Look, I get it. You clicked on this article hoping for another major thinkpiece about how Europe is going to build the next ChatGPT killer, probably powered by renewable energy and GDPR-compliant wet dreams. But sorry to disappoint you my friend. That ship sailed already caught fire, and sank somewhere between the American frontier model labs that are burning billions in compute and the Chinese manufacturing juggernaut that is already embedding AI into everything they have from cars to those suspiciously cheap smart toothbrushes that I see flooding Amazon.

So I have been thinking, what is Europe’s actual move in this AI arms race?

Well, it is not what you think.

And it’s definitely not what Ursula von der Leyen’s PowerPoint deck promised at the last Digital Europe Summit‡. But before I get there, let’s take a clear-eyed look at what’s actually happening in the global AI landscape, starting with the truth that does not fit on a slide about Europe’s current A-“strategy”.

‡ AI, bureaucrats & lots of broken promises | LinkedIn

---

More rants after the messages:

1. Connect with me on Linkedin 🙏
2. Subscribe to TechTonic Shifts to get your daily dose of tech 📰
3. Please comment, like or clap the article. Whatever you fancy.

---

Three players and three completely different games

When you squint your eyes a bit and look at he global AI landscape, you can vaguely see that each continent has a focus on whatever part of the AI industry. America is basically a venture capital-fueled casino that is betting billions that if they just make the models bigger then surely AGI will emerge like some kind of messiah which will immediately crown the country and its maker to the winner of the AI competition. But because that isn’t enough of a strategy, they’re simultaneously kneecapping China’s chip supply because free markets are for winners.

Now China being, um, China, they looked at this competitive masturbation and said “cool pitch el presidente” before proceeding to put AI in literally everything from cars to the seventeen million trillion $4.99 gadgets that are pouring out of Shenzhen, that somehow actually work until they don’t and generate obscene amounts of revenue. Through Tik Tok.

And meanwhile our Europe – ye blessed olde continente – it sat in the corner with its books full of regulatory frameworks and ‘ethical’ principles and they discovered that its most lucrative AI export isn’t innovation or deployment but . . . fines.

It’s not ASML, though their lithography machines are genuinely impressive. It’s not Mistral either, though they’re trying very hard and burning through investor euros at an admirable pace.

Ursula has turned enforcement into a business model that is so freaking successful that the penalties we levy against AI companies exceed the export value of every European AI startup combined. They’ve made close to half a billion last year on fines they sent to Zucky, Noel Skum and the rest of the bro-ligarchy. Now, that may either be the most depressing indictment of our technological irrelevance or the most brilliant accidentally-on-purpose strategy for monetizing other people’s innovation, depending on how many glasses of French wine you’ve had and whether you’re trying to pitch this to the 60.000 European Commission technocrats.

The US → The “let’s find AGI” approach

The American AI strategy is beautifully simple and completely insane.

It goes something like this . . .

You throw ungodly amounts of venture capital at anything that might lead to AGI. Then you let a handful of companies monopolize the frontier model space, and then export the entire stack as a geopolitical product and simultaneously kneecapping China’s access to the chips that make it all possible.

Voila.

OpenAI, Anthropic, Google DeepMind, these are ‘manifestations’ of American capital’s terminal case of “what, uh, if we just kept, um, scaling?”. That strategy is literally “bigger model, more compute, surely AGI must be just around the corner”.

And you know what. It’s kind of working. GPT-4 writes better marketing copy than most marketing directors. Claude can debug code that human developers gave up on. Gemini can… well, Gemini exists and scores very well on benchmarks no sane working-class human ever reads anyway.

The U.S. produced 40 “notable AI models” in 2024. Get a load of that you wine drinking, croissants rolling, tulip farming has-beens, because Europe has three when I last counted, and I can only recall Mistral because as it happens I am implementing it somewhere. And before you ask . . . no, “notable” doesn’t mean “actually deployed anywhere”, it’s just that someone wrote a paper about it and got some press coverage. But the point stands, America owns the capability layer. They’re building the engines.

