Fiduciary Duty in the Age of Agency: What Family Offices Must Decide Before Their Agents Act

A family office exists to do one thing well: look after a family's capital, and its interests, across time. Its people are trusted because they are accountable. When a payment goes out, a counterparty is approved or a trust distribution is made, someone can say who decided, on what basis, and within what authority.

AI agents put that arrangement under strain. Systems that once answered questions can now prepare instructions, route approvals, reconcile accounts and, increasingly, act. For a lean office with growing reporting and compliance work, the appeal is obvious. The risk is less obvious, and it is not mainly technical. It is a question of fiduciary duty.

This essay sets out what changes when agents act on behalf of a family or a trust, and what a family office should decide before it lets them.

1. Delegation does not transfer responsibility

When a family office delegates work to an employee, an external manager or a service provider, the law and the family both expect it to choose carefully, set limits and supervise. Delegating to software is no different in principle. The office, its directors and its trustees remain responsible for what the system does.

What is different is speed and scale. An agent can take hundreds of small actions in the time a person takes to review one. A mistake that a person would make once, an agent can repeat across every account it touches before anyone notices. A delegation that would be prudent for a careful employee may be imprudent for a system that never pauses to ask.

The starting point, then, is not "what can the agent do?" but "what are we prepared to answer for?"

2. The mandate comes before the model

Most family offices have an investment policy, a governance charter or a family constitution. Few have a written mandate for AI. Without one, decisions about automation are made one tool at a time, by whoever is closest to the problem.

A mandate need not be long. It should say, in terms the family can read:

• which kinds of work AI may support, and which it may not touch
• which actions it may take on its own, which it may only recommend, and which always require a named person
• the value limits, counterparties and accounts that trigger human approval
• how the family will be told when something goes wrong

This is the same discipline I describe elsewhere as an AI delegation map: for each material system, a clear statement of what it may execute, what it may recommend and what it must never decide.

3. Trust deeds and mandates are rules, so encode them as rules

Discretionary trusts, investment mandates and family arrangements already contain boundaries: permitted assets, distribution conditions, required consents, reporting duties. A human administrator carries those boundaries in their head and in their checklists. An agent carries only what it has been given.

Where an agent prepares or routes instructions that touch trust assets, the relevant terms should be written into the system as explicit checks, not left to a general instruction to "comply with the trust deed". An action that falls outside them should stop, and a person should decide. Every decision, including the ones that were stopped, should leave a record that a trustee, an auditor or a family member could follow.

4. Count the value honestly

Family offices are rightly sceptical of technology narratives. The test for an agent should be the same as for any other expense: what does it do to the office's costs, risks and returns, at its full cost?

That means counting the cost of model usage, data connections, oversight and review, not just the licence. It means counting saved hours only when they are redeployed or replace a hire that would otherwise happen. And it means being cautious about claims that an agent improves investment performance. Better information can support better decisions, but the decision, and the accountability for it, remains with the investment committee.

For the companies a family owns, the same discipline applies. An AI initiative in a portfolio company should be tied to a line in its accounts, with a baseline and a named owner, before it is funded.

5. Regulation sets the frame, not the answer

In Singapore, family offices already operate within a defined framework. The Monetary Authority of Singapore's revised framework for single family offices took effect on 15 June 2026, and the fund tax incentive schemes set conditions on investment professionals, local spending and capital deployment. None of these were written with AI agents in mind, but all of them shape what a prudent office should automate, and how it should evidence its decisions.

Personal data obligations, including Singapore's Personal Data Protection Act, apply to family data as to any other. And the general expectation in Singapore's published AI governance guidance is consistent with fiduciary practice: human involvement proportionate to risk, clear accountability and controls across the life of the system.

Regulation will not tell a family office which agents to deploy. It does make clear that the office will be expected to explain what it deployed, and why.

6. Three questions for the next family meeting

Before an agent acts on a family's behalf, the principals and their advisers should be able to answer three questions:

1. What have we delegated? A list of the AI systems in use, including those inside vendor software, and what each may do.
2. Within what limits? The mandate, the approval thresholds and the trust and investment terms each system must respect.
3. Who answers for it? A named person for each material system, with the authority to switch it off.

If those answers are clear, AI can take real weight off a lean office and make its work more consistent, more transparent and easier to hand on to the next generation. If they are not, the office has not delegated work to a machine. It has delegated its judgment, and that is not something a fiduciary can do.

Views are the author's own. The frameworks described are applied methodologies of Kiros Intelligence Group, not regulatory standards.