AI Insights · Risk & Governance
Who Controls Your AI? The Data Sovereignty Risk Every Business Should Understand
In June 2026, Anthropic suspended access to two of its most capable AI models within three days of launch, following a US government export control directive. Businesses outside the United States that had integrated these tools had no warning. This is not a hypothetical risk. It happened. Here is what it means for how you build with AI.
What Happened in June 2026
On 12 June 2026, Anthropic suspended access to Claude Fable 5 and Mythos 5 — models it had released just three days earlier — in response to an export control directive from the US government. The directive prohibited use of the models by anyone who is not a US national.
Mythos 5 is Anthropic's most powerful frontier model. When it was first announced in April 2026, the company acknowledged it was too capable in certain sensitive areas to release publicly without restriction. It had initially been made available only to a small group of organisations — predominantly US technology companies — to help identify and patch vulnerabilities in critical digital infrastructure.
For businesses that had accessed the models during the brief window they were available, the suspension was abrupt. No migration path. No transition period. No alternative provided. Access: off.
The lesson here is not that Anthropic behaved badly. They were responding to a government directive they had no choice but to comply with. The lesson is that when your AI capability sits entirely within a single provider's infrastructure, in a jurisdiction governed by another country's law, your continuity of access is not under your control.
The Four Categories of AI Sovereignty Risk
The June 2026 incident is the most visible example of AI access risk, but it is one of four distinct categories every business should understand.
Risk 01
Geopolitical Access Risk
A government — your own or the provider's — can restrict, suspend, or revoke access to an AI model at any time. Export controls, national security directives, and trade policy can all affect which AI tools are legally accessible from your location.
Risk 02
Provider Continuity Risk
AI providers deprecate models. Pricing changes overnight. A provider can alter its terms of service, limit API access, pivot its product strategy, or simply cease operating. If your workflows are built on a specific model, any of these events creates an operational problem.
Risk 03
Data Training Risk
Many AI services — particularly consumer and small business tiers — use your inputs to improve their models. If your team is submitting client data, commercially sensitive information, or intellectual property to an AI tool on a standard plan, that data may be retained and used in ways your contract does not clearly prohibit.
Risk 04
Jurisdiction and Residency Risk
The country in which your data is processed is not always the country in which the servers are located. A UK business using a US-owned AI platform may find that its data is subject to US legal jurisdiction, regardless of where the servers sit. This has direct implications for GDPR compliance and client confidentiality obligations.
Why Single-Provider Dependency Is the Core Problem
In any other area of technology, dependency on a single supplier without a fallback is considered poor architecture. A company that ran its entire operation on one email server with no backup would be considered negligent. The same logic applies to AI — but many businesses are building AI workflows with exactly this level of dependency.
The issue is compounded by how AI integration is typically sold. Vendors make it easy to embed their tools directly into your processes. Copilot sits inside your Microsoft 365 environment. ChatGPT plugins connect to your workflows. Each integration creates a dependency that is painful to remove if the underlying service changes.
This is not an argument against using AI. It is an argument for building on a platform layer — something that sits between your core business data and the AI models you access — so that the model can change without disrupting the workflow underneath it.
The analogy: think of it as the difference between buying fuel from one specific petrol station that only you can access versus owning a car that can run on multiple fuel types from any supplier. The capability (driving) remains consistent even if one supplier becomes unavailable.
What "Your Data" Actually Means in an AI Context
When we talk about data sovereignty in AI, we mean several distinct things that businesses often conflate.
Input data: the prompts, documents, and queries you submit to an AI model. On many plans, this is retained and may be reviewed by the provider for safety and quality purposes. On enterprise plans with appropriate contractual terms, retention and use of input data is typically restricted.
Output data: the AI's responses. These are generated by the model and do not, in themselves, expose your input data to third parties — but they may carry the provider's intellectual property considerations depending on your plan and jurisdiction.
Embedded data: if you connect an AI to your CRM, finance system, or document store via an integration or agent, the AI may access and process significantly more of your business data than you intended. Understanding what data an AI agent can reach — and what it does with that access — is a governance question, not just a technical one.
Training data: whether your inputs are used to improve the underlying model. Most enterprise contracts explicitly exclude this. Most consumer and SME plans do not.
A practical check: look at the AI tools your team is using today. For each one, identify which tier of plan you are on, whether your data is contractually excluded from training, and where the data is processed. Many businesses discover they are on consumer or growth plans — not enterprise plans — for tools their employees use with client data every day.
The Geopolitical Dimension Is Not Going Away
The June 2026 suspension was not an isolated incident. It reflects a broader and accelerating pattern: governments are treating advanced AI as a strategic asset, subject to the same export control frameworks applied to defence technology and semiconductor equipment.
The US Bureau of Industry and Security has progressively tightened controls on AI-related exports since 2022. The EU AI Act introduces regulatory requirements that affect how AI can be deployed in European businesses. The UK is developing its own regulatory framework. China has introduced requirements that affect how AI providers operating in China handle data and model access.
For a business in the UK — or anywhere outside the United States — relying on a US-headquartered AI provider for a core operational capability means accepting that a US government decision could affect your access to that capability. That is a geopolitical risk, and it belongs in your risk register in the same way that single-supplier risk or regulatory compliance risk does.
Building with Sovereignty in Mind: A Practical Framework
The goal is not to avoid AI. It is to use AI in a way that keeps your data under appropriate control and your workflows resilient to supplier changes. The following practical steps apply regardless of the AI tools you currently use.
- Audit your AI tool inventory. List every AI tool in use across your business, including tools employees are using informally. Identify the plan tier, the provider's jurisdiction, and whether your data is excluded from training under your current contract.
- Separate your platform from your model. Your core business data — CRM records, finance data, documents — should sit on a platform you control, not inside an AI provider's environment. AI models should connect to your data, not hold it.
- Review what data your AI agents can access. If you are using AI agents that connect to your systems via integrations, map exactly what data those agents can reach. Limit access to what is genuinely necessary for each task.
- Have a fallback for critical workflows. Any AI-assisted workflow that is genuinely critical to your operations should have a non-AI fallback. This is not defeatism — it is the same resilience planning you would apply to any third-party service dependency.
- Choose providers with demonstrable privacy commitments. Not all AI is equal on data privacy. Providers who have made verifiable, contractual commitments about data handling — and who have independent certifications to back those commitments up — offer materially different risk profiles from those who rely on general terms of service.
Why Zoho's architecture is relevant here: Zoho's platform is built around the principle that your data belongs to you. It does not use customer data to train its AI models. Data residency is configurable. And because Zoho's AI (Zia) operates within the Zoho ecosystem rather than relying on a single external model provider, it is insulated from the kind of overnight access suspension that affected Anthropic's models in June 2026. Read more on the Zoho Privacy-First AI page.
The Responsible Position
We use Claude, Gemini, and other external AI models in our own work at 1 Cloud Consultants. We are not suggesting these tools are unsafe. We are saying that the organisations building on top of them — including us — need to understand what they are taking on when they do so, and to build accordingly.
The June 2026 suspension affected access to frontier models with particular sensitivity around cybersecurity capabilities. Most businesses were not using those specific models for daily operations. But the mechanism by which access was suspended is the same mechanism that applies to any AI capability sitting in a provider's infrastructure under another country's jurisdiction.
Understanding that mechanism, and building AI into your business in a way that accounts for it, is not paranoia. It is ordinary due diligence — the same kind of thinking you apply to every other supplier relationship that your operations depend on.
Frequently Asked Questions
Want to review how AI is currently integrated into your Zoho environment and what your data exposure looks like?