Enterprise AI in Europe runs largely on infrastructure owned by a handful of US hyperscalers. That dependency carries risk: providers can change their terms, prices or availability, and access can be affected by decisions made outside your organisation. The EU is responding with legislation meant to strengthen European tech capacity and business autonomy.
For European leaders, that creates a balancing act. AI is fast becoming a core part of how businesses operate, not a nice-to-have, and that small shift in operating model brings a big shift in risk and compliance. This guide shows how to make the move without getting burned.
The new landscape: dependency and legislation
Most enterprise AI in Europe runs on a small number of providers, most of them based outside the EU. That creates concentration risk: when one provider changes course, every customer built on it is affected. Providers can raise prices, update or retire models, change their terms or suffer outages, and they must comply with the laws of the countries where they operate.
A recent example shows what this looks like in practice. On 9 June 2026, Anthropic released its most advanced models, Claude Fable 5 and Mythos 5. Three days later, the US Department of Commerce placed both models under export controls that barred access by foreign nationals. To comply, Anthropic disabled the two models for all customers, while its other models stayed available. Access was restored on 1 July. For nearly three weeks, businesses that had built on those models had to switch to alternatives or wait.
For a pilot, that is an inconvenience. For a process that depends on a single model, it is an outage. The lesson isn’t about one provider or one country: any service you don’t control can become unavailable for reasons outside your organisation.
The regulatory framework
At the same time, the EU has built up a body of legislation that shapes how enterprises use AI and the cloud services behind it. Here is what each piece means in practice, as of October 2026.
The EU AI Act sorts AI systems by risk and sets obligations accordingly. Bans on prohibited practices and the AI literacy duty already apply. The July 2026 “Digital Omnibus on AI” softened that duty to taking measures that support staff literacy, and moved most high-risk obligations to December 2027, and to August 2028 for AI built into regulated products.
GDPR governs any personal data that flows into or through an AI system, including customer questions and the data used to train or ground a model. A separate Digital Omnibus proposal would amend parts of it; it is still being negotiated, so today’s rules apply in full.
The Data Act has applied since September 2025 and gives you the right to switch cloud and data-processing providers. From 12 January 2027, providers can no longer charge you for switching.
NIS2 requires organisations in critical and important sectors to manage cybersecurity risks across their supply chain, which includes their AI and cloud providers. A January 2026 proposal would add EU-level supply-chain risk assessments for critical sectors.
DORA has applied to financial entities since January 2025. It requires them to manage ICT third-party risk, including exit strategies for critical providers, and brings the largest cloud providers under direct EU oversight.
New initiatives aim to strengthen European cloud and AI capacity. In June 2026, the Commission proposed the Cloud and AI Development Act, which would introduce four EU-wide sovereignty assurance levels for cloud and AI services, starting with public-sector and critical workloads. It is still a proposal, but it signals where procurement standards are heading.
The rules are meant to increase resilience and choice, but in the short term they add complexity for companies trying to move fast.
What’s wrong with existing thinking: AI as a feature vs. AI as an operating model
Most enterprises still treat AI as something to sprinkle on top of existing processes: a chatbot here, a copilot there. But as AI moves into core functions such as customer service, pricing, content production or decision support, it stops being a tool and becomes part of how the business operates. That brings new requirements for governance, data, skills, vendor management and compliance that a “sprinkle on top” approach never had to consider.
Take a hypothetical mid-sized Belgian retailer with an online shop of 20,000 products. In year one, the marketing team starts using an AI writing assistant. Copywriters use it to draft product descriptions faster, then edit them and translate them by hand into Dutch and French. If the tool goes down for a day, the team writes the old way. If the provider changes its model, someone notices the tone feels different and adjusts the prompt. The risk is real but small, and it sits with a handful of people who know exactly what they’re using.
Two years later, the retailer has connected that AI model directly to its product database. New products are described, translated and published automatically, in three languages, within minutes of being added. Customer service uses the same model to answer questions about those products. The copywriting team has shrunk from six people to two, who now review a sample of the output rather than writing it.
The technology hasn’t changed, but the risk profile has. If the provider goes offline, as Anthropic’s most advanced models did in June 2026, new products can’t go live and customer service slows to a crawl. If the provider quietly updates its model, thousands of descriptions might change tone or start containing errors before anyone notices.
Customer questions now flow through a third party, which raises GDPR questions that never came up when copywriters were drafting text. And if the retailer wants to switch providers, it discovers that its prompts, workflows and quality checks are all built around one vendor’s model, and the team that used to write descriptions by hand no longer exists.
