Picture a junior analyst late on a Thursday. The data room is open in one tab and a consumer chatbot in another. A few pastes later, the first screen of a confidential deal is done, and nobody at the fund can say where that data went, how long it will be kept, or which country's law now governs it.

When we shared what we learned from 50 investment teams across Europe, that scene was one of the most common. We have since passed 100 conversations, and it has not changed. Ungoverned AI is the de facto standard in private capital. At the same time, the COOs and partners we meet tell us, in different words, that European data sovereignty has moved from a preference to a purchasing criterion.

This article is about the space between those two realities: the regulation that frames it, the infrastructure that makes it hard to close, and why we built Clarenza around it from the first line of code.

Six years in the making

The EU AI Act is often described as sudden. It was not. It began with the Commission's 2018 AI strategy and an expert group whose 2019 Ethics Guidelines for Trustworthy AI set out principles that still run through the law: human oversight, transparency, data governance and accountability. A 2020 White Paper proposed the risk-based logic, and the Commission tabled the Act in April 2021. Then ChatGPT arrived, legislators rewrote the draft to cover general-purpose models, and political agreement followed in December 2023. The Act entered into force on 1 August 2024.

Since then it has been phased in. Prohibited practices and the AI literacy duty have applied since February 2025, obligations for general-purpose model providers since August 2025, and transparency rules since August 2026. This year's Digital Omnibus moved obligations for stand-alone high-risk systems to December 2027, which legal commentators have consistently read as a deferral rather than a retreat. Alongside it runs DORA, the Digital Operational Resilience Act, which has applied since January 2025 to almost every regulated financial entity in the EU, authorised fund managers included. Most deal screening tools will never be classed as high-risk AI, but every fund that uses AI is a deployer under one law and an ICT customer under the other, and both roles carry obligations today.

Two continents, two philosophies

The United States chose the opposite path. There is no federal AI law. The 2023 executive order on AI safety was revoked in January 2025, and Washington has since worked to preempt state rules rather than write its own. Colorado's pioneering AI Act was repealed and replaced in May 2026 with a narrower law, and the SEC withdrew its proposed rule on predictive analytics for advisers in June 2025.

For a European fund, though, the decisive difference is not AI rules. It is data jurisdiction. The CLOUD Act lets US authorities compel US companies to hand over data in their control wherever it is stored, and FISA Section 702 permits surveillance of non-US persons through US providers. The EU's top court has struck down two transatlantic data frameworks on these grounds, and the current one is under appeal. When the French Senate asked Microsoft France in June 2025 whether it could guarantee that French data would never reach US authorities without French consent, the answer came under oath.

‍

"No, I cannot guarantee it."
,Anton Carniaux, Director of Public and Legal Affairs, Microsoft France, before the French Senate, 10 June 2025

‍

We think that answer is honest, and it captures the point. A server in Frankfurt run by a company under US law gives you data residency. Sovereignty means only parties under European law have legal, operational and technical control.

When your edge leaves the building

To be fair to the leading US labs, their enterprise offerings have matured. OpenAI offers European data residency with zero retention for eligible API customers, and the major labs do not train on enterprise data by default. For many workloads, that is enough.

Confidential deal work is not most workloads. Residency does not remove CLOUD Act exposure. Retention defaults and feature exceptions vary, and in 2025 a US court ordered OpenAI to preserve user output logs, deleted chats included, as part of the New York Times litigation. A deletion promise is only as strong as the legal system behind it. Data rooms are bound by NDAs and may hold material non-public information. And a fund's real edge is not its documents but its thesis, screening logic and committee reasoning, which one participant in our research described as an "uncontrolled transfer of that edge" whenever it flows through external models.

The larger exposure is that most of this traffic is not sanctioned at all. MIT's 2025 GenAI Divide study found that only 40% of companies had bought an official LLM subscription, while workers at more than 90% of them used personal AI tools for work. Gartner found that 69% of organisations suspect or know that staff use prohibited public AI tools. Our conversations matched both numbers: analysts building first screens in consumer chatbots, and at one large buyout platform, nobody using the firm's sanctioned enterprise assistant for investment documents at all. Shadow AI is a demand signal. A ban does not close it; a better, governed tool does.

DORA readiness, in plain terms

DORA rests on five pillars: ICT risk management, incident reporting, resilience testing, third-party risk and information sharing. The one most funds underestimate is the fourth, because every AI tool a fund adopts is an ICT third-party service. An LLM API, a chatbot subscription and a diligence platform all belong on the register of information that firms first reported to supervisors in April 2025.

In practice, readiness comes down to a handful of questions a COO should be able to answer for every AI tool. Was it assessed before signing, including for concentration risk? Does the contract name the countries where data is processed and stored? Are there audit and access rights, and a tested exit plan with a transition period? Is the subcontracting chain known? A vendor reselling a US model on a US hyperscaler is a chain of at least three. A personal chatbot account processing deal data fails every one of these questions by default.

Supervisors are watching the concentration question closely. In November 2025, the European Supervisory Authorities designated the first 19 critical ICT third-party providers, including AWS, Google Cloud and Microsoft, using data from the very registers funds had submitted. Most AI tools on the market were not built for any of this. They give an analyst a chat window, but not a COO a register entry, a defined data location, an audit trail or an exit path.

