Seven Technical Decisions That Make or Break AI Sovereignty
Salesforce bought Slack for $27.7B in 2021, inherited 750 million daily conversations, and gained the contractual right to train AI on all of it. Because the data was now "Salesforce data." No new consent required. No notice to the organizations that had been planning strategy, closing deals, and managing their teams on Slack for years. One acquisition — a purely financial transaction between two companies — transferred AI training rights to three-quarters of a billion daily conversations without the people in those conversations making a single decision about it.
That is what sovereignty lost looks like. Not a breach. Not a hack. A business transaction. And it happened because the organizations using Slack had never answered seven questions that determine whether AI sovereignty holds.
What the Seven Non-Negotiables Actually Are
The Sovereign Intelligence Architecture standard defines seven requirements for any AI deployment to qualify as sovereign. They are binary. Each one either passes or fails. There is no partial credit.
Data Residency means data never leaves the organization's infrastructure — not just at rest, but during inference, during retrieval, during logging. "European servers" in a contract means nothing if queries are processed in Virginia.
Model Sovereignty means the organization uses open-weight models it can inspect, run, and reproduce independently. If the model provider changes their API, deprecates the version, or gets acquired, a sovereign organization can switch models without rebuilding its architecture.
Vendor Independence means no lock-in. The organization can exit any vendor relationship without losing its data, its processes, or its AI capability. Single-vendor dependency creates exactly the risk that materialized in the Salesforce-Slack case.
Audit Completeness means every AI interaction is logged — who asked, which model answered, what data was accessed, what was produced — with logs the organization owns and controls, not logs held by a vendor on vendor infrastructure.
Hybrid Intelligence means the organization has a Router: a system that classifies every AI request by sensitivity before deciding where to send it. Non-sensitive queries can use cloud models safely and efficiently. Sensitive queries — those involving strategy, client data, personnel, or intellectual property — go to sovereign infrastructure and never leave it.
Governance by Design means compliance is built into the architecture, not added afterward through policy documents. A policy that says "don't paste client data into ChatGPT" doesn't work when AI is embedded in the tools your teams already use. Architecture that physically prevents sensitive data from reaching non-sovereign endpoints does.
LLM Agnosticism means the architecture is designed so the model is replaceable. Today it's Llama. Next year it might be Mistral, or something that doesn't exist yet. An LLM-agnostic architecture treats model choice as an operational variable, not a structural commitment. The architecture is the asset. The model is the component.
Why Partial Compliance Creates Full Exposure
An organization that passes six of seven non-negotiables is still exposed on the seventh. Every failed non-negotiable opens a compliance door a regulator can walk through.
Consider Audit Completeness alone. HIPAA — the US law governing healthcare data — requires organizations to show exactly what happened to patient information and when. The EU AI Act, which begins enforcement for high-risk AI categories in 2026, requires documentation of AI decision-making. Penalties reach €35M or 7% of global revenue — whichever is higher. If an organization cannot produce a complete AI audit trail, it cannot satisfy either framework, regardless of how well it has handled the other six requirements.
Or consider Vendor Independence failing on its own. TikTok received a €530M fine from Ireland's Data Protection Commission in May 2025 — the largest single data protection penalty of 2025 — specifically for data transfers without equivalent protection. The mechanism was vendor dependency: TikTok's architecture required data to flow to infrastructure outside declared protections, and TikTok couldn't stop it because the infrastructure decisions weren't theirs to control.
The seven non-negotiables are interdependent. Without Data Residency, Audit Completeness becomes harder to achieve — you can't fully audit a system you don't control. Without Vendor Independence, Governance by Design is compromised — the architecture belongs to the vendor, not to you. Without LLM Agnosticism, Model Sovereignty becomes fragile — the model is sovereign but the architecture requires it.
The Compliance Story That Breaks Under Audit
Most organizations that believe they have AI governance have documented two things: where their data is stored, and what their AI policy says. These are the wrong two things.
Regulators auditing for AI sovereignty look at the actual data flow — where inference happens, where embeddings are indexed, where logs are retained, who can access those logs. A contract that says "European servers" while routing inference to available GPU capacity in Virginia is what Ireland's Data Protection Commission called the basis for a €530M fine. A governance policy document that says "handle data responsibly" while 89% of enterprise AI usage leaves no audit trail — according to LayerX's 2025 research — is what EU AI Act auditors will classify as architectural non-compliance.
