Every two weeks, we publish a sharp and honest overview on the Xylos blog of what’s really happening in the world of generative AI, complete with the necessary context. The foundation for this remains the biweekly LinkedIn roundup by Tom Van ‘t veld, Learning Innovator at OASE (powered by Xylos). Tom closely follows AI developments, and we translate his observations into what they actually mean for organizations and the people who work there. Welcome to edition seven.
ISSUE 7 • JULY 6, 2026
Below are the stories from the past two weeks that will be most beneficial to your organization.
Loop engineering: between real engineering and token accounting
“Loop engineering” is the term you’ve been seeing pop up everywhere in recent weeks. The idea is simple: instead of typing individual prompts, you let an AI agent run in a loop that adjusts itself. Try, check the result, try again, until a goal is achieved. Boris Cherny, the man behind Claude Code, summed it up by saying, “I don’t prompt anymore; I write loops.” Anyone who’s worked with agent-based tools will recognize this immediately. It boils down to giving a specific instruction in which you ask the agent to review its own work.
Still, there’s more to it than just a better prompt. The difference lies in memory and control. A prompt runs once; a loop remembers what went wrong and tries to do better next time. The real work lies in two things: clearly defining when something is finished, and building in brakes. Without those brakes, an agent keeps running, keeps consuming tokens, and racks up a hefty bill. Anthropic itself admits that such a multi-step setup can quickly consume fifteen times more tokens than a regular conversation. Choosing the right model for each task is also crucial here.
For organizations, this theme embodies a principle that goes beyond just a buzzword. The technology is real, the cost is just as real, and the benefit lies in conscious choices. From now on, you’ll focus on the finish line rather than every single step, and you’ll retain your human judgment to decide when it’s good enough. That’s why you should build agent-based workflows with a clear endpoint, a usage limit per use case, and a human review step. This way, you gain speed without the token bill becoming a surprise. Our AI experts will help you make those choices before you roll out the solution on a large scale.
More affordable standard models, but not every model is suitable for every task
Model developers are competing with each other by releasing faster and cheaper versions. Anthropic released Claude Sonnet 5, a model that comes close to the more expensive Opus 4.8 at 40 to 60 percent lower cost and immediately becomes the default for free and Pro users. OpenAI did something similar and made GPT-5.5 Instant the new standard for all ChatGPT users, with fewer hallucinations and more concise responses.
It sounds good on paper, though real-world experience reveals some nuances. When Tom tried to build a complex interface himself, Sonnet 5 got stuck, while the more expensive Opus solved the task on the first try. For more standard tasks in the Claude app, however, the cheaper model works just fine. Meanwhile, Google is pushing Gemini further from being a chatbot toward something that actually performs actions on your screen. Gemini 3.5 Flash now includes built-in computer usage, allowing an agent to view screens, click, and type on its own—at about one-third the token cost of GPT-5.5.
In addition, Gemini received a series of smaller features for the daily workflow. It has recently gained the ability to take notes during your calls in Google Meet—and now also in Zoom and Teams—and neatly compiles the decisions and action items into a document. This is handy, though the question remains: who grants you access to this data, and where does it end up?
The lesson for organizations is that the cheapest model is rarely the best solution for every task. A simple summary requires a different model than complex reasoning or sensitive code. By consciously choosing the right model for each use case, you’ll improve quality and keep costs under control. Document this decision-making process in clear guidelines so that employees know when to use a more robust model and when a simpler one will suffice.
Access to top models is becoming a strategic issue
The focus in recent weeks has been on governments that help determine who is allowed to use the most powerful models. OpenAI initially had to have its new GPT-5.6 approved on a customer-by-customer basis by the U.S. government, which meant that, for the time being, only about twenty vetted companies were granted access, while European customers had to wait. Shortly thereafter, the situation reversed: the U.S. government lifted the export restrictions on Mythos 5 and Fable 5, after which Anthropic began restoring access for foreign users with additional security measures.
At the same time, tensions between Anthropic and China continued to escalate. Anthropic accuses operators at Alibaba’s Qwen Lab of conducting nearly 29 million conversations with Claude through some 25,000 fake accounts in order to mimic its strongest capabilities. Alibaba, in turn, banned its employees from using Claude Code. The major players also realigned their strategies. Microsoft announced Frontier Co., a 6,000-person unit tasked with helping customers adopt AI, while Meta and Amazon indicated they wanted to reduce their reliance on a single model provider.