China → The Artificial Industrial blitzkrieg

‘China, China, China’ – to quote my favorite geopolitical chaos agent – Xi’s men looked at the American obsession with AGI and they went “that’s nice sweetie, now watch us put AI in literally everything while you MAGA”. Their AI+ strategy is definitely not about building the smartest model because when a new model appears they simply suck it dry, copy it and regurgitate it in the form of a new release of their Qwen, Kimi, DeepSeek, Doubao, Yi, Baichuan, ChatGLM, Ernie, Tongyi Qianwen, or whichever of the seventeen virtually identical LLMs happened to launch that week…

They are all about building the most useful products. And that is a heck of a cool strategy.

And holy shit, have they delivered.

Manus AI.

Say no more. It has been around for a long long time now, and they still rain supreme. Not because of the model though, but the tooling, the practical stuff – the blue collar AI. These guys keep on amazing me and no-one comes close, not even whatsitsname that I recently de-installed because too chaotic, um, oh yeah, the clawd-thingy. I still think 20 coin for Ebenezer Altman’s first love-baby is a tad too much, especially because this AI has caught a special case of ‘token frugality’, simply because Sam ‘the Scam’ needs to save money because he ain’t making enough, but I cough up 950 coin a month without a sweat for a Manus subscription – made possible by mr. Zucky who recently bought M’anus because his own AI strategy sucks big time.

And it’s not only Manus. Also Abacus and Skywork, they’re all trying to be useful to us users. These things actually solve problems people have. The platform integrates into workflows so smoothly you’d think it was designed by people who actually use software for work instead of just theorizing about “the future of work” in Medium posts (ahem…😅).

China has weaponized AI at every layer of society. Have you seen the tsunami of cars that are hitting the shores of Europe and Canada (sort of Europe)? The word ‘car’ is a dysphemism in this case – the opposite of a euphemism – because these things are basically driving computers that married an LED factory, had a four-wheeled love-baby, and somehow still cost less than a used Honda Civic with “minor water damage”.

Yes people, it’s in their cars, their factories and also in the seventeen million trillion billion gadgets that are pouring out of Shenzhen you guys can buy on AliExpress and somehow include “AI-powered” features that kinda work – until they don’t.

And the Chinese are not trying to build HAL 9000, but building ten thousand specialized, practical, revenue-generating applications that make existing industries more efficient. It’s unglamorous and it is not going to get you on the cover of Wired but, hell yeah, it’s making them filthy rich and technologically indispensable.

Um, Europe. The “we definitely have an, um, strategy” confusion

And then there’s Europe.

Oh, Europe.

Land of the gee-dee-pee-are. Home of the Digital Markets Act. Birthplace of the world’s first comprehensive AI regulation. The place where innovation goes to get 47 different compliance approvals before it’s allowed to proceed and eventually suffocates under the weight of the approval forms.

And, as I stated earlier, Europe’s most lucrative AI product—the one that generates more revenue than all European AI companies combined – is fines.

I’m serious. The penalties that the EU levies against AI companies for various regulatory violations exceed the export value of every European AI startup put together.

You don’t have to take this statement at face value because I did what any responsible researcher would do in 2026 – I unleashed my AI agents – my little Oompa Loompas built on CrewAI and I sent them out with the mission to find out exactly how much money Europe makes from AI exports versus how much it rakes in from fining Big Tech companies.

I gave them access to Eurostat databases, European Commission enforcement records, company financial statements, and told them to come back when they had numbers.

And they have, and I call it . . .

---

The spreadsheet of shame

I am going to paint you a picture with some numbers that would make any European Commissioner immediately schedule an emergency meeting to discuss “strategic communication frameworks around digital economy metrics”.

Europe’s AI market was valued at roughly €25 billion in 2023.

That’s impressive. It is the kind of number you put in a press release with three exclamation marks and maybe a stock photo of a diverse group of people pointing at a holographic display.

The EU exported €219 billion worth of high-tech products in 2023, with stuff like semiconductors but also pharmaceuticals, aerospace, the whole nine yards. That’s a real economy doing real things with companies building actual products that employs people and generates tax revenue and doesn’t require a Medium post to explain why it matters.

Now let’s talk about the other revenue stream. Regulatory vengeance.

The EU imposed at least €3.2 billion in fines on Big Tech companies in 2024 alone. The EU has crossed €10 billion in total penalties since enforcement began in earnest around 2017.

