The Sovereignty Clause: Buying Automation Under Europe's New Data Rules

Sovereignty stopped being an infrastructure decision

The numbers make the shift hard to dismiss. In a forecast published on 9 February 2026, Gartner put worldwide sovereign cloud infrastructure-as-a-service spending at roughly 80 billion dollars for 2026, a 35.6% increase over the previous year, rising further in 2027. Europe is among the fastest-growing regions in that forecast, at around 83% growth, and Gartner expects the trend to move about a fifth of current workloads from global cloud providers to local ones. Governments remain the largest buyers, followed by regulated industries and critical infrastructure operators.

The supply side has moved to meet it. AWS committed 7.8 billion euros to its European Sovereign Cloud, a deliberately isolated infrastructure with its first region in Brandenburg, Germany, operating independently of every other AWS region worldwide. Microsoft has extended its sovereign cloud line with disconnected operations for Azure Local, letting an organisation run a private cloud whose control plane never leaves its own premises, and has committed that data processed by its AI services for EU customers stays inside the EU Data Boundary. Google Cloud has announced multi-billion-dollar European investments in Belgium and Germany. Between them, the hyperscalers have put well over 25 billion dollars behind the proposition that where your workload runs is now a product feature rather than an implementation detail.

None of that, on its own, would change how you buy automation. What changed is that the regulatory layer stopped talking about infrastructure and started talking about services, contracts and supply chains — the layer where automation actually lives.

Residency is not sovereignty, and the distinction is now legal

The single most useful thing a buyer can learn this year is that residency and sovereignty are different properties, and that satisfying one says nothing about the other. Residency is a question of geography: where the bytes physically sit. Sovereignty is a question of jurisdiction: which legal system can compel access to them. A European region operated by a United States company gives you residency without sovereignty, because the operator remains subject to United States law regardless of where the disk is.

Until recently this was an argument between consultants. On 3 June 2026 it became a legislative proposal. The European Commission published the Cloud and AI Development Act, known as CADA, as the centrepiece of a broader Tech Sovereignty Package aimed at two stated vulnerabilities: a structural shortfall in European data centre capacity, and dependence on a small number of non-EU cloud providers. Alongside a target of roughly tripling EU data centre capacity over five to seven years, the proposal defines four Union assurance levels for cloud and AI sovereignty, to be used by public sector bodies according to their own risk assessments.

CADA is a proposal, not yet law, and its assurance levels are aimed at public procurement rather than at every private buyer. Treating it as an immediate compliance obligation would be a misreading. Treating it as the vocabulary the market is about to adopt would not. Once public bodies start specifying a level in tenders, vendors will start advertising one, and private buyers will inherit a ready-made way to describe requirements they previously had to invent from scratch. Level 1 is roughly the honest description of what most "EU hosted" marketing means today. Level 3 is where a large number of currently popular tools would fail.

What actually changed in the last twelve months

Four developments matter for anyone buying automation, and they compound rather than overlap.

• The EU Data Act became applicable on 12 September 2025. Customers of in-scope data processing services now hold a statutory right to switch providers and to receive technical cooperation in porting their data out. Crucially, those rights and the provider's corresponding obligations must be set out in a written agreement, which turns them into something you can inspect before signing rather than discover during a migration.
• Switching charges are being withdrawn on a published timetable. Between January 2024 and January 2027, providers may charge for switching but only up to the direct costs they actually incur. From 12 January 2027, switching charges — including egress fees — are prohibited outright. Genuinely optional premium migration services can still be billed; the baseline cannot.
• NIS2 enforcement arrived. The directive's supply-chain provisions make an in-scope organisation accountable for the security of its direct suppliers, with obligations that have to be enforced contractually and flowed down to sub-contractors. 2026 is the year supervisory authorities began inspections, which turns sub-processor disclosure from a nice-to-have into an audit finding.
• CADA gave the market a shared vocabulary. Four levels, publicly defined, replacing a decade of vendor-specific sovereignty claims that could not be compared with one another.