For Belgian organizations, a principle takes precedence over the specific model in this context. Access to top-tier models is no longer just a commercial choice; it is also shaped by geopolitics. VRT expert Tim Verheyden draws the same conclusion: AI and geopolitics have become intertwined, and Europe would do well to reduce its dependence on a single supplier. The right approach is to actively avoid vendor lock-in. Build your AI stack flexibly enough to switch to a different model, and maintain critical processes using a well-thought-out mix of suppliers with a clear migration path in case one provider drops out.
AI is shifting jobs, and humans remain the driving force
The impact on the workplace has become very real in recent weeks. Oracle cut about 21,000 jobs last year and explicitly cited AI as a factor. Ford did the opposite and rehired hundreds of experienced engineers after its AI-powered quality control system fell short. The company subsequently topped a major quality ranking for the first time since 2010. The most memorable example comes from IKEA: instead of laying off about 8,500 call center employees, the company retrained them as remote interior design consultants after its AI assistant “Billie” took over nearly half of the simple inquiries.
At the same time, the issue of costs keeps coming up. Gartner warns that AI programming could become more expensive than the average developer’s salary by 2028, mainly due to rising token consumption. A discussion about the value of human interaction in sales paints a reassuring picture: AI works best as an enhancer for employees, as long as a human retains final control. According to LinkedIn, messages drafted by AI but reviewed by a human generate 25 to 40 percent more responses, while fully automated messages are quickly recognized as generic.
The message for employers is twofold. AI is taking over tasks, and that is precisely why the ability to guide and evaluate AI is becoming a core competency at every level. Invest in reskilling rather than just in tools, and ensure that employees learn to create prompts, evaluate, and read critically. This aligns with what Tom and his colleagues at OASE and Xylos Learning build every day: learning programs that empower people and evolve alongside the new tools that emerge.
Creators are demanding to regain control over their work
A striking number of creators are taking concrete steps these past few weeks to regain control. Cate Blanchett launched a free registry at the European Parliament that allows artists to use color codes to specify the extent to which their face, voice, or movements may be used by AI, with support from figures such as Tom Hanks and Emma Thompson. Dutch publishers and authors are developing their own “book pact” to establish agreements on how their work may be used. And streaming service Tidal is the first major player to stop paying royalties for music generated entirely by AI.
What these three steps have in common is that the process of seeking permission is gradually taking shape. Whereas the discussion had long focused on principles, there are now records, agreements, and payment rules that define what is and isn’t allowed.
For organizations, this directly affects your own AI output and your brand. The same questions about origin and permission will soon apply to your content. Start establishing guidelines today for labeling AI-generated material and specifying which sources your models are allowed to use. This will help you build trust with customers and employees, and prevent surprises when copyright rules become even stricter.
The downside of too much AI: quality, privacy, and energy consumption
Meanwhile, companies that rely heavily on AI are feeling the downside. The now-familiar “workslop”—low-quality output that erodes trust and forces colleagues to spend time correcting errors—is becoming more common. Researchers also point to “AI brain fry” among some heavy users, caused by constantly checking AI responses.
Signs were also mounting regarding data and privacy. At Meta, internal keystrokes, screen recordings, and even employee evaluations—which the company had collected to train models—were leaked, after which the program was shut down. Starting July 8, Anthropic will ask users to verify their identity with a live selfie and an ID document to ensure that no minors are participating in the conversations, and says it will not use those images for training. Finally, the physical cost became tangible: during the recent heat wave, large data centers were found to be contributing to the warming of their surroundings, with cooling systems already consuming about 40 percent of their energy. Researchers at Cambridge estimate that this effect could impact more than 340 million people.
The lesson is that more AI doesn’t automatically mean more value. Without governance and human review, the output erodes trust and knowledge, and the risk increases that company data will leave the organization via consumer tools. A practical solution combines technology and awareness: clear AI guidelines, an officially supported platform that eliminates the need for shadow tools, and an ongoing training program that sets boundaries for what you entrust to an AI assistant. Our cybersecurity and learning teams often collaborate at that intersection.
We will be back in two weeks with the next edition.
About Tom Van ‘t Veld
Tom has been working at Xylos for years, where he started as a Microsoft Office trainer and went on to become the driving force behind innovative learning concepts. He is one of the founders of 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 how to use AI prompts in a fun way. As a Learning Innovator, he has increasingly focused his attention in recent years on what AI means for the way we learn and work. Want to respond or continue the conversation? Find him on LinkedIn.