That’s billion with a B. Ten of them.

And who are the lucky recipients of these strongly-worded invoices with very large numbers at the bottom?

Google leads the pack with over €9 billion in fines since 2017 for “crimes” against competition including Android bundling, Google Shopping favoritism, and the apparently illegal act of being really, really good at search advertising while also selling ads.

Apple clocked in at around €2.5 billion for App Store anti-steering violations and that whole music streaming antitrust thing where they had the audacity to charge a commission on their own platform. Scoundrels!

Meta accumulated roughly €2.8 billion for their innovative “pay or consent” model (which the EU politely suggested was maybe not super legal), GDPR data transfer violations, and the usual competition law shenanigans that happen when you’re a social media monopoly.

Amazon got off relatively light with €746 million for GDPR violations related to advertising and data processing—basically a rounding error for a company that size, roughly equivalent to what Jeff Bezos spends on rocket fuel in a slow month.

Even Xwitter managed to score €550 million as the first company fined under the Digital Services Act for transparency failures and generally being terrible at literally everything since Elon took over.

And these are just the headline acts. Qualcomm, Intel, and a rotating cast of supporting players have all paid their dues to the European regulatory machine.

Ok, those €3.2 billion in fines from 2024 is only 1.46% of the €219 billion in high-tech exports, so pocket change in comparison, a rounding error. And if you were a nitpicking SoBi you’d say that it’s the kind of number that makes you go “oh, okay, so the fines are significant but obviously not comparable to actual economic output”.

But wait. Let’s get more specific about the Big Tech and AI part. Because not all high-tech exports are AI-related. The €219 billion includes pharmaceuticals (a huge chunk), aerospace, telecommunications equipment, and a bunch of other stuff that has nothing to do with machine learning or neural networks or any of the buzzwords that make venture capitalists reach for their checkbooks.

The real AI champions in Europe, the companies actually generating AI-specific revenue, are basically ASML (the Dutch lithography monopoly with €28 billion in 2024 revenue and a €300 billion market cap). They have a global monopoly on EUV lithography machines, which are essential for making the chips that run AI. That’s real strategic power and they have the position you want to be in when the semiconductor wars start heating up. And then there’s Mistral AI (the French LLM upstart with $40 million in revenue and a €6.5 billion valuation as of September 2024). They’re trying, and they surely raised a lot of money (recently 1,5 Billion Euro from ASML), they’re building open-source models, and they’re positioning themselves as the European alternative to OpenAI. Yeah, right. But whether they actually generate sustainable revenue at scale remains… an open question requiring more investor capital and possibly divine intervention.

SAP and Siemens are also in the European Tech & AI game where they’re integrating AI into their existing enterprise software stacks. Which is a smart move and they’re actually making money instead of just burning through venture funding and promising that AGI is just around the corner.

Here’s the thing that made me genuinely laugh out loud when my Oompa Loompas brought back the data, neatly formatted in a PDF that somehow made the absurdity even more absurd.

The €10+ billion in total fines since enforcement began is a genuinely impressive number. It represents real money extracted from real companies for them misbehaving slightly.

But what’s wild is that it’s more than the entire annual revenue of most European AI companies combined.

Let me say that again, slower, for the people in the back.

The cumulative fines we’ve extracted from American tech companies for violating European regulations exceed the export value generated by European AI-specific companies.

Europe has created a business model called regulatory arbitrage as economic policy.

Think about it from a pure ROI perspective, the European Commission employs, what, sixty thousand people working on competition and digital regulation? Their combined salary spend is maybe four to five billion per year? I’m being generous here.

And with that relatively modest investment, they’ve generated €10 billion in revenue by… sending very official-looking letters to Sundar Pichai explaining why Google’s behavior is illegal under Article Whatever of Treaty Something-or-Other. And meanwhile, building a competitive LLM from scratch requires hundred of millions in compute and a team of PhD-level researchers with stock options, years of iterations and constant angst that OpenAI will release GPT-6 tomorrow and make your model worthless overnight.

Which business would you rather be in?

Option A. You hire 50,000 engineers, build massive datacenters, burn through billions in training runs, compete with companies that have effectively infinite funding, and hope you don’t get disrupted six months after launch.