None of these risks were visible when AI was a tool on the side. They appeared gradually, as AI moved from helping people work to doing the work itself. That is the shift most organisations miss: not a single decision to make AI critical, but a series of small, sensible steps that add up to a new operating model, without a new approach to risk to match.
A small shift in operating model is a big shift in risk
This might sound like a reason to slow down, hire a team of lawyers or launch a multi-year transformation programme, but it isn’t. Adopting AI without getting burned doesn’t require a complete overhaul. It requires knowing three things: where you are, what applies to you, and what could go wrong. Get those right, and most of the decisions that follow become obvious.
1. Know where you are in the adoption journey
In our previous post, we described three phases of enterprise AI adoption. Each phase delivers more value, and each raises the stakes when something goes wrong.
Phase 1: Get to know AI
Most enterprises are here. AI arrives through the tools you already use: an assistant in your CRM, a copilot in your office suite, staff experimenting with chatbots. The productivity gains are real but modest, and so are the risks.
They mostly sit with individuals: an employee pasting customer data into an unapproved tool, or trusting an answer they should have checked. What you need at this stage is a clear usage policy, an overview of which tools are actually in use, and AI literacy training, which the AI Act also requires.
Phase 2: Transform core functions
This is where AI starts to redesign how work gets done, as in the retailer example, and where most of the value lies. It’s also where risk shifts from the individual to the organisation. An outage or a model update now affects customers and revenue, data flows through third parties at scale, and some use cases may fall into the AI Act’s high-risk categories.
At this stage, you need an inventory of your AI systems and the processes that depend on them, a risk classification for each use case, vendor assessments that cover where your data is processed and under which jurisdiction, fallback plans for critical models, and clear ownership.
Phase 3: Reinvent the business
With core functions transformed, enterprises start building new offers, services and experiences on top of AI. AI is no longer just how you work; it’s part of what you sell. That brings a new level of exposure.
Depending on how you build and brand your AI services, you may become a provider under the AI Act rather than just a deployer, with heavier obligations. Customer contracts and trust depend on AI behaving as promised, and dependency on a single provider becomes a strategic risk to the business model itself. At this stage, AI governance belongs on the board agenda, and sovereignty needs to be built into your architecture choices rather than checked afterwards.
Moving between phases is rarely a deliberate decision. It happens through small, sensible steps, which is why many organisations operate in phase 2 while still governing as if they were in phase 1. Knowing which phase you’re really in tells you which of the next two steps need your attention now.
2. Understand your compliance obligations
You don’t need to master every regulation at once. Your obligations grow with your adoption phase.
- In phase 1, the essentials are GDPR and the AI Act’s AI literacy duty: know what data goes into which tools, and make sure your people know how to use them responsibly.
- In phase 2, as AI moves into core processes, the AI Act’s risk categories start to matter, and so do the resilience rules: the Data Act’s switching rights, and supply-chain requirements under NIS2 or, in financial services, DORA.
- In phase 3, when you build AI into what you sell, you may take on obligations as a provider rather than just a deployer.
A practical first step at every stage is to list your AI use cases and classify each one against the AI Act’s risk levels. It takes less time than most expect and immediately shows where your real exposure lies.
3. Map and manage your AI risks
Once you know your phase and your obligations, the next step is to understand what could actually go wrong. Most AI risk discussions focus on hallucinations and bias. Those matter, but for European enterprises in phase 2 and beyond, the biggest risks are often structural: who you depend on, where your data goes, and how fast the ground is shifting under your feet.
Vendor lock-in and concentration risk
When a core process runs on a single AI provider, that provider’s decisions become your business decisions. Prices can rise, terms can change and services can be withdrawn, all without your input. Lock-in rarely comes from contracts alone. It builds up in prompts, workflows and quality checks designed around one vendor’s model.
The Anthropic suspension showed the flip side: organisations with a fallback to another model or provider kept working, and the directive didn’t even touch Anthropic’s other models.
What to do: identify which business-critical processes depend on a single provider, design your architecture so models can be swapped, and test your fallback before you need it. The Data Act helps: from January 2027, cloud providers can no longer charge you for switching. But a legal right to leave is only useful if your systems are built to let you.
Geopolitical exposure and data sovereignty
Storing data in an EU data centre is not the same as keeping it under EU control. Microsoft has acknowledged it cannot guarantee European data sovereignty under the US CLOUD Act, which allows American authorities to request data held by US companies regardless of where it is stored. And as the Anthropic suspension showed, access itself can be restricted by decisions made outside your organisation.
