Seven TSI Principles. Seven Non-Negotiable Requirements. Here's How We Meet All of Them.
Stralevo's TSI Compliance Architecture, Explained for the CFO Who Has to Answer for It
---
A European Banking Authority (EBA) examiner walks into your quarterly audit and asks one question: "Which TSI compliance principles does your financial AI vendor satisfy?" Most CFOs cannot answer — not because they deployed AI carelessly — their vendor never told them. TSI stands for The Sovereign Institute, an independent certification body that defines seven technical requirements for sovereign AI. Here is the complete answer for Stralevo: seven principles, seven requirements, all met by architecture, not by policy document.
"Ask your AI vendor which TSI principles they meet. If they answer with marketing language, that is your answer — and it is no."
---
Why TSI Compliance Matters More Than "Sovereign AI"
"Sovereign AI" has been used to describe anything from GDPR-compliant data storage to full on-premise deployments — which makes the term nearly useless for a CFO evaluating vendors. "TSI-compliant" means something precise: seven named principles, each with a technical implementation requirement, each independently verifiable.
Calling a European data center "sovereign AI" is a geography claim, not an architecture claim. True TSI sovereignty requires seven specific conditions — met simultaneously, not partially. A system that meets six of the seven principles and allows training on client data fails TSI compliance entirely, and gives its users the false confidence of assumed protection on the five or six principles it does satisfy.
The seven principles defined by The Sovereign Institute — Data Residency, Model Sovereignty, Vendor Independence, Audit Completeness, Hybrid Intelligence, Governance by Design, and LLM Agnosticism — are not AI ethics guidelines. They are regulatory protection mechanisms. A CFO who can map every principle to a documented architecture decision closes an EBA examination in an afternoon. One who cannot opens a multi-week remediation process.
---
Principle 1: Data Residency — Your Data Never Leaves EU Legal Jurisdiction
SAP, Sage, and Cegid all have AI features that route data through external cloud services — often outside EU jurisdiction — for processing. "EU-hosted" is not the same as "EU-jurisdiction." A US-headquartered vendor with servers in Frankfurt still subjects your financial data to the CLOUD Act, which allows US law enforcement to compel data production from US companies regardless of where the data is stored.
Stralevo runs on EU-sovereign infrastructure with no US parent company, no US acquisition or investment that transfers jurisdictional exposure, and no routing to US-domiciled infrastructure for any processing. Every financial query is processed within EU legal jurisdiction. That is the complete answer to Principle 1 — one that most vendors cannot give because they either have US infrastructure in the processing chain or a US parent company that creates indirect jurisdictional exposure.
---
Principle 2: Model Sovereignty — The AI Can Be Replaced Without Redeployment
Account for the scenario: your AI vendor's model provider raises prices, introduces data use policies you cannot accept, or is acquired by a company in an unfavorable jurisdiction. If your financial AI is architecturally coupled to one model provider, the vendor's answer to that scenario is: wait for our timeline, accept the new terms, or rebuild from scratch.
Our model-agnostic orchestration layer allows hot-swapping any LLM (large language model — the AI engine that processes your financial queries) — OpenAI, Anthropic, EU-hosted models, sovereign private models — in seconds, without changing workflows, controls, or integrations. The model is a component, not the architecture. Your financial intelligence capability does not depend on any single provider's terms of service remaining acceptable.
---
Principle 3: Vendor Independence — Your Data Is Never Used to Train Their Model
Samsung engineers pasted chip design source code into ChatGPT in March 2023 — within weeks of the company authorizing AI use. The same dynamic plays out in finance: accounting teams paste margin data and client financials into AI tools for "quick analysis." BMW, Société Générale, and Deutsche Telekom have each introduced policies explicitly banning unauthorized AI tool use — because the pattern of financial data flowing into training datasets is systematic, not exceptional.
Once financial data enters a vendor's training pipeline, nothing removes it. Deleting your account does not delete the learned patterns derived from your margin data, your supplier pricing, your cash flow queries. Principle 3 is the only real protection: a contractual and architectural guarantee that your data was never used for training in the first place.
We carry a documented, contractually binding policy: customer financial data is never used to train AI models, improve vendor models, or generate training datasets for any purpose. That policy is not a marketing claim — it is a documented obligation in the service agreement, and it is architecturally enforced by the fact that all model training occurs independently of customer data environments.
