%, 26% or 54.5%? Official statistics and institutes measure AI use very differently - and agree on what holds companies back.→ The range is not a contradiction but a matter of definition - and the bottlenecks are the same everywhere.Our take →
Article · 5 min

Public AI without a cost trap: how we limit abuse

Every question to an AI costs money. How our website stays fair to use - with session tokens, open limits and a separate budget per feature - and what of this fits your projects.

Article · 7 min

30 minutes stay 30 minutes: how much developer time AI can actually reach

Meetings are blocked time, and the rest is fragmented. What studies say about which part of the week an AI lever can act on at all.

Recommended4 studiesEvidence review · agent-sii.com · 2026„Evidence review: how much does AI really save in coding?“From 55.8% on one task to around 5 to 14% in the project: what experiments and longitudinal studies show about AI in coding.→ A single percentage is no basis for planning - what matters is what was measured.Our take →
Article · 8 min

Where this website runs - and what happens to your data

Hosting in Frankfurt, no cookies, AI models only in EU data centres - no retention at the provider, no training, no sharing. Our privacy concept in plain words.

Article · 4 min

Agents that write tests - how we approach it

Coding agents can dramatically speed up test automation. How this works even in TISAX environments - with human approval for every step.

Recommended9 studiesEvidence review · agent-sii.com · 2026„Evidence review: AI in everyday development - usage, trust, code quality“84% use AI, 46% distrust it: what surveys, repository analyses and measured test generators show about AI in daily work.→ Usage is not the bottleneck - review is.Our take →
Article · 4 min

AI adoption: we make your teams capable, not dependent

How AI really takes hold in an organisation: with internal champions, a central core that stays current, and effects that are measured rather than felt.

Article · 11 min

Agentic AI: 85% on the slide, 36% at best

Headline figures describe single tasks. What reaches the project is the product of three factors - typically single digits according to the studies, about a third at best.

Recommended6 studiesEvidence review · agent-sii.com · 2026„Evidence review: staying visible when AI answers the search“Fewer clicks, different sources, measurable levers: what independent experiments and studies show about visibility in AI search.→ What is not cited is not seen - and AI search mainly cites third parties.Our take →
Article · 4 min

Usage statistics without a cookie banner: the question first, then the data

Data protection forbids less than many people think. If you first decide which decision you want to make and only measure for that, simple audience measurement needs neither cookies nor consent.

Article · 4 min

The product may be complex. The experience may not.

Customers think in needs, offerings think in variants. Why advice is becoming a conversation - and what streaming, retail and banking have set as the benchmark.

Recommended3 studiesEvidence review · agent-sii.com · 2026„Evidence review: how far do open models lag behind the frontier?“Independent benchmarks by Epoch AI: four months on average, six to seven for cyber capabilities, six to twelve on a single graphics card.→ For AI on your own infrastructure what counts is the gap to the frontier - and whether the model can be swapped.Our take →
Article · 5 min

From the artefact to the tool: controlled autonomy in software development

Traditional, AI-assisted, agentic - the ways of working coexist. What shifts: people work less on the code and more on the system that produces it.

Article · 4 min

From shadow AI to official approval: a roadmap

Generative AI is banned but used anyway? Five steps to an official assistant that data protection, information security and the works council sign off on.

Recommendedgotocph.comTalk · GOTO Copenhagen · Adam Tornhill · 2026„AI-Friendly Code: The Missing Link for Speed with Quality“Agents work far more reliably on healthy code than on grown legacy. Code quality decides whether AI speed lasts or fades after a few months.→ The bottleneck isn’t the model - it’s the code it meets.Our take →
Article · 4 min

What we learned from six GenAI projects

Specifications, price lists, user comments, construction documents - six lessons from projects that made it past the demo.

Article · 4 min

On-premises AI: three to ten months behind the cloud

If you may only run AI in your own data centre, you work with models that trail the frontier. How big the gap is - and what follows from it.

Recommendedveracode.comReport · Veracode · 2026„2026 GenAI Code Security Report“Around 44% of tested code-generation tasks contained a vulnerability - barely any progress on the previous year despite better models.→ Security checks belong in every AI pipeline, not at the end.Our take →
Article · 4 min

Start. Prove. Scale.

No big project to begin with - one real problem is enough. How we get from workshop to prototype in a few weeks, and what happens next.

Article · 5 min

What the EU AI Act means for your engineering

Less panic, more plan. Five things engineering teams should know and do now.

Recommendedarxiv.orgStudy · arXiv · 2026„Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild“Large-scale study of around 300,000 verified AI-authored commits in more than 6,000 repositories: a relevant share of AI-introduced issues stays in the code.→ Review and refactoring must scale with AI use.Our take →
Article · 6 min

Agentic development in everyday enterprise work

Coding agents are no longer a demo. What works in large organisations - and where it fails.

Article · 4 min

Websites for agents: MCP is the new SEO

When customers delegate research to AI assistants, an interface decides who gets seen.

Recommendedai-act-service-desk.ec.europa.euStandard · Europäische Kommission · 2026„Timeline for the implementation of the EU AI Act“The European Commission’s official timeline: which AI Act obligations apply from when.→ The basis of any planning - orientation, not legal advice.Our take →
Article · 4 min

AI governance without bureaucracy

Good governance speeds things up. Four principles for rules that teams actually follow.

Article · 5 min

TISAX meets GenAI

Using generative AI in the automotive supply chain without risking the label - a practical checklist.

Recommendeddevelopers.googleblog.comStandard · Google · 2025„Introducing A2UI: An open project for agent-driven interfaces“An open protocol that lets AI agents return interfaces instead of text - safely, because the client only renders known components.→ The foundation of the AI assistant on this website.Our take →
Recommendedresearch.googleReport · DORA / Google Cloud · 2025„2025 State of AI-assisted Software Development“The large annual DORA study shows that AI amplifies what is already there - good teams get better, weak processes become more visible.→ Confirms our approach: the engineering frame first, then autonomy.Our take →
Recommendedmetr.orgStudy · METR · 2025„Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity“In a randomised field study, experienced open-source developers were slower with AI tools - while believing they were faster.→ Measure effects instead of estimating them - which is what we do in our projects.Our take →
Recommendedmodelcontextprotocol.ioStandard · Model Context Protocol · 2025„Model Context Protocol - Specification“The open standard through which AI applications access data and tools - from IDEs to company websites.→ If you want to be visible to agents, speak MCP.Our take →
Recommendedanthropic.comGuide · Anthropic · 2024„Building effective agents“Practical patterns for agent systems - and the case for starting with the simplest setup that works.→ How we build agents: simple, verifiable, with clear tools.Our take →
Recommendedcyber.gouv.frGuide · BSI & ANSSI · 2024„AI Coding Assistants“Joint recommendations by the German and French cybersecurity agencies on using AI coding assistants securely.→ The benchmark for AI development in regulated environments.Our take →
Part of the SII Germany team

Let’s
talk.