What to do: for each AI use case, know where your data is processed, which jurisdiction the provider falls under, and what would happen if access were cut off. For your most sensitive data and critical processes, consider European providers or open-weight models you can host on infrastructure you control.
Model changes
AI models are not static products. Providers regularly update, replace and retire them, and a new version can change the tone, accuracy or behaviour of every output that depends on it. In a phase 1 pilot, someone notices and adjusts. In a phase 2 pipeline producing thousands of outputs a day, errors can spread before anyone spots them.
What to do: pin model versions where possible, track provider deprecation schedules, and set up automated quality checks that flag changes in output before they reach customers.
Shadow AI
If you don’t give employees approved AI tools, they’ll find their own. Staff pasting customer data, contracts or internal strategy into free consumer tools is one of the most common and least visible AI risks, and it can create GDPR exposure without anyone in management knowing.
What to do: treat shadow AI as a signal of demand, not just a compliance problem. Offer approved tools good enough that people don’t need workarounds, publish a clear usage policy, and make it easy to ask whether a tool is allowed.
Regulatory risk
Under the AI Act, AI used in areas such as hiring, credit scoring, education and critical infrastructure counts as high-risk, with strict requirements for risk management, documentation and human oversight. Those obligations now apply from 2 December 2027 for most high-risk systems, and from 2 August 2028 for AI built into regulated products.
That might sound like plenty of time, but fines for breaching them can reach €15 million or 3% of global annual turnover, whichever is higher, and building compliant processes takes longer than most teams expect. The risk also grows as you move through the phases: a company building its own AI services in phase 3 may take on the heavier obligations of a provider.
What to do: classify every AI use case against the AI Act’s risk categories now, and assign someone to own each high-risk system before the deadline, not after.
Pace of change and skills
Few technologies have moved as fast as AI. Models, tools, prices and regulations change every few months, so what your team learned last year may already be out of date. That makes skills a risk in their own right: people make mistakes with tools they don’t understand, and organisations lose the knowledge to do the work without AI, as the retailer’s shrinking copywriting team showed
The AI Act makes this a legal matter too: every company deploying AI must take measures to support AI literacy among its staff.
What to do: treat AI literacy as an ongoing programme rather than a one-off training, tailor it to each role, and update it whenever your tools or obligations change.
None of these risks is a reason to slow down. They’re a reason to be systematic. The organisations getting the most out of AI aren’t the ones avoiding risk. They’re the ones that see it clearly and manage it from the start.
What this means for you
If you recognise your organisation in phase 2 or 3, where AI is already part of how core work gets done, you don’t need a large programme to bring your risk approach up to speed. You need a few deliberate steps, taken now rather than after something goes wrong.
- Find your critical AI: List the processes where AI is no longer optional: the ones where an outage would stop work, affect customers or cost revenue. Most organisations find more than they expect, because AI became critical through small, sensible steps rather than one big decision.
- Map your dependencies: For each critical use case, write down which provider and model it runs on, where your data is processed, and under which jurisdiction. Then ask the question Anthropic’s customers had to answer in June 2026: what happens if access disappears tomorrow?
- Build your exits before you need them: Design critical processes so the model can be swapped, and test the fallback. For your most sensitive workloads, look at European providers or open-weight models you can run yourself. With switching charges disappearing in January 2027, now is a good moment to plan.
- Classify and assign: Sort your use cases by AI Act risk level, and give every critical or high-risk system a named owner. Governance without ownership is just a document.
- Make it a rhythm, not a project: Models, providers, prices and rules all change every few months. Review your AI landscape quarterly, update your training when your tools change, and treat your risk map as a living document.
None of this needs to slow down your AI ambitions. In fact, it’s what lets you scale them with confidence. The organisations that struggle are the ones that change their operating model first and think about risk later. The ones that succeed change both together.
Not sure which phase you’re in, or where your biggest exposure lies? Our AI readiness benchmark gives you a clear picture in under an hour, covering your data, your team’s AI fluency and the use cases closest to real value.
A final thought
Every European enterprise is playing with matches now. AI is too valuable to ignore, and your competitors have already struck theirs. The question isn’t whether to light the fire, but whether you know what you’re holding.
In June 2026, Anthropic’s customers learned that a frontier model could become unavailable overnight, for reasons entirely outside their control. They hadn’t done anything wrong. They simply depended on something they didn’t own.
So before your next AI initiative, ask one question: if access were cut off tomorrow, would your business keep running? If the answer is yes, you’re ready to move faster. If it’s no, you know exactly where to start.