Use AI on fund material with controls you run and evidence you retain.

The dilemma, stated plainly

Sending a prompt to a hosted foreign model may disclose confidential and personal data to a processor or other recipient outside the firm's environment. It can engage professional-secrecy duties, GDPR requirements, third-country transfer rules where applicable, and MAR controls over inside information. The analysis depends on the service and data flow.

EU hosting does not settle every issue. Residency concerns location; sovereignty concerns legal and operational control. A provider subject to US jurisdiction may face valid CLOUD Act process for data in its possession, custody or control wherever stored. US ownership alone is not conclusive.

If you are an EU-authorised AIFM (in Greece, generally a ΔΟΕΕ supervised by the Hellenic Capital Market Commission), these assessments remain yours. A vendor can support them, not assume them.

The regimes that bind you

AI interacts with material governed by several existing regimes. The practical questions are consistent: who used what material, where did it go, what controls applied and who approved the result?

• AIFMD / AIFMD II. AIFMD's delegation framework, including Article 20, does not transfer an AIFM's responsibility. AIFMD II (Directive (EU) 2024/927) entered into force on 15 April 2024; Member States' transposition deadline was 16 April 2026, and specified reporting amendments apply from 16 April 2027.
• DORA. DORA governs ICT third-party risk for in-scope AIFMs: assessment, contracts, monitoring, concentration-risk evaluation and, for ICT services supporting critical or important functions, documented and tested exit arrangements.
• GDPR. Investor and other personal data require a lawful basis, appropriate safeguards and risk-based security. Transfers to third countries require compliance with Chapter V where a transfer occurs; supplementary measures may be necessary. Avoiding unnecessary cleartext disclosure can reduce, not remove, obligations.
• MiFID II / MiFIR, MAR, AML, SFDR. These frameworks can affect AI-assisted material through record-keeping and conduct requirements, controls over inside information, AML/CFT confidentiality and sustainability disclosures. Applicability depends on the entity, activity and material.
• EU AI Act. The AI Act applies in phases. GPAI provider obligations applied from 2 August 2025; Article 50 transparency obligations generally apply from 2 August 2026. Regulation (EU) 2026/1744 postpones the main high-risk provisions to 2 December 2027 for Article 6(2)/Annex III systems and 2 August 2028 for Article 6(1)/Annex I systems. Whether a use is high-risk is fact-specific.

How GAIA-CORE is designed to control egress

The gateway is intended to run on customer-controlled EU infrastructure, not a GAIA25 cloud. Actual control depends on deployment, support access, integrations and configuration.

In the intended protected path, supported sensitive values are replaced with reversible tokens before external dispatch and the mapping stays within the customer boundary. The mechanism is built but default-off pending governed activation. It does not mean zero data leaves, every value is detected or re-identification is impossible; residual context and provider obligations remain.

GAIA-CORE is planned for AGPLv3 release. Until then, pilot buyers can negotiate controlled source review, dependency and test evidence, reproducible builds and continuity terms rather than rely on black-box claims.

Article-by-article evidence, not assurance

Controls matter only if you can demonstrate them. The gateway is designed to produce evidence that a compliance function, auditor or supervisor can review.

A configured tamper-evident audit trail can record model destination, timing and tokenised request data. This can support, not itself satisfy, AIFMD oversight and MiFID II record-keeping. DPIA and RoPA endpoints are designed to provide current processing information for Article 30 records and Article 35 assessments. Customer-hosting and provider abstraction can support third-party registers and exit planning; the customer must assess and maintain both.

The maturity signal is internal: 130,595 tracked TypeScript lines, 1,430/1,430 cumulative regression criteria and separate 10/10 and 18/18 HTTP suites at the assessed source state. This is engineering evidence, not external certification or customer proof.

The objective boundary and the first controlled pilot

GAIA Brain is just short of MVP, with no official customer deployment. Protected-egress controls remain default-off until production hardware-key binding, witnessed activation and customer validation. There is no general-availability claim or public pilot date.

We state this plainly because firms evaluating AI controls need candour about maturity. No regulator has endorsed GAIA25 or GAIA Brain. If this is your problem, discuss a bounded design-partner pilot with us. Your funds, data and supervisory relationship remain yours.