Who is accountable for the family office’s AI?

Ask a family office whether it uses AI and the answer is often no. Ask how the principal’s assistant drafts letters, how the analyst summarises bank statements, or what the new feature in the portfolio software does, and the answer changes. AI tends to arrive through habits and vendor updates rather than a board decision. The question that follows is simple: when one of these tools leaks a document or produces a wrong figure that someone relies on, who answers for it?

Monaco law already gives most of the answer. This note reads the texts available on 16 June 2026.

Start from what the office is

Law No. 1.439 created the multi family office activity. It is open only to Monegasque public limited companies authorised by ministerial order, with approval from the CCAF where the office carries out certain financial activities under Law No. 1.338. Article 9 subjects anyone carrying out the activity to professional secrecy under article 308 of the Criminal Code. The law excludes family office activity carried out privately among members of a single family.

The structure sets the accountable body. In a multi family office, the board of the company answers for its systems, as it answers for everything else, and the secrecy duty travels with each document an AI tool touches. A single family office outside the law’s scope still answers under data protection law, and to the family’s own governance.

The law already assigns accountability

Law No. 1.565 defines the controller as the body that determines the purposes and means of processing. Article 22 requires the controller to ensure, and be able to demonstrate, that processing complies with the law. Using an AI tool on client or staff data is a choice of means. The office cannot hand that accountability to its vendor.

Several articles turn this into concrete work:

• Article 26. A vendor that processes data for the office is a processor. It needs a written contract, including that it acts only on documented instructions.
• Article 35. Processing that, in particular through new technologies, is likely to create a high risk requires an impact assessment before it starts. The listed cases include systematic evaluation of personal aspects, including profiling, that leads to decisions with legal or similarly significant effects.
• Article 19. A person has the right not to be subject to a decision producing legal effects or similarly significant effects based solely on automated processing, subject to exceptions. Where an exception applies on the basis of contract or explicit consent, the controller must at least offer human intervention, a chance to state a view and a way to contest the decision.

A plain example: an office that lets a tool rank applicants for household or office staff, and sends rejections without a person reviewing them, has brought itself within article 19 and probably article 35.

When EU rules reach Monaco

Monaco is not in the EU, but Regulation (EU) 2024/1689, the AI Act, reaches beyond it. Article 2 applies it to providers and deployers located in a third country where the output of the AI system is used in the Union. An office that serves family members or entities in France or Italy, or produces reports used there, should check each use against that test.

Two parts matter most for a family office:

• AI literacy. Article 4 requires providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy among staff who operate or use AI systems. It has applied since 2 February 2025.
• High-risk uses. Annex III lists, among others, AI systems for recruitment and selection of people, and for evaluating the creditworthiness of natural persons. Portfolio analytics and document drafting do not appear there. Hiring tools might.

The dates were under renegotiation in 2026. As published, the Act applies generally from 2 August 2026, under article 113. On 7 May 2026 the Council presidency and Parliament negotiators reached a provisional agreement that would move the high-risk rules to 2 December 2027 for stand-alone systems and 2 August 2028 for systems embedded in products. When this note’s research closed, that agreement still awaited endorsement, legal-linguistic revision and formal adoption. A provisional agreement is not law: until an amending act is published in the Official Journal, the published dates stand. The safer planning question is not which date applies, but whether any use falls within Annex III at all.

Name owners, not committees

NIST’s AI Risk Management Framework, released in January 2023 for voluntary use, offers a useful vocabulary. Its GOVERN function asks for documented roles and lines of communication, for executive leadership to take responsibility for decisions about AI risk, for mechanisms to inventory AI systems, and for policies on third-party software and data.

In a family office of fifteen people, that reduces to a few named roles for each use:

• a business owner who decided the tool is worth using and answers for its outputs;
• a data owner who decides which files and which family members’ data it may reach;
• an operator, internal or contracted, who keeps it patched, logged and backed up;
• a reviewer for any output that leaves the office or moves money.

The board, or the principal in a single family office, approves uses that touch sensitive categories and reviews the list once a year.

A register the board can read

The register is one page per use:

• purpose and the people it affects;
• data categories and where they are stored;
• the model, its version and where it runs;
• the vendor and its contract under article 26;
• the human review point;
• impact assessment status under article 35;
• AI Act position, if any output is used in the EU;
• the incident path, linked to the office’s breach procedure;
• a review or retirement date.

Vendor features count. A document platform that switches on a summariser in an update has added a use, and the register should show it.

A human where it matters

ANSSI’s security recommendations for generative AI systems state that a generative AI system should not be able to take critical decisions with a strong impact on the business, or on the protection of property and people, without human control. For a family office the list of such decisions is short and clear: payment instructions, changes to bank details, signatures, messages to counterparties in the principal’s name. AI may draft them. A person approves them, in a system that records who approved what.

Where the system runs

A private deployment, on hardware the office controls, makes the register easier to keep true. The office can state where data went because it did not leave. It still carries every duty above. Running AI on premises changes the list of processors and transfers. It does not change who is accountable.

Limits of this note

This is not legal advice. It reflects texts available on 16 June 2026. The EU agreement described above had not been formally adopted at that date, and later changes are outside this note. Offices should confirm their position with Monegasque and, where relevant, EU counsel. To discuss a governance model, write to contact@desgarde.com.