A workflow is a supply chain, and that is the automation-specific problem

Sovereignty guidance is almost always written about applications and infrastructure: one system, one vendor, one region. Automation does not behave like that, and this is where buyers who have done everything correctly at the platform level still end up exposed.

Consider an ordinary workflow. A form submission arrives from one vendor, is enriched by a second, classified by a language model hosted by a third, written to a CRM operated by a fourth, notified into a chat tool run by a fifth, and archived to storage owned by a sixth. Every one of those hops is a transfer. The orchestration platform sitting in the middle also keeps a run history, which typically contains the full payload of each step — meaning the automation tool holds a copy of data from six systems in one place, often for months, under a retention policy nobody chose deliberately.

The consequence is that the sovereignty posture of your automation platform tells you very little about the sovereignty posture of the work it performs. A platform hosted in Frankfurt that calls a United States inference endpoint on every execution has moved your data to the United States on every execution. This is the same structural issue that makes exit planning hard, and it is worth reading alongside our guide to avoiding automation vendor lock-in, because the inventory you need for one is the inventory you need for the other.

The practical instrument here is boring and effective: a connector inventory. For each workflow, list every external system it touches, the vendor behind it, the region that vendor processes in, and whether that region is configurable. Most organisations discover two or three surprises in the first pass. The most common is an inference step nobody registered as a data transfer, because it was added as a feature rather than as a vendor.

What sovereignty costs, in numbers you can budget

One genuinely useful development of 2026 is that the price of residency stopped being a negotiation and started being a published number. For AI inference, four major vendors now publish an explicit, separable surcharge for pinning processing to a region or data zone, and they have converged on the same figure: roughly ten percent, a 1.1x multiplier on standard list pricing.

Two things follow from this table. The first is that ten percent on the inference line of an automation budget is, for most organisations, a rounding error next to the cost of the incident it prevents — which makes the common objection that sovereignty is expensive largely unfounded at this layer. The second is that the exception is instructive. Anthropic's first-party API does not currently offer a guaranteed EU-only processing region, so European teams that want Claude inside the EU route it through AWS Bedrock in Frankfurt, Ireland, Paris or Stockholm, or through Google Vertex AI in European regions — and on Bedrock's published price list there is no residency premium at all. The lesson is not that one vendor is cheaper. It is that residency pricing is a packaging decision, not a cost of physics, and it is therefore negotiable and subject to change.

The costs that are harder to see sit elsewhere: feature lag in regional deployments, smaller model selections, and support teams in a different time zone. Ask about all three. A ten percent surcharge on a model you cannot use for four months after general availability is not a ten percent decision.

Where the major automation platforms actually sit

Platform positions differ more than the marketing suggests, and the differences that matter to a buyer are usually about defaults and reversibility rather than about whether a European region exists at all.

Notice the pattern. The differentiator is rarely whether a vendor can host in Europe; almost all of them can. It is whether Europe is the default, whether the choice is reversible, and whether the guarantee extends past the platform to the connectors. A decision made irreversibly at signup is a decision worth making slowly. If you are still weighing managed hosting against running the platform yourself, our comparison of cloud versus self-hosted automation covers the operational trade-offs that sit underneath this table.

Five clauses worth putting in the contract

Sovereignty requirements that live only in a slide deck do not survive a vendor's next architecture change. The following belong in the agreement itself, and all five are reasonable asks that serious vendors answer without friction.

1. Named processing regions, including for sub-processors. Not "the EU" but the specific regions, for the platform and for every sub-processor in the chain. Add an obligation to notify you before a new sub-processor is added, with a right to object.
2. Run history and log residency, stated separately. Execution logs, error payloads and support diagnostics frequently follow different rules from production data. Specify where they live and how long they are retained, and remember that a support engineer pulling a failed execution is a data access event.
3. Exit mechanics with a tested format and a deadline. The Data Act gives you the right; the contract should give you the specifics. What format, covering which objects, delivered within how many days, and at what cost before January 2027, when the charge must go to zero anyway.
4. Change of control. Sovereignty commitments made by an independent European vendor mean something different after an acquisition by a non-European parent. Given how active software M&A has been this year, a clause that lets you exit without penalty if the sovereignty posture materially changes is no longer an exotic request. We have written separately about what to do when your automation vendor gets acquired.
5. Model and inference routing. If the platform ships AI features, pin down which provider and which region serve them, and require notice before that changes. This is the clause most often missing, and the one most likely to move without anyone telling you.

