A Reader Asked a Simple Question

After the last edition, a reader left a comment that deserved more than a thumbs up. Show me the European companies that actually own their orchestration layer. Real ones. Case studies.

Fair. I’d just spent an entire edition telling Europe to own the rails. So I went looking for the people who already do.

Full disclosure before I start. I work personally with several EU enterprises that are on exactly this journey right now. They’re under NDA, so I can’t name them or share a single detail. That’s the reason for this edition. I know what this looks like from the inside. What I wanted to find out is who else out there shares the mindset and has the technical capability to back it up.

I found plenty of press releases. I found a lot of vendors selling “sovereign agentic platforms” with a straight face. What I didn’t find was a single European company that owns its entire agentic operating system, top to bottom, on the public record.

Before I show you who came closest, let me show you why it matters. Because the kill switch isn’t theoretical anymore. Somebody already pulled it.

In May 2025, after US sanctions landed on the International Criminal Court, Chief Prosecutor Karim Khan lost access to his Microsoft email. Microsoft says it never cut off the ICC itself. Then in February 2026, Microsoft asked the UK Parliament to correct testimony one of its own executives had given about the episode. [5] By then, the court had already decided to move roughly 1,800 workstations to openDesk, the open-source suite from Germany’s ZenDiS. [6]

That was an inbox. Now picture the same sanctions order landing on the agents that approve your claims, route your payments and answer your customers.

That’s the question I took to the field. Who in Europe could survive that call?

The Test Nobody Passes

I’m not handing out participation trophies. To own your agentic OS, six layers have to be yours:

• Compute. The GPUs, in a building whose lease you control.
• Model. Weights you hold and can post-train, not an endpoint you call.
• Gateway and routing. The layer that decides which model gets which request, and meters it.
• Agent runtime and orchestration. The code that decides what your agents do with your business.
• Identity, policy and audit. Who an agent acts for, what it’s allowed to touch, and the record that proves it.
• Context, memory and knowledge. What your agents remember, the knowledge they retrieve and the format it’s stored in.

That last layer is the one most vendors would rather you didn’t notice. Everything else you can swap with a migration. Memory is where switching cost compounds, because every conversation your agents learn from makes leaving more expensive. Rent it, and the day you leave, your agents wake up with amnesia.

Own all six and you can lose any vendor on a Tuesday and keep operating on Wednesday. Own three and you have a strong position with a known gap. Own one and you have a press release.

On the public record, nobody in Europe owns all six. Two companies own enough to be worth studying. A lot more rented the whole thing and found out what that costs.

The Closest Bank: BNP Paribas

The most sovereign agentic stack I found in Europe doesn’t belong to a startup or a government. It belongs to a bank with roots going back two centuries.

Su Yang runs AI transformation at BNP Paribas. His interview with The Big Whale in June carried a headline I’d frame and hang in every CIO’s office: “Agentifying a bad process just burns tokens.” [1] What he described underneath it is the closest thing to my playbook I’ve seen a European enterprise put on the record:

• Compute: GPUs bought directly from Nvidia and racked in BNP’s own datacenters. Not rented from a hyperscaler. [1]
• Gateway: an internal LLM-as-a-Service layer that most of the bank’s tokens run through, so any application can call a model without being built for AI. [1][2]
• Orchestration: built in-house, across about 15 models in production and 100 in experimentation. [1]
• Model: Mistral, picked partly because it sells downloadable weights the customer controls. [1]

Then comes the line that should make every frontier lab’s CFO sweat. Su Yang argues against going all-in on operating expense, because model providers don’t make money and may raise prices. BNP would rather spend capex now to shrink opex later. [1] That’s my OSS-baseline, frontier-burst thesis, executed by a bank with an IT budget of roughly €8 billion a year. And they publish the receipts: €635 million of AI value in 2025, €750 million targeted this year. [1]

Now the asterisk, because I promised not to give credit that wasn’t earned. BNP also buys Microsoft Copilot licenses and Microsoft’s Agent Builder. [1] I won’t pile on for that. Nearly every enterprise in Europe runs Copilot, and most treat it as an acceptable risk for everyday productivity work. That’s a reasonable call. It still dents the sovereignty posture, and a board should know exactly where. Its corporate and investment bank signed with Google to build agents on Gemini Enterprise, and wired Gemini models into LLM@CIB, the internal platform serving more than 65,000 employees. [3] The agent-building layer, the part that turns into your workflows, is partly rented.

