Yes. GLBA applies to financial institutions, not to technologies, and an agent that reads, uses or shares nonpublic personal information about consumers on a financial institution's behalf is part of how that institution handles the data. Its privacy notices, sharing limits and information security programme must cover the agent, its tools and the vendors behind it.
Who does GLBA apply to?
The Gramm-Leach-Bliley Act covers US financial institutions, a category that is wider than banks. Lenders, mortgage brokers, investment advisers, insurers, payment firms and many fintechs significantly engaged in financial activities are included. Supervision and enforcement are split among agencies by type of institution, including the federal banking agencies, the CFPB, the SEC and state insurance regulators.
In substance, GLBA does three things that matter for agents:
- Privacy. Institutions must tell customers how they share nonpublic personal information and, for many kinds of sharing with non-affiliated third parties, let them opt out. Sharing with service providers that act for the institution is permitted under conditions.
- Safeguards. Institutions must run an information security