The seven non-negotiables replace documentation with architecture. Governance by Design means the architecture physically cannot route sensitive data to non-sovereign endpoints, not that a policy says it shouldn't.
The Framework Vendors Don't Want You To Apply
Apply the seven non-negotiables as a checklist to the AI products your organization currently uses. For each product, answer seven binary questions:
Does your data stay within your infrastructure during inference, not just storage? Does the model run on open weights you can inspect? Can you exit the vendor without losing your AI capability? Do you own the audit logs? Does a Router classify queries before routing them? Is compliance enforced by architecture rather than policy? Can you replace the model without rebuilding the architecture?
Most enterprise AI products fail at minimum three — and they fail by design, not by accident. Audit Completeness logs are held by the vendor because those logs have commercial value. Data Residency isn't guaranteed because routing inference to available GPU capacity is cheaper. Vendor Independence is structurally undermined because lock-in increases retention.
The pattern that describes this: dependency theater. Contracts that specify data center locations. Privacy policies that describe responsible data handling. Governance documentation that sits in a binder. Every layer creates the appearance of control while the architecture behaves differently.
What the SIA Standard Requires Architecturally
The Sovereign Intelligence Architecture methodology treats all seven non-negotiables as binary requirements, not aspirational targets. An organization either has Data Residency or it doesn't. Either every inference is logged or some aren't. Either it can swap models without rebuilding or it can't.
The SIA standard's four core components address the non-negotiables directly:
The Router — a system that classifies every AI request by sensitivity before routing it — is the technical implementation of Hybrid Intelligence. Think of it as a mail room that reads the sensitivity label on every envelope before choosing which courier to use. The Router also enables Governance by Design, because classification and routing rules are architectural, not optional.
The Vault — on-premise knowledge storage for the organization's documents, data, and institutional knowledge — implements Data Residency at the knowledge layer. Documents indexed in the Vault are used by the AI without ever leaving the organization's perimeter. Your competitive intelligence stays yours.
The Recorder — immutable logging of every AI interaction — implements Audit Completeness. The organization owns these logs. They are not held by a vendor, not processed through a third-party analytics platform, not subject to a vendor's retention policy. When a regulator asks what the AI did with specific data on a specific date, the Recorder provides the answer.
The Firewall — egress control that prevents AI models from sending data outward — makes Data Residency hold even when a model is configured incorrectly or attempts to reach external endpoints. Your data doesn't leave your perimeter unless you explicitly permit it.
Together, these four components implement all seven non-negotiables simultaneously — not as separate features but as interlocking architectural requirements.
The Question Regulators Will Ask
Work backwards from this: "Provide a complete audit trail of every AI interaction involving personal data for the last 24 months."
If your current AI setup cannot produce that trail — with logs your organization controls, timestamped, tied to specific users and specific data — you have failed Audit Completeness. Not technically. Architecturally.
EU AI Act Article 26 makes the deploying organization responsible for compliance, not the company that built the model. If your AI system is used in a high-risk category — healthcare, legal, financial services, employment decisions — and it fails the Act's requirements, the fine comes to you. The model provider's terms of service do not transfer that liability.
The FTC has already cited inadequate AI data governance — specifically the absence of audit trails and documented data controls — as grounds for enforcement action. €7.9B in cumulative data protection fines worldwide since 2018. The trajectory is established and accelerating.
The Competitive Dimension
Organizations that implement the seven non-negotiables now gain a procurement advantage that is already visible in regulated markets. European enterprise contracts are including sovereignty requirements in RFPs: "Can you prove your AI never processed our data on US infrastructure?" The organizations that can answer yes — with documentation, with logs, with a named architectural standard — are closing contracts that organizations relying on dependency theater are losing.
EU AI Act enforcement for high-risk categories begins in 2026. Organizations building sovereign architecture now have 12-18 months to build it correctly, document it thoroughly, and demonstrate it under audit. Organizations that wait will be remediating under regulatory pressure, which costs more and produces worse architecture.
The seven non-negotiables are not an advanced option for mature organizations. They are the baseline. Any AI deployment that fails even one of them is not "good enough for now." Every day a non-negotiable goes unaddressed, more AI interactions accumulate outside the architecture that was supposed to govern them — building a compliance liability that grows with every prompt.
The most valuable thing a CTO can do this quarter is answer seven binary questions about every AI system their organization currently runs. The answers won't be comfortable. The organizations that find the gaps now and close them architecturally will be ready when regulators ask.
The Seven Non-Negotiables define what sovereignty actually means. Anything less is dependency theater.