A due-diligence sequence that fits in one meeting

You do not need a formal programme to do this well. You need an ordered set of questions and the discipline to stop when the answers get vague.

1. Classify the data the workflow carries: regulated, commercially sensitive, or ordinary operational data.
2. Inventory every connector and inference call the workflow makes, and the vendor behind each one.
3. For each vendor, establish the processing region and the controlling legal entity — not the same question.
4. Ask where run history and support diagnostics are stored, and for how long.
5. Ask whether the region is fixed at provisioning or changeable afterwards.
6. Request the export format and a realistic time-to-export for a full account.
7. Price the residency options explicitly rather than assuming they are unaffordable.
8. Decide the assurance level you actually need for this data class, and buy to it — no higher.

Frequently asked questions

What is the difference between data residency and data sovereignty?

Residency is where your data physically sits; sovereignty is which government has legal authority over it. Storing data in a European region of a United States provider satisfies residency but leaves the provider subject to United States law, which is why the European Commission's Cloud and AI Development Act separates them into distinct assurance levels rather than treating a European region as sufficient.

What are the four sovereignty levels in the Cloud and AI Development Act?

The Commission's proposal, published on 3 June 2026, defines four Union assurance levels. Level 1 requires processing and storage on Union infrastructure. Level 2 adds demonstrated independence from third-country control. Level 3 adds European ownership, control and personnel criteria. Level 4 requires full transparency over the software supply chain with no third-country interference. The levels condition access to public sector contracts rather than applying to every buyer.

How much does data residency actually cost?

For AI inference the market has converged on roughly ten percent. OpenAI, Azure OpenAI, Google Vertex AI and Mistral each publish a separate, isolable price for pinning inference to a region or data zone, and all four work out to about a 1.1x multiplier on standard list pricing. AWS Bedrock is the notable exception: its published rates for Claude are the same in Frankfurt as in Virginia.

Does the EU Data Act help me leave an automation vendor?

Yes. Since 12 September 2025, customers of in-scope data processing services have a statutory right to switch providers and to receive technical cooperation for porting their data out. Switching charges, including egress fees, are being withdrawn in stages and are prohibited entirely from 12 January 2027. The rights and the provider's obligations must be set out in writing in the contract, which makes them something you can check before you sign.

Which automation platforms offer European data residency?

Make processes through European infrastructure by default and lets you pick a region when the organisation is created, though the choice cannot be changed later without a migration. n8n Cloud runs in Azure Germany West Central in Frankfurt, and the fair-code licence means you can self-host anywhere. Zapier hosts in United States AWS regions by default with European residency available on Enterprise plans and explicit configuration. Power Automate follows the tenant geography and the EU Data Boundary.

Why does sovereignty matter more for automation than for ordinary software?

Because a single workflow is a supply chain. One automation may touch six vendors, an inference provider and several sub-processors in one execution, and the run history stores the payloads it moved. The sovereignty posture of the orchestrator tells you almost nothing about the sovereignty posture of the run as a whole, which is why the connector inventory matters more than the platform's marketing page.

Is buying a sovereign stack always the right call?

No. Sovereignty is a control that costs money, narrows your vendor choice and often reduces feature availability, so it should be applied to the data flows that genuinely need it rather than to the whole estate. The practical approach is to classify workflows by the data they carry, apply strict requirements to the small share that handle regulated or sensitive records, and leave the rest on standard infrastructure.

What should I ask a vendor before signing?

Ask for the full sub-processor list with locations, the legal entity that controls each one, where run history and logs are stored and for how long, whether the region can be changed after provisioning, what the export format is and how long a full export takes, and what happens to your data if the vendor is acquired. Vague answers to any of these are the finding, not the follow-up question.