Follow only the keynotes and you would think GenAI can already do everything. Read only the incidents and you would think it is best avoided. The reality of 2026 is more interesting — and asks for a level head.
What genuinely works
- Agents that keep going. AI now completes tasks of several hours on its own — research, administration, software changes. Reliability on such long tasks sits around 50%: impressive, and at the same time the reason a human must stay accountable.
- Proven workhorses. Coding assistance and customer service are the most mature applications. Klarna let its AI assistant do the work of hundreds of agents — and then hired people back for the complex and emotional conversations. Both halves of that story are true.
- Collapsing costs. The AI performance that cost twenty dollars per million tokens in late 2022 now costs cents. Mind you: cheap per token is not the same as a predictable bill — 73% of companies saw AI costs exceed budget (FinOps Foundation, 2026).
For perspective: in 2025 about 33% of Dutch companies used AI (Eurostat) — a European front-runner, but among small firms it is 17%. Start in a structured way now and you are not behind.
What genuinely goes wrong
- Fabricated output is here to stay. Hallucination is not a teething problem but a property. The global tally of court cases in which AI fabrications surfaced in filings stands at roughly 1,400. And since the Air Canada ruling, the principle is clear: your chatbot, your liability.
- Quiet data leaks. Over a third of what employees paste into AI tools is sensitive information (Cyberhaven, 2025), often through personal accounts. Banning shadow AI does not work; clear agreements and business licences do.
- Regulation with dates. The EU AI Act already applies: prohibited practices and AI literacy since February 2025, transparency duties for chatbots from 2 August 2026. The heavier obligations shift to late 2027 — postponed, not cancelled.
Our advice
Treat GenAI like a new hire: plenty of talent, no sense of responsibility. Give it real work, check the output, and record who is accountable. Between hype and fear lies plain good management.
Tags
- genai
- strategy
- risk
Related
- Education
GenAI in the lecture hall: rules students actually use
Banning fails, and so does a free-for-all. Our research at Leiden University shows what GenAI guidelines look like when students and teachers both accept them.
2 min read
- Advisory
Forecasting as a routine: technology forecasting in practice
Which technology deserves your investment — and when? Our action research combines text mining and scenario planning into an approach that is fast, deep, and repeatable.
1 min read
- Advisory
Beyond AI: the SME technology agenda for 2026
While everyone watches AI, 2026 is the year compliance sets the digital agenda for SMEs — with four deadlines that have an actual date.
2 min read