EU AI Act going live
Europe's AI rulebook is moving from theory to practice and it's already forcing firms to rethink product design rather than legal memos. Organisations that deploy partly autonomous or “agentic” AI will soon face detailed governance, logging and documentation duties that are harder to map to real products than summaries suggest. That matters because buyers in regulated markets will expect audit trails and clear human oversight as a baseline, not an afterthought. (artificialintelligence-news.com)
Europe’s artificial intelligence law is not arriving in one dramatic launch. It has been switching on in stages since August 1, 2024, with the biggest block of rules for many companies due on August 2, 2026. (ec.europa.eu, eur-lex.europa.eu) The first hard edge already landed on February 2, 2025. That date turned on the ban for certain uses, including social scoring, untargeted scraping of faces to build recognition databases, and emotion-reading systems in workplaces or schools except for medical or safety reasons. (ec.europa.eu, ec.europa.eu, ec.europa.eu) The next stage hit on August 2, 2025, and it pulled in the companies behind general-purpose models. The European Commission says those providers now need to meet model-level duties, and the most advanced models that pose “systemic risk” must notify the European Union’s new Artificial Intelligence Office. (ec.europa.eu) That matters because most businesses do not build a foundation model from scratch. They buy one, fine-tune it, connect it to company data, and wrap it in a tool that can draft emails, approve claims, rank job applicants, or act on a worker’s instructions. (ec.europa.eu, europarl.europa.eu) The law splits those tools by risk, not by hype. A chatbot that answers routine questions is treated differently from software used in hiring, education, critical infrastructure, or access to essential services, where mistakes can change a person’s job, loan, school place, or medical pathway. (eur-lex.europa.eu, europarl.europa.eu) August 2, 2026 is the date many product teams are now circling. The Commission’s implementation page says that is when most of the Act starts to bite, including rules for high-risk systems in Annex Three, transparency duties in Article 50, and national and European enforcement. (ec.europa.eu) This is why “agentic” software is awkward to map onto the law. If a system can plan steps, call tools, retrieve data, and trigger actions across a workflow, a company has to show who is supervising it, what logs exist, what instructions it followed, and where a human can step in before harm spreads. (eur-lex.europa.eu, europarl.europa.eu) The practical problem is that the rulebook is ahead of the plumbing. The European Parliament’s research service said in June 2025 that harmonised standards were slipping into 2026 even though they are crucial before the high-risk obligations take effect in August 2026. (europarl.europa.eu) So companies are redesigning products now, not waiting for lawyers to write a memo later. Logging, documentation, model cards, human review checkpoints, and clear handoffs between a model provider and the business deploying the tool are becoming product features in their own right. (ec.europa.eu, ec.europa.eu) The fines explain why nobody wants to improvise. The Act allows penalties up to 35 million euros or 7 percent of worldwide annual turnover for prohibited practices, and up to 15 million euros or 3 percent for many other breaches of operator obligations. (artificialintelligenceact.eu, ai-act-law.eu) The end result is simple even if the legal text is not. In Europe, an artificial intelligence tool that cannot show its work, record its actions, and leave a human clearly in charge is becoming much harder to sell into banks, hospitals, insurers, schools, and government buyers. (ec.europa.eu, eur-lex.europa.eu)
