When Customers Under Guardianship or Power of Attorney Use Your AI Agent

Quick answer

When a meaningful share of your customers may be under guardianship or have a power of attorney on file, your AI agent needs a documented recognition and escalation policy: how it identifies these accounts, how it authenticates the authorized representative, what the agent can and cannot do without human review, and how it handles a conflict between what the customer says they want and what the legal record says the representative is authorized to decide.

Why this requires specific policy, not just generic vulnerable-customer rules

General vulnerable-customer guidelines tell your AI agent to be patient, clear, and to escalate when the customer seems confused or distressed. That is necessary but not sufficient for customers under a legal instrument like guardianship or power of attorney.

The difference is legal authority. A guardian appointed by a court or a person holding a financial power of attorney has a legally recognized right to make decisions on behalf of the customer. Your AI agent has no mechanism to verify that authority in a real-time conversation. But ignoring the issue creates risk in two directions.

If you act on a guardianship holder's instructions without verification, and the instruction is fraudulent or outside the scope of the instrument, you may have participated in financial or personal harm to a vulnerable person. If you ignore a legitimate guardian's instructions because your agent has no policy for handling them, you are potentially obstructing a court-ordered arrangement.

The accounts where this matters

Guardianship and POA most commonly arise in three contexts relevant to consumer AI agents.

Elderly customers with cognitive decline. A family member holds financial POA and may contact your agent on the customer's behalf or alongside the customer.

Adults under disability-related guardianship. A court-appointed guardian manages certain decisions. The customer may themselves be active on the account.

Customers in a medical or temporary incapacity. A healthcare POA or a durable POA becomes active during a period when the customer cannot manage their own affairs.

In each case, your agent may encounter a caller or user who is not the account holder but claims authority to act on their behalf.

The four policy elements you need

Recognition. How does your system flag an account as having a legal representative on file? This requires a process for customers or their representatives to register the instrument, and a flag in your CRM or account system that is visible to the AI agent's context.

Authentication. Your AI agent should not authenticate a legal representative through standard customer authentication. The AI cannot verify a court order or a POA document in a conversation. The policy should route any interaction claiming a legal representative authority to a human agent or a dedicated verification process that can review the document.

Scope limits. Even a verified legal representative has a scope. A financial POA does not necessarily authorize changes to a healthcare account. A guardian may have authority over financial decisions but not communications preferences. Your policy should define what actions a verified representative can take via the AI agent channel and which require additional review.

Conflict escalation. When the person interacting with your AI agent is the customer themselves, and the customer's instructions conflict with the representative's registered preferences, your agent needs a defined escalation path. It should not resolve that conflict unilaterally.

What the AI agent itself should do

The AI agent should not attempt to adjudicate legal representative authority in real-time. Its role is to recognize the situation and route correctly. Specifically:

When a caller or user identifies themselves as acting on behalf of an account holder under a legal instrument, the agent should pause the transactional interaction, acknowledge the situation, and transfer to a human agent or a verification queue. It should not confirm, deny, or attempt to validate the instrument in the conversation.

When an account has a registered legal representative flag, the agent should be configured to apply heightened care: slower pacing on transactional decisions, confirmation requests before any account changes, and lower escalation thresholds.

When the account holder themselves expresses confusion or contradicts a registered representative preference, the agent should escalate rather than proceeding based on either input.

FAQ

Do we have to accept POA on file for all account types? No, but you should have an explicit policy for each account type. "We do not accept legal representative registrations for this account type; customers in this situation should contact us through [channel]" is a valid policy. What is not valid is having no policy and handling it inconsistently.

What if the guardian calls and the customer is also active on the account? This is the hardest case. Both have some claim to interact with the account. The AI agent should not adjudicate this. Route to human review with a note that both parties are active.

Can we use a transcript of a phone call to verify a POA? No. A person saying "I have a POA" in a phone call is not verification of the instrument. The verification process must include document review, and that document review must happen outside the AI agent conversation.

Is there regulatory guidance on this? In the US, the Consumer Financial Protection Bureau has published guidance on elder financial exploitation that is relevant to this area. In the UK, the Financial Conduct Authority has vulnerability guidance that covers this. Neither specifically addresses AI agents, but both establish expectations for how companies handle customers who may have diminished capacity or legal representatives. Apply those expectations to your AI channel.

Related reading:

• AI Agent Foundation Model Inherited Bias Management
• When Your AI Agent Gives the Correct Answer in the Wrong Tone
• AI Agent Human Agent Conflicting Answers to the Same Question