BNP owns three of the six layers outright and a fourth in part. It hasn’t said a word in public about the sixth, where its agents’ memory and knowledge live. That’s still the best score in Europe I can prove. It’s still not six.

The Closest Builder: Zalando

If BNP owns the basement, Zalando owns the control room.

Zalando’s engineering team published a remarkably candid post in August about 2.5 years of agentic engineering. [4] Strip away the developer-blog modesty and here’s what they built:

• Gateway: an LLM proxy on open-source LiteLLM, live since January 2024, with their own cost tracking bolted on. It serves about 2,000 monthly users.
• Agent runtime: an agent platform assembled from open-source parts, with kagent running agents on Kubernetes.
• Identity and audit: an in-house Agentic Identity Broker that tracks delegation chains (which agent is acting for which human), brokers between OAuth systems and holds a token vault. It sits in the call path between agents and the tools they reach.

That last one matters more than anything else in this edition. Identity is the layer every “agentic platform” vendor wants to own, because whoever brokers the identity owns the agent. Zalando built theirs.

Their own summary of why: “Vendor independence is key in a fast-moving environment.” [4] Somebody should put that on a slide for every board in Europe.

The asterisk: every brain behind that proxy is rented, from OpenAI, AWS Bedrock and Google Vertex. So is the compute those models run on. And the platform serves Zalando’s own engineers, not customer-facing agents. Zalando owns the agentic OS layer and rents the model lineage. It’s the exact mirror image of the problem I flagged last edition, where companies obsess over the weights and forget the layer above them.

And like BNP, Zalando hasn’t published where its agents’ memory and knowledge live. For a company that built its own identity broker, that’s the obvious next layer to show.

Want proof that owning the routing pays? Look at Ecosia, which I covered in October. [15] Europe’s sovereign search engine swapped OpenAI for Mistral in May, then left Mistral in October, with CEO Christian Kroll telling Politico he was “disappointed with the quality of Mistral.” [13] Mistral answered by inviting Ecosia to try its newest model. [14] Whoever was right, Ecosia changed suppliers twice in one year. That’s what switching costs look like when leaving is a routing change, not a rebuild.

The Rented and the Burned

That’s the honor roll. It’s two names long. Here’s everyone else.

Builder.ai: the unicorn whose customers owned nothing. The London startup, backed by Microsoft and once valued at $1.5 billion, promised apps built by AI. The Wall Street Journal reported the platform leaned heavily on human engineers. Bloomberg reported former staff alleging sales were inflated by more than 20%. In May 2025, a lender swept its accounts and the company entered insolvency. [7][8] On Builder.ai’s own LinkedIn page, customers asked what the company was doing to get them the code for apps they had already paid for. [9] If you don’t own the code, you don’t own a product. You own an invoice.

Klarna: rent the brain, own the blame. Klarna built its assistant with OpenAI, bragged it was doing the work of 700 agents, and its CEO declared AI could already do all the jobs. Then quality fell and Klarna hired humans back. [10][11][12]

The Lesson From Paris

The most useful failure in Europe belongs to the French state.

France’s civil-service ministry admitted that its first AI assistant, built on free open-source models, was very experimental and gave poor results. [16] The state then signed a paid contract with Mistral to rebuild it as Assistant IA, rolled out to 10,000 civil servants. [17][18] On the surface, that’s a sovereignty retreat: open models out, a vendor in.

Look one layer down. Underneath the chatbot sits Albert API, the interministerial inference platform the state’s digital directorate built on government infrastructure. When the app on top flopped, the platform underneath didn’t. It became the base layer everything else is now built on. [17]

That’s the whole thesis in one government IT project. The model was disposable because the state owned the layer above it. Swap the brain, keep the nervous system.

