On top of that, the debate on European AI sovereignty landed back at the top of the agenda, Flemish minister Demir announced a new AI roadmap for education, and the European privacy watchdog warned about shadow AI within organizations. An edition that puts strategic choices on the table.

Every two weeks on the Xylos blog, we bring you a sharp and honest overview of what's really moving in the world of generative AI, with the context you need. The foundation remains the biweekly LinkedIn overview by Tom Van 't veld, Learning Innovator at OASE (powered by Xylos). Tom follows AI developments closely, and we translate his observations into what they concretely mean for organizations and the people working in them. Welcome to edition six.

Below are five stories from the past two weeks that matter most for your organization.

On June 9, Anthropic launched Claude Fable 5, the first model in the Mythos class. Three days later, Anthropic switched off both models worldwide. The US Department of Commerce had issued a directive to deny foreign nationals access, citing export rules and national security. In a shared cloud, real-time control over nationality is unworkable. Switching it off worldwide turned out to be the only practical option.

The trigger turned out to be trivial. Amazon CEO Andy Jassy had mentioned to the US government, in conversation, that his researchers had pulled sensitive information out of Fable. That information could allegedly be used for cyberattacks. According to one researcher involved, the feared 'jailbreak' came down to having the model read and improve a piece of code — something that works just as well with other models, GPT-5.5 included. Security researchers called the ban 'dangerous' and the model 'no unique risk.'

For Europe, the debate on technological sovereignty immediately shot back to the top of the agenda. France openly advocated for Mistral. The European Commission is looking into the practical consequences. Anthropic is adjusting its privacy policy so that US citizens regain access via an identity check (starting July 8). For business users outside the US, that offers little relief.

For Belgian organizations, there's a principle here that weighs heavier than the specific model. A foreign government can shut down a commercial AI model for hundreds of millions of users with a single directive. That affects the continuity of any organization that builds critical workflows on a single AI vendor. The right reflex is to actively avoid vendor lock-in. Build your AI stack flexible enough to switch to a different model when needed. Keep critical processes built around a well-considered mix of vendors, with a clear migration path in case one vendor falls away.

On June 16, Microsoft made Copilot Cowork generally available. Cowork is the agentic layer of Copilot. It carries out complex, long-running tasks independently instead of just offering a suggestion. A new button appeared in the Microsoft 365 Copilot app to switch from chat to Cowork mode. Anyone without a license yet sees a 'Request Access' button.

An important detail for European organizations: Cowork currently runs on Anthropic models. That data is processed outside the EU. GPT models will be added soon (already testable in the Frontier program), and those would run within the EU. Microsoft is also announcing its own, cheaper model, 'Cowork 1,' for standard tasks.

The biggest change lies in the cost model. Cowork isn't included in the regular Copilot license — you pay for it separately through Copilot Credits. Every task is billed on four factors: the model chosen, the context retrieved, the number of tools used and the duration. Microsoft divides tasks into light, medium and heavy categories. A light task costs roughly 1 to 3 dollars, a short summary for example. A heavy task can run up to 7 dollars or more, think of comparing thousands of files between two product versions. Both pay-as-you-go and a pre-committed P3 volume with a discount are possible.

For IT leaders and CFOs, something fundamental is changing in how you budget for AI. A variable cost structure requires monitoring and clear agreements you didn't need with fixed monthly licenses. Which teams or roles are allowed to trigger agent tasks? Which types of tasks do you actually route to Cowork instead of classic Copilot? A structured rollout across a few use cases makes the added value tangible, without letting the bill spiral.

Cowork's cost model fits into a broader trend. Uber recently imposed a cap of 1,500 dollars per employee per month on AI usage. The internally grown practice of 'tokenmaxxing' (burning as many tokens as possible, sometimes with internal leaderboards) is colliding with the reality of the token bills. OpenAI reportedly booked a loss of 38.5 billion dollars in 2025, almost eight times as much as the year before. Anthropic paused a planned price increase for its Claude Agent SDK at the last minute. The pressure comes from price cuts by OpenAI and a planned IPO.

The message for Belgian organizations is that AI costs are becoming less predictable than a classic software license. A well-considered approach keeps it manageable. Work with a fixed AI budget per team or per use case and monitor consumption weekly. Evaluate every quarter whether the added value matches the cost. Test heavier agent tasks on a small scale first before rolling out more broadly. Structured governance around AI budgets is by now just as much a part of the digital workplace as the license choice itself.

The European privacy watchdog is warning about shadow AI: employees who use AI tools on their own initiative and, in doing so, unintentionally leak company data. The problem grows as consumer tools become more accessible and seem more convenient than the officially approved alternatives within the organization.

That same week, it emerged that ChatGPT recommends fake shops in its shopping results that are after payment data. The cause is what researchers call AI poisoning: bad actors manipulate the sources AI models train on or search with, so that harmful results end up in official answers.

Two risks that feed into each other. Shadow AI lets data leave the organization uncontrolled, AI poisoning lets unreliable answers enter the organization. A workable answer combines technology and awareness. Provide clear AI guidelines that give employees something to hold on to. Facilitate an officially supported AI platform so the need for shadow tools disappears. And invest in an ongoing training program so people know where the line is for what you can entrust to an AI assistant. Our cybersecurity and learning teams often work together at that intersection.

Norway recently became the first European country to largely ban generative AI from primary schools. The fear: children skip crucial learning steps and don't sufficiently develop their critical thinking skills. Flanders is choosing a different path. Minister Demir is freeing up 10 million euros for an AI roadmap for education. Responsible use and human oversight are central to it.

A decision with an impact beyond the school walls. Whoever joins the workforce in the coming years brings along the AI habits formed today in the classroom. For organizations, that means a double task. On the inflow side, employers can support the Flemish initiative by working together with education partners. On the internal side, AI literacy remains a basic skill that needs to be maintained at every level of the organization. That aligns with what Tom and his colleagues at OASE and Xylos Learning build every day: learning programs that teach people to prompt, evaluate and read critically, and that grow along with the incoming tooling.

We'll be back in two weeks with the next edition.

About Tom Van 't veld

Tom has worked for Xylos for years, where he started as a Microsoft Office trainer and grew into the driving force behind innovative learning concepts. He co-founded OASE, Xylos's online learning platform, and PlayForward, Xylos's new gamified learning brand. He also developed, among other things, the Digital Coach concept, a Microsoft Teams Escape Room app and the mAindset game, which helps employees learn to prompt AI in a playful way. As Learning Innovator, his focus in recent years has increasingly turned to what AI means for the way we learn and work. Want to respond or chat further? Find him on LinkedIn.