---
Principle 4: Audit Completeness — Every Query Is Logged With Source Documentation
Ask your current AI vendor to generate a log of every financial query your team made in the last 90 days, with the source document for each answer. Most cannot produce it — not because the data does not exist. It exists in their format, on their systems, accessible through their processes, on their timeline.
Every Stralevo session generates a complete audit log automatically — every query, every answer, every source document referenced, every timestamp — stored on infrastructure you control, accessible through your audit procedures, available on demand. When France's General Directorate of Public Finance (DGFiP) or an internal auditor asks for data lineage on a specific financial figure, your team produces it from your own systems in minutes.
Under DORA (Digital Operational Resilience Act, mandatory January 2025), financial institutions must document information and communications technology (ICT) third-party risk assessments. Stralevo's TSI compliance report — generated automatically from the audit log — satisfies that documentation requirement in a single step. Organizations that deployed non-compliant AI in 2024 and are now building their DORA documentation file are discovering that reconstruction is expensive; generation at the time of each inference costs nothing.
---
Principle 5: Hybrid Intelligence — You Can Export Everything, Anytime
Data lock-in is an overlooked risk in financial AI: the longer you use a vendor, the more of your financial data they hold, and the harder it becomes to switch. Principle 5 requires that you can export your complete data at any time, in standard formats, without vendor intervention — so that your financial intelligence capability belongs to you, not to the vendor relationship.
Complete data export is available on demand — all documents, all extracted data, all query histories, all generated reports — in open, portable formats that any accounting system can read. No data is held captive by proprietary structures. If your organization changes direction, merges, or simply wants to move to a different platform, every piece of financial intelligence you built with Stralevo travels with you. The decision to continue or transition belongs entirely to your organization.
---
Principles 6 and 7: Governance by Design and LLM Agnosticism
Governance by Design means AI governance is built into the architecture, not added as a policy layer after deployment. Every Stralevo deployment includes audit trails, access controls, data residency enforcement, and compliance reporting as default infrastructure — not optional features. A CFO who deploys Stralevo does not need to add a governance layer after the fact; it is present from the first transaction.
LLM Agnosticism means the system runs on any qualified large language model without rewiring workflows or compliance controls. As the open-source model market evolves — and EU-sovereign foundation models continue to develop — Stralevo's architecture adopts them without disruption. No regulatory uncertainty about a particular model provider's compliance posture becomes an operational constraint.
---
The Real Question Behind the Principles
Three requirements connect into one clear action: EBA Guidelines on Outsourcing (EBA/GL/2019/02) require documented third-party AI risk assessments. DORA requires financial institutions to assess ICT vendor concentration risk. TSI Principles 4 and 2 directly satisfy both requirements in a single architecture decision — and every organization using AI in financial workflows needs to be able to produce that documentation on request.
Only 15% of financial AI vendors satisfy all seven TSI principles when assessed through a formal vendor review — yet most market themselves as sovereign. Nearly all fail on Principle 3 (no training on customer data) and Principle 4 (complete audit trail) — the two that most directly create personal liability for the CFO who signed off on the selection.
Flip the standard compliance question: instead of "does our AI vendor comply with TSI?" ask "could we prove it to an EBA examiner by 10am tomorrow?" CFOs who can produce the compliance map in minutes are in a different regulatory position than those who need three weeks and two law firms to construct a response.
---
The Trade-Off That No Longer Exists
Speed versus control was the real tension in financial AI adoption: CFOs wanted fast, intelligent analysis, and their compliance teams wanted zero regulatory exposure. Before TSI-compliant financial AI existed, accepting one meant conceding the other.
That trade-off is gone. Seven principles. All met by architecture. An answer in 8 seconds, sourced and logged, on infrastructure you control. The CFO who walks into any examination — EBA, France's data protection authority (CNIL), DGFiP, or the banking and insurance supervisor ACPR — with a complete TSI compliance file is not navigating a compliance conversation. They are ending one.
---
Stralevo is TSI-certified across all seven principles: Data Residency, Model Sovereignty, Vendor Independence, Audit Completeness, Hybrid Intelligence, Governance by Design, and LLM Agnosticism. Compliance documentation is available on request and generated automatically by the system for every deployment.