The Five Pillars: Grading Europe’s Best

Same standard for everyone. I’m grading the two companies that came closest, BNP Paribas and Zalando, because grading the ones that rented everything would be a very short section.

1. Territorial

The question: Where do the data and compute physically sit?

The story: BNP racks its own Nvidia GPUs in its own datacenters. [1] Zalando sends every model call to OpenAI, AWS Bedrock or Google Vertex. [4]

Under pressure: My standing rule applies. A hosted frontier model anywhere in the request path fails Territorial by default. BNP’s core LLM-as-a-Service path passes. Its Gemini-powered CIB agents don’t. [3]

Verdict: BNP Paribas, Partial. Zalando, Not Demonstrated.

2. Operational

The question: Who holds the keys, runs the environment and keeps the logs?

The story: The ICC learned what happens when someone else holds the switch. [5] Zalando built its own identity broker and token vault, so it knows which agent acted for which human, and it holds the keys. [4] BNP runs its own datacenters and its own orchestration. [1]

Under pressure: Neither company has published how it audits agent actions end to end. Zalando’s broker is the closest thing to it in the public record.

Verdict: BNP Paribas, Demonstrated for infrastructure. Zalando, Demonstrated at the agent layer.

3. Technological

The question: Can you audit it, fork it and run it without the vendor in the room?

The story: Builder.ai’s customers couldn’t get their own code back. [9] Zalando’s agent stack is assembled from open source it can fork, LiteLLM and kagent included. [4] BNP built orchestration in-house and picked a model supplier partly for downloadable weights. [1]

Under pressure: Zalando’s brains are hosted models it doesn’t control. BNP’s Copilot seats are the acceptable risk almost every enterprise takes. Its Google-built CIB agents sit closer to the crown jewels, and that’s the bigger dent. [1][3] Each owns a different half of the stack. And neither has said where agent memory lives, which is the layer that decides whether you can actually leave.

Verdict: BNP Paribas, Partial. Zalando, Partial.

4. Legal

The question: Which jurisdiction can compel access, or switch you off?

The story: The ICC proved the scarier risk isn’t disclosure. It’s a sanctions designation that turns a service off overnight. [5][6]

Under pressure: BNP’s own stack sits in its own EU datacenters. Every Microsoft and Google path it uses carries US jurisdiction with it. For Copilot on everyday documents, most boards accept that trade, and fairly so. For agents drafting credit memos, the question deserves a harder look. Zalando’s entire model layer carries it.

Verdict: BNP Paribas, Partial. Zalando, Not Demonstrated.

5. Financial

The question: Who controls your unit economics?

The story: BNP is the only European enterprise I found that says out loud it won’t hand its AI budget to someone else’s meter. Capex now, less opex later, because providers may raise prices. [1] Klarna ran the opposite experiment. Cost came first, quality fell, and the humans came back. [11]

Under pressure: Zalando tracks every token through its own proxy, which is smart. But tracking someone else’s meter isn’t owning it. Under my rule, hosted frontier models in the path fail Financial by default.

Verdict: BNP Paribas, Demonstrated for its core platform. Zalando, Not Demonstrated.

The Executive TCO

Every AI budget I see has a line for tokens. Almost none has a line for leaving. That’s the one that kills you.

• Builder.ai’s customers found out the exit cost was a full rebuild, plus whatever they’d already paid. [9]
• Klarna optimized cost per conversation, then paid again to bring humans back when quality dropped. [11]
• Ecosia left twice in a year for the price of a routing change. Kroll says it halved costs, though he hasn’t published the numbers. [13]
• Zalando can see every token on its own proxy, which is the first step. Seeing the meter isn’t owning it.
• BNP Paribas is buying its way off the meter: capex on its own GPUs today so opex shrinks tomorrow. [1]

The executive version is one sentence. Your real AI cost is tokens plus the price of leaving, and only the companies that own the layer above the model get to set the second number.

And the most expensive thing to leave behind isn’t code. It’s memory: everything your agents learned about your business, stored in someone else’s format.