Option B. You hire 5,000 lawyers and regulators instead, write very stern laws, wait for American companies to violate them (they will), then send invoices with very large numbers.

Europe chose Option B. And honestly, from a pure profit-margin perspective, it’s kind of genius.

Europe has turned enforcement into their primary value-add to the global AI economy. They’re the unwanted, scoffed at parking meter attendants of the digital age, except instead of quarters they’re extracting hundreds of millions in penalties for cookie consent violations.

Is this really what Europe brings to the table?

Legislation and silicon wafers? Are they just the world’s most expensive compliance department with a side business in semiconductor manufacturing equipment?

What if this IS the strategy? Not deliberately, necessarily. But through bureaucratic incentives, and the path of least resistance, what if Europe has accidentally discovered that it’s better at extracting value from American tech companies than building competitive alternatives?

---

Here’s a modest proposal for a novel European AI strategy

Okay my dear over-intelligent friend, it’s fantasy time. Let’s say tomorrow morning I wake up and somehow I’m running European AI policy. My first move would be to fire 50,000 technocrats. Not because I’m a cruel person, but nobody needs that many people to “coordinate stakeholder alignment on AI readiness frameworks version 6.x.y”, and quickly following that, my second move would be to take the remaining 10,000 bureaucrats and retrain them as AI engineers. Okay, realistically they’d probably end up as content labelers – you know, the people who teach AI what a stop sign looks like by clicking on grainy images eight hours a day – but “AI engineer” sounds better in the press release.

But my third move. Oh baby.

This is where I actually figure out what Europe is genuinely good at that isn’t “sending strongly worded letters about GDPR compliance” and then suing the heck out of Big Tech so I can buy my staff their Christmas presents.

And the thing is that Europe is good at something, not only silicon wafers and paper pushing. We are actually excellent at thinking about ethics and governance and safety and yes, also rules. That’s not sarcasm – we (they, still in denial) actually are! The problem with “ethics” is that it isn’t exactly a revenue-generating product category unless you’re an “AI Ethics Consultant” billing the Amsterdam municipality €350/hour to attend meetings about responsible chatbot deployment‡.

But what if – and hear me out here – what if ethics, safety, and governance could actually be infrastructure AND make you money?

‡ Amsterdam’s Ethical AI fairy tale went spectacularly tits up | LinkedIn

---

Everyone’s AI is a black box on fire

Here’s what’s happening right now in every serious AI deployment I’m watching. Companies are taking processes that should be automated by deterministic systems and they’re now running them through a probabilistic Russian roulette engine just because this automation happens to have a ‘brain’ with a tendency to lie in your face, but with confidence.

You remember the time of the good old-fashioned rules-based automation systems where you know exactly what’s going to happen? You’ve probably heard of these knowledge management systems with explicit rules that determine only a few outcomes, and not a superposition of a whole Schroedinger equation. I am talking about the stuff that is Symbolic, deterministic systems where IF condition A THEN action B, every single time, no surprises, no “the model decided differently today because Mercury is in retrograde”, just clean, predictable, boring-as-hell logic trees that actually let you sleep at night because when something breaks you can trace it back to the exact line of code that fucked up instead of staring at a 175-billion-parameter black box wondering why it suddenly started hallucinating insurance approvals for fictional customers.

And they’re replacing them with probabilistic agentic AI. Or worse.

Why?

Because it’s 2026 and if you’re not “leveraging LLM capabilities to unlock transformative business value” then what are you even doing with your project funding?

The result is predictable chaos.

You have AI agents making decisions that affect real humans with real money and oh so real outcomes. And when something goes wrong, which it will, because these systems are statistical models, not oracles, then nobody is able to explain what happened.

The model made a decision.

But why “the model made a decision” totally eludes you.

Yeah, that’s helpful.

This is where observability comes in. And explainability. And all the other -ilities that sound boring until your AI agent accidentally approves a $3 million insurance claim because it misunderstood a pdf with a coffee stain on page 47.

Got it?

Now get this, I’m building an AI factory right now. The entire architecture is systems that need to be watched, monitored, controlled, and explained when they inevitably go sideways. And all of those controls, the monitoring infrastructure and all of those safety mechanisms, that is what I call scaffolding.

