Language model (LLM)
A program that understands and produces language: it writes, summarises, translates, classifies and answers questions in natural language.
AI knowledge · as of June 2026
AI Compass 2026: what actually works now
Evidence instead of hype – so you know where AI creates value today, what becomes mandatory, and how to start sensibly. Every claim with a source. Not legal advice.
AI in plain words
The key terms – made simple
People who don't know what modern AI can do won't use it. Here are the essentials in plain language.
AI agent
AI that doesn't just answer but completes tasks: it plans steps, uses tools (e.g. your systems) and delivers a result – with a human in control.
Knowledge bot (RAG)
An assistant that answers based on your own documents and cites sources – instead of guessing. No model retraining required.
"Reasoning"
Newer models "think" step by step, solving more demanding analysis, maths and planning tasks – at the cost of a bit more time.
Numbers that matter
Where mid-sized companies stand in 2026
Adoption rates vary by survey and AI definition (official ~26 %, industry surveys up to ~41 %). The "95 %" figure is widely cited but methodologically debated – treat it as an indication, not a precise value.
Effectiveness
Where AI works most – with proven effects
The largest, best-documented productivity effects come from controlled studies. Important: these are task-level effects – business value only emerges when you redesign the process.
Customer service
more issues resolved per hour – for newcomers around +34 %. Field study with 5,172 support agents.
Brynjolfsson et al., QJE 2025 ↗Knowledge work & writing
on typical writing tasks, with +18 % quality. Controlled experiment with 453 professionals.
Noy & Zhang, Science 2023 ↗Professional & advisory tasks
and over 40 % higher quality – within AI's strengths. Field experiment with 758 consultants.
BCG/Harvard 2023 ↗Software development
more tasks completed per week; largest gains for junior developers. Three field studies, ~4,800 people.
Cui, Demirer et al. 2024 ↗Benefits are greatest for entry-level and routine tasks and for less experienced staff. Outside AI's strengths, plausible-sounding errors loom ("jagged frontier"). What matters: choosing the right tasks, human review – and redesigning the process, not the tool alone.
Law & obligations
EU AI Act & KI-MIG: the essentials
KI-MIG is still in the legislative process; details may change. Not legal advice.
- Since 02 Feb 2025 the AI literacy obligation applies (Art. 4 EU AI Act) – even for companies that only use ready-made tools like ChatGPT.
- From 02 Aug 2026 transparency obligations apply: chatbots must be recognisable as AI, AI content must be labelled.
- Germany implements via KI-MIG (AI Market Surveillance and Innovation Promotion Act); the Federal Network Agency becomes the central authority.
Society & work
Why AI matters now
Demographics as a driver
Germany's working-age population shrinks by 3.2–4.9 million by the mid-2030s. AI and automation become a plausible answer to the skills shortage.
Destatis 2025 ↗Tasks, not jobs
AI mainly changes tasks. In Germany ~1.6 million roles are in transition (created and reduced); total employment stays broadly stable per projection.
IAB 2025 ↗Reskilling is the bottleneck
Globally, a net +78 million jobs could emerge by 2030 – but 59 % of workers will need new skills by then.
WEF Future of Jobs 2025 ↗Honestly named
Risks you should know
AI can be wrong
Language models sometimes "hallucinate" convincingly. Knowledge bots with sources and human review reduce the risk significantly – but do not eliminate it.
Stanford RegLab 2024 ↗Shadow AI
About half of employees use unsanctioned AI tools – a real data-protection and compliance risk that needs clear rules and secure tools.
Cisco / CIO 2025 ↗Prompt injection
The top security risk for AI applications – especially for agents with access to systems and data. Guardrails and oversight are mandatory.
OWASP Top 10 (LLM) 2025 ↗How to start
From not knowing to real impact – a proven path
Perception gap: 67 % of leaders, but only 40 % of employees even know what AI agents are. People who don't know AI won't adopt it – so awareness and enablement come first.
Microsoft Work Trend Index 2025 ↗
- 01
Understand
Awareness at leadership and staff level: what can (agentic) AI do, and what not?
- 02
Enable
Build broad AI literacy – also a duty under Art. 4 of the EU AI Act.
- 03
Focus
Choose use cases from the business problem, not "AI because everyone does it".
- 04
Pilot
Build one use case – and redesign the process instead of bolting AI on top.
- 05
Scale
Into productive operation with governance, clear roles and trust.
Transparency
Sources & methodology
Research as of June 2026. We deliberately distinguish model estimates (scenarios) from measured studies (experiments); adoption figures vary by survey. This page is factual orientation, not legal advice.
Make AI work in your company?
In a free initial check we assess your starting point and show the most sensible first step – honestly and without sales pressure.