What to Do Monday

1. Run the six-layer test on yourself. Compute, model, gateway, agent runtime, identity and audit, and context and memory. Write down who owns each one. Be honest. Nobody’s grading you but your board.
2. Price the exit. For every AI vendor in production, estimate what it costs to leave in 90 days. If the answer is “rebuild,” that’s your Builder.ai exposure.
3. Own the gateway first. It’s the cheapest layer to take back and it makes every other switch a configuration change. Ecosia proved it. Zalando built it in 2024.
4. Own identity next. Whoever brokers the identity owns the agent. Don’t let a vendor’s platform be the only record of who your agents acted for.
5. Find out where your agents’ memory lives. Ask every vendor: can I export what my agents have learned, in a format I can load somewhere else? If not, every day you stay makes leaving harder.
6. Ask the kill-switch question in writing. If a sanctions order hit your provider tomorrow, which agents stop? Get the answer from your vendors before a regulator asks you.
7. Stop measuring cost per conversation alone. Klarna did that. Measure cost per resolved outcome, including the humans you’ll need when the model gets it wrong.

The Layer You Own Survives Your Mistakes

So, dear reader, here’s your answer. Show me a European company that owns its whole agentic OS? I can’t. Not on the record.

But I can show you a bank that owns the basement, an e-commerce company that owns the control room, and a government that flopped its chatbot and kept the platform. I can also show you a $1.5 billion unicorn whose customers owned nothing, a fintech that rented the brain and kept the blame, and a court whose prosecutor found out who really held the switch.

Notice the pattern. Every company that owned a layer survived its mistakes on that layer. Every company that rented the stack inherited someone else’s.

Europe doesn’t need a hero that owns all five layers by Friday. It needs a lot more BNPs and Zalandos owning two or three, and a lot fewer press releases claiming six.

And if you’re one of the companies doing this quietly and you’re ready to go on the record, my inbox is open. The next edition has room.

Amit

About the Author and References

About the author: Amit Eyal Govrin is CEO and co-founder of Agentcy Labs, which architects sovereign AI platforms for enterprises in Europe and the US. That work is why this analysis comes from the field: the five-pillar test is the one we work through with buyers every week. This piece follows theCUBE Research editorial standards.

1. Su Yang (BNP Paribas): “Agentifying a bad process just burns tokens”, The Big Whale, June 10, 2026.
2. BNP Paribas, LLM as a Service platform press release, June 5, 2025.
3. FinTech Futures, BNP Paribas inks new Google Cloud deal to advance agentic AI deployment.
4. Zalando Engineering, Agentic Engineering at Zalando: a snapshot, Aug 14, 2026.
5. The Register, Microsoft asks UK Parliament to correct record on ICC email, Feb 18, 2026.
6. heise, International Criminal Court kicks out Microsoft.
7. TechCrunch, Once worth over $1B, Microsoft-backed Builder.ai is running out of money, May 20, 2025.
8. Bloomberg, Builder.ai files for bankruptcy after creditors seize accounts, June 5, 2025.
9. The National, Builder.ai blame game begins after AI start-up goes insolvent, May 23, 2025.
10. Maginative, Klarna dials back its AI customer service strategy.
11. Entrepreneur, Klarna is hiring customer service agents after AI couldn’t cut it, citing Bloomberg.
12. Outlook Business, AI took their jobs, but Swedish fintech now wants human workforce back.
13. Cybernews, European Google Search alternative drops Mistral, citing Politico, Oct 2026.
14. Presse-citron, Mistral répond aux critiques d’Ecosia.
15. Amit Eyal Govrin, Ecosia Broke Up With Europe’s Sovereign Champion. It’s Not Me, It’s You, theCUBE Research, Oct 11, 2026.
16. Banque des Territoires, En 2026, le ministère de la Fonction publique promet d’accélérer sur l’IA.
17. Acteurs publics, Derrière l’échec médiatique d’Albert, un projet d’IA plus global, Feb 13, 2026.
18. Acteurs publics, L’État signe un contrat avec Mistral pour booster son chatbot.