And scaffolding, my friends, might be where Europe can actually win – because the future is in wrapping your AI in a straight jacket and giving it unlimited access to every thing that pushes electrons around.

---

What the heck is scaffolding anyway?

Scaffolding is everything that surrounds an AI model to make it safe, auditable, and deployable in the real world. And remember, it’s not the model, but the cage around the model. The rules and permissions and the logging or the human oversight triggers, the audit trails, safety checks, rrrrollback mechanisms and the “are you absolutely sure you want to do that?” system-prompts that prevent your AI from nuking the production database. And yes, it also includes the tools- and connectivity layer.

Think of it this way, if the AI model is a brilliant but chaotic worker with a philosophy degree and access to unlimited coffee, then scaffolding is the entire corporate structure that prevents that person from accidentally emailing the CEO’s draft resignation letter to the entire company.

And there’s a more technical term for scaffolding, and I call it . . .

The achitecture of not fucking up

Scaffolding breaks down into a few critical layers, and each one of those layers is preventing a different flavor of disaster.

Let me bulletize this, without using bullets.

You have . . .

Procedural scaffolding which is the “how should it behave” layer. This is your workflows and guardrails and the step-by-step processes that prevents the AI from doing whatever the hell it wants and it is the difference between telling your workers to “increase revenue” versus “increase revenue by implementing these specific tested strategies, getting approval at each stage, and documenting your decisions for the audit committee”.

And then there’s semantic scaffolding. This is the “what does it actually understand” layer. This is where you use ontologies‡ and knowledge graphs that let the model actually know what it’s talking about because an LLM trained on the internet doesn’t know that “turkey” in “I’m cooking a turkey” is different from “turkey” in “Turkey shot down a Russian jet”. Semantic scaffolding prevents that kind of conceptual chaos.

Then there’s tool access control, which is the “what can it touch” layer. Your AI agent needs access to tools and APIs, databases, MCP stuff, the works. But do you really want it to have unrestricted access to everything? I do. But I use a different machine for it so it cannot screw up my main body of work. And scaffolding defines exactly which tools are available, with what parameters and under what conditions they’re allowed to be used.

And then there’s observability and logging. I call it the “what the fuck happened” layer. Every decision or tool call and every piece of data that was touched gets logged and is traceable for compliance and audits, and the inevitable moment when someone demands to know why the AI did that stupid thing it did.

And last and still quite necessary is the human oversight ‘integration’. This is the “emergency brake” layer. For high-stakes decisions, scaffolding can force the system to stop and ask a human before proceeding.

The seven-layer burrito of production AI

If you want to get properly nerdy about it – and of course you do, you read this far – a production-grade agentic system is a seven-layer architectural stack.

Allow me to present a ChatGPT worthy summary of ‘stuff’ that is ‘making the AI actually do it’s magic’.

1. You got your data persistence layer. Where you store everything the agent needs to remember.
2. Then there’s the security & safeguards layer. Rate limiting, input sanitization, all the stuff that prevents obvious attacks.
3. The AI service layer. The actual connection to your LLM, including fallback handling when OpenAI inevitably has an outage.
4. A multi-agent orchestration. How multiple agents coordinate without turning into a knife fight.
5. An API gateway. Secure endpoints, authentication, the front door to your system.
6. Observability & testing stuff. Metrics, evaluation, knowing what’s actually happening.
7. And the User Interface. The part humans interact with, hopefully without screaming.

Notice what’s missing from that stack?

The actual model is just one layer. Everything else is scaffolding.

And for the nitpicking reader, yes, what I now call ‘the model’ also has a lot of ‘scaffolding’, but you have to draw a line somewhere to make a point. Life ain’t all black and white you know.

‡ Generative models guess, ontologies clean up the mess | LinkedIn

---

The scaffolding winners

There are already some companies on the old continent that figured this out. They’re not building models, but creating the industrial infrastructure around models.

Take LangChain. They started as a simple wrapper library and evolved into a full orchestration framework. They don’t train models themselves and they simply don’t care which model you use, but what they do care about are the pipes and chains, the workflows that make models actually useful.