‍

The hidden cost of going it alone

Faced with those gaps, some firms decide to build their own stack: self-hosted open models, retrieval over the data room, logging, access controls, evaluations and an internal interface. It is a legitimate choice and rarely a cheap one. Practitioner breakdowns put GPU hardware for a single large-model deployment at $120,000 to $240,000 before hosting and redundancy, and the ongoing cost of an enterprise in-house LLM platform at $760,000 to $2 million a year, most of it people. Then the stack keeps moving. Models are superseded every few months, and every upgrade to drivers, frameworks or guardrails has to be retested and documented again for DORA and the AI Act.

And that is only the infrastructure. It does not yet include what makes AI useful in an investment process: extraction that reconciles messy financials, deterministic calculations, traceability to the page and screening against a fund's own thesis. For a large bank, these costs are a rounding error. For a mid-market fund without ML engineers, they are prohibitive, which is exactly why shadow AI fills the gap.

Sovereign AI has been a big-player privilege

The European market is moving fast toward sovereign, confidential AI, and the clearest validation comes from its most regulated institutions. Mistral AI has become the reference partner for European finance. BNP Paribas renewed its group-wide partnership for three years in May 2026, HSBC runs Mistral's models on a self-hosted basis, and ABN AMRO signed in August 2026 with the explicit aim of reducing reliance on AI providers outside Europe. Underneath, Mistral is building its own compute, from an 18,000-chip data centre south of Paris to a planned facility in Sweden.

Aleph Alpha has taken a complementary route. On 3 October 2026 it released Kolibri, an open-weights model built in Germany for deployment on infrastructure its users control. Its technical report documents measures for copyright, data protection and AI Act requirements, and in retrieval tasks it is trained to abstain when the supplied documents lack sufficient evidence, a design choice any analyst will appreciate.

These efforts sit next to ours rather than against it. Mistral and Aleph Alpha build models and infrastructure; Clarenza builds the investment workflow on top. Together they confirm that in the world's most regulated economic area, secure and compliant AI is a market in its own right. They also expose the gap. A multi-year bank partnership assumes an AI team, a procurement department and data centres. Open weights still need someone to host, secure, evaluate and maintain them. For the lean mid-market fund, the family office or the five-person venture team, sovereign AI has stayed out of reach.

Designed for Compliance from day one

We did not add compliance to Clarenza after the fact. The governance findings from our research shaped the architecture before the first agent was built.

Confidential deal data is processed only by EU-sovereign models on European infrastructure with zero data retention. It never touches a US LLM or US cloud provider, which keeps it outside the reach of the CLOUD Act and FISA 702. Where teams want richer public market context, they can choose to connect frontier models, but never for a deal's private data. Every figure links to its exact location in the source, the agents' reasoning is visible end to end, and financial calculations run on fixed solvers rather than a language model, so the analyst stays the decision-maker, as the AI Act's principles of transparency and human oversight intend. Each fund's memory lives in a tenant-isolated environment that never trains anyone's model. Clarenza produces audit-ready reporting for DORA and deploys as a managed EU cloud, a dedicated environment or on-premise, so a fund can match the architecture to its risk tolerance and keep a credible exit plan. Our SOC 2, ISO 27001 and ISO 42001 programmes are under way.

The economics are the other half of the answer. Clarenza runs the sovereign infrastructure layer once and shares it across the teams that use it, while each fund keeps its own isolated environment, thesis and memory. A mid-market fund gets the class of governed AI that today sits mostly inside Europe's largest banks, without hiring a platform team to run it. That is also how we think shadow AI actually ends: with a tool analysts prefer to the consumer chatbot, and one their COO can sign off on.

What this means for your fund

The high-risk deadline has moved, but AI literacy, transparency and DORA apply now. Start by mapping which tools already touch deal data. Put every AI tool through the same third-party scrutiny as any other critical vendor, and draw a firm line between public research and confidential data. Above all, be honest about build versus buy: if you cannot staff a platform team for years, do not start building one.

Sources

1. Gibson Dunn: EU AI Act Omnibus agreement and postponed high-risk deadlines
2. Cloud Security Alliance: High-risk deadline deferred, not cancelled
3. Goodwin: What DORA means for fund managers
4. DORA Article 30: key contractual provisions
5. ESAs designate critical ICT third-party providers under DORA
6. CASRAI: The US federal AI preemption fight
7. The Register: Microsoft admits it cannot guarantee data sovereignty
8. DataGuidance: CJEU appeal on the EU-US Data Privacy Framework
9. OpenAI: Introducing data residency in Europe
10. MIT NANDA and Gartner shadow AI findings, summarised by WorkJam
11. Thinking of hosting your own enterprise LLM? It will cost you
12. BNP Paribas and Mistral AI extend their partnership
13. FinTech Global: ABN AMRO taps Mistral
14. Aleph Alpha releases Kolibri

This article is for general information and does not constitute legal advice. Regulatory timelines reflect the position as of October 2026.