Observations.
What is becoming visible right now — across the market, in research and in product decisions.
2026
-
More agents, same results — Stanford plus an honest scope note
A single AI agent with the same compute budget delivers at least as good results as a whole team of agents. Often better. A Stanford group demonstrated this in a controlled study in April.
-
Commerzbank puts AI KPIs on the table
Oliver Dörler, Chief Data and AI Officer at Commerzbank, committed to specific numbers and KPIs for AI in the Börsen-Zeitung at the end of April — a level of clarity I rarely hear from large German corporations. By 2028, the bank plans to invest 140 million euros in AI, generate a value contribution of 300 million…
-
Knowledge Flywheel: The memory makes the difference
In two years, many professional services firms will likely be running the same AI assistant. On the same frontier models, with comparable capabilities.
-
IAB Establishment Panel: Adoption Without Guardrails
One in four German businesses now uses generative AI. Sounds like a success story. Until you read the second set of numbers in the same study.
-
GDPR is the sharper blade for everyday AI use
The sharpest AI regulation affecting everyday operations at most companies isn't in the AI Act. It sits in GDPR and has been in force for eight years.
-
AI Architecture: Switching Vendors Has to Be Possible
The best AI model is not worth much if you cannot switch away from it.
-
Shadow AI and the training gap: read the order of the numbers
Bitkom surveyed 1,003 working adults in Germany on AI at work. 48% use it. 21% have ever received employer-provided training on it. 12% use it covertly.
-
One Model Is a Risk Bearer, Not a Control Mechanism
Virtually every complex task I've given to a single AI model so far has contained flaws, substantial ones in the vast majority of cases. I know this because I have the output checked by a second model each time.
-
Agents prefer guessing to asking — ProactiveBench + frontier differentiation
Your AI agent would rather deliver a wrong answer than admit it's missing data. For most models in the wild, that's still true — even as the picture at the top is starting to shift.
-
AI Readiness Self-Check, open instead of gated
Many "AI readiness tests" on the web are only half self-assessment. I've taken a closer look at a few of them over the past days.
-
Aleph Alpha + Cohere — the real signal isn't the merger
Europe has built plenty of AI models. What's been missing: a provider that doesn't just get regulated — but actually sells.
-
Verification Gap — why AI structurally cannot replace strategy
AI can write code at senior level. But it can't develop strategy. That's not a coincidence — and based on the current state of technology, it's not a problem that can simply be solved with more compute.
-
Profitless Prosperity — 81% with no measurable AI success
81% of companies deploying AI see no measurable business impact. That's not a skeptic talking. That's McKinsey, Roland Berger, MIT, Bain, and Deloitte — all in the last few weeks, independently.
-
The Mapping Problem — it's not the AI, it's where you look
Same AI, same tools, same APIs. One makes almost twice the revenue. The difference has nothing to do with technology.
-
Copilot cranked up — suddenly everyone sees everything
40% of IT leaders have delayed their Copilot rollout by three months or more, according to Gartner. Not because Copilot failed — but because it did exactly what it was built to do: find information.
-
Prompt Injection: Model Selection is Security Decision
Someone writes a single sentence in a document. The AI system evaluating this document then changes its judgment by 20 percentage points.
-
Klarna — Automate. Fail. Adapt.
"We focused too much on efficiency and cost. Quality suffered." — Klarna CEO Siemiatkowski. What sounds like AI failure is the opposite.
-
Secret Cyborgs — Your Business Model Punishes Efficiency
Your employees are already using AI. You don't know it — and your business model is the reason why nobody tells you about it.
-
Next-Gen Models Ready — Vendor Evaluations Built on Sand
Anyone evaluating AI vendors right now is making decisions based on models that won't exist in a few months.
-
Xiaomi Hunter Alpha — is Europe paying eight times more for AI?
Is your company paying eight times more for AI — without knowing it?
-
LLM Homogenisation — Why Your LLM Makes You Average
ChatGPT makes you average. Not because it's bad — but because it does exactly what it was built to do.
-
Should the CFO Lead AI — Not the CTO?
A recent HBR study compared over 1,000 companies. The result is remarkably clear: AI initiatives under finance leadership deliver measurable business value in 76% of cases. Under IT leadership, it's 53%.
-
AI Agents in Production — The First Documented Incidents Reveal a Systemic Pattern
AI agents in the enterprise are no longer pilot projects. They make decisions, modify data, access systems — with the same permissions as the person who launched them. No confirmation. No limits.
-
3 Warning Signs Your Company Isn't AI-Ready
"We need to do something with AI."
-
AI Agents: Brilliant at Clear Tasks, But Not for DIY Builders
A new employee asks when something is unclear. An AI agent doesn't. It works with what it gets — and when in doubt, delivers convincingly sounding wrong results.
-
Your Data Isn't AI-Ready — And That's Your Biggest Lever
McKinsey reports: 88% of companies use AI. At the same time, Gartner says: 57% of organisations have data that isn't AI-ready. That's not a contradiction — that's the norm.
-
The Self-Driving Question — When Do You Delegate to AI?
Self-driving cars have fewer accidents than humans. Yet we still sit behind the wheel.
-
Hallucinations: Measurably Improving, But You Need to Know
Knowledge workers spend an average of 4.3 hours per week verifying AI answers for accuracy. Not because they're paranoid — but because hallucinated answers sound just as convincing as correct ones. Often even more so: an MIT study shows that AI uses more confident language in wrong answers than in correct ones.
-
Where AI Already Works Today — the Spec+DoD Formula
If you want to understand where AI works and where it doesn't, two questions are essential: Can you precisely describe what you want beforehand? And can you verify whether the result is correct afterwards?
-
The Software Revolution Nobody Saw Coming
I come from software development and built a software company myself. What is happening in this industry right now, I have not seen in 30 years.
-
AI in 2026: Why This Year Is Decisive
In conversations with executives, I'm seeing two reactions to AI right now: anxiety, because they don't know where to start. Or complacency, because ChatGPT seemed so easy.