And when I joked about OpenClaw, it was just for a bit of pun, because I’m genuinely impressed by this thing. And yes, OpenClaw is an example of pure scaffolding with zero AI. They’re an integration layer, a control plane, a policy engine. They are selling the cage but not the tiger.

And there’s also Agent0 – another true open-source hero who builds the runtime environment for autonomous agents. They’re not the ones that are teaching the agent what to do, but they’re building the environment that keeps the agent from going rogue, and offering a lot of cool integrations that really do a lot of things a human would otherwise have to do.

These companies understand something fundamental. They know that in the long run, models become commoditized. Every six months someone releases a new “state of the art” model. But the infrastructure – the scaffolding that makes those models safe and useful – that is durable and in the end that is the layer that provides the ROI. Because of the tools-stack and because of the evidence it provides.

And that my friends, is where the moat for Europe is.

---

Europe should bet the farm on this

Here’s my strategic argument.

The U.S. owns capabilities. They’re going to keep pushing the frontier. Fine. Let them spend the billions on energy for training.

China owns deployment. They’re integrating AI into everything. Great. Good for them. We can always use another toothbrush that thinks for me.

And Europe can own trust infrastructure.

We’re already good at the things that scaffolding requires. . .

• Like writing detailed rules… ✅
• Thinking about edge cases and risks… ✅
• Building standards and certification frameworks… ✅
• Actually caring about transparency and accountability… ✅

The EU AI Act is already creating demand for exactly this kind of infrastructure. Every company deploying AI systems at scale needs risk management processes (scaffolding) and automatic logging systems (yup, scaffolding), human oversight mechanisms (aha – scaffolding) and audit trails for regulators (also scaffolding).

Europe is writing the rules, and yes, they suck, but they created a market for the infrastructure to enforce those rules at the same time.

And the beautiful thing is that this infrastructure is exportable.

The U.S. wants to export their full-stack AI platforms, and that’s great, those platforms will need European-style compliance scaffolding to sell into regulated markets. And Europe can be the ones who build that layer.

China wants to sell AI products globally. Wow, fantastic 👍 but those products will need trustworthiness certification, and Europe can provide the scaffolding that makes that certification possible.

The genius move here is turning the regulatory strength from a barrier into a platform.

AI compliance is mostly theater right now. Companies write “AI ethics principles” that are like corporate horoscopes, they have “responsible AI teams” that get overruled when there’s a deadline and they publish transparency reports that reveal approximately, um, nothing.

But scaffolding makes compliance executable.

At runtime.

Scaffolding is not a PDF. It’s code that runs before your AI agent can do anything dangerous.

Just think of it . . .

• Policy-as-code that automatically blocks prohibited actions before they happen
• Audit-grade observability that generates compliance evidence in real-time
• Risk-based human oversight that intelligently escalates decisions based on actual impact
• Portable sovereignty that works across any cloud, any model, any infrastructure

And that, my friend, is a control plane for the agentic economy.

And it’s an export product.

A certification standard.

A service layer.

An infrastructure category.

Need I go on?

---

The Eigenvector play. The architects of trust

Hahaaa – I got you a second time. Lured you into a great looking article, only to start selling you stuff like it’s an advertorial. But hey, at least I’m transparent about it, huh. The European way. You feel pain, but at least you know why I made it hurt.

So, if you’re still here, and you’re wondering where our company‡ fits into this grand vision -and of course you are, he said with a smug grin – here’s the pitch. We at Eigenvector, we’re not building the models, but we’re building model-agnostic scaffolding that makes agentic AI predictable, auditable, tool-rich and compliant-by-default. Portable across infrastructure of course, and works with any model, and turns EU AI Act requirements or any other corporate rule-book from legal obligations into runtime guarantees.

And that’s why we have a research agenda.

We’re researching and developing. . .

1. Compliance-as-code runtime engines – policies that execute before your AI can do stupid shit
2. Auditability-by-design logging architectures – because “the model made a decision” isn’t an acceptable answer to regulators
3. Human oversight engineering (thresholds, UX, anti-automation-bias) – the “are you absolutely sure?” layer that prevents dumb shit from happening.
4. Tool governance and blast-radius reduction (yup, that’s a real name) – sandboxing so your agent can’t accidentally DROP TABLE everything
5. Continuous assurance loops – ongoing risk monitoring instead of “we tested it once in 2023”
6. Reliable tool calling – making sure your agent actually invokes the correct API instead of improvising
7. Multi-agent coordination protocols – so your agents don’t turn into a knife fight over shared resources
8. Context-aware permission systems – different access rights based on what the agent is actually trying to do
9. Semantic constraint modeling – ontologies and knowledge graphs that prevent conceptual chaos
10. Rollback and recovery mechanisms – the undo button for when your agent goes rogue
11. Policy DSL engines – domain-specific languages for expressing “you may NOT do that” in executable code
12. Cross-model portability layers – scaffolding that works whether you’re running GPT, Claude, or whatever open-source model is trendy this week
13. Deterministic workflow orchestration – explicit control flow instead of “let the model figure it out”
14. Audit artifact generation – automatic compliance evidence that regulators can actually read
15. Agent behavioral fingerprinting – detecting when your agent starts acting weird before it causes damage

The deliverable is not the next chatbot though, but the reference architecture for corporate and EU-ready agents in production, a policy engine for tool permissions and an audit pack that regulators can actually use.

This, my friend, is what I call ‘Scaffolding’. With a capital S. And if Europe plays this right, they stop only regulating the AI, but build the layer where regulation becomes reality. Europe should not invest time and money and wasting smart people’s energies on the next best thing, and certainly not try to beat America or China at capabilities, nor hyper-specialize in building AI that matches the EU-country’s brand like the Dutch with their water-management AI – BARF!. Europe should be building the trust layer that makes those capabilities deployable and the governance and tool-calling infrastructure that makes that scale acceptable in democratic societies.

Models determine what’s possible but scaffolding determines what’s allowed and what’s allowed shapes what’s real and if Europe builds that layer and if they turn that regulatory instincts into exportable infrastructure, that is the chance of Europe becoming strategically relevant. And if they just resort to writing more stupid rules without building the tech to enforce them, they’re a museum.

Choose wisely, Brussels.

Signing off,

Marco

‡ How a rubber toy taught me more about AI and politics than any tech conference | LinkedIn

---

I build AI by day and warn about it by night. I call it job security. Big Tech keeps inflating its promises, and I just bring the pins and clean up the mess.

---

👉 Think a friend would enjoy this too? Share the newsletter and let them join the conversation. LinkedIn, Google and the AI engines appreciates your likes by making my articles available to more readers.

To keep you doomscrolling 👇

1. I may have found a solution to Vibe Coding’s technical debt problem | LinkedIn
2. Shadow AI isn’t rebellion it’s office survival | LinkedIn
3. Macrohard is Musk’s middle finger to Microsoft | LinkedIn
4. We are in the midst of an incremental apocalypse and only the 1% are prepared | LinkedIn
5. Did ChatGPT actually steal your job? (Including job risk-assessment tool) | LinkedIn
6. Living in the post-human economy | LinkedIn
7. Vibe Coding is gonna spawn the most braindead software generation ever | LinkedIn
8. Workslop is the new office plague | LinkedIn
9. The funniest comments ever left in source code | LinkedIn
10. The Sloppiverse is here, and what are the consequences for writing and speaking? | LinkedIn
11. OpenAI finally confesses their bots are chronic liars | LinkedIn
12. Money, the final frontier. . . | LinkedIn
13. Kickstarter exposed. The ultimate honeytrap for investors | LinkedIn
14. China’s AI+ plan and the Manus middle finger | LinkedIn
15. Autopsy of an algorithm – Is building an audience still worth it these days? | LinkedIn
16. AI is screwing with your résumé and you’re letting it happen | LinkedIn
17. Oops! I did it again. . . | LinkedIn
18. Palantir turns your life into a spreadsheet | LinkedIn
19. Another nail in the coffin – AI’s not ‘reasoning’ at all | LinkedIn
20. How AI went from miracle to bubble. An interactive timeline | LinkedIn
21. The day vibe coding jobs got real and half the dev world cried into their keyboards | LinkedIn
22. The Buy Now – Cry Later company learns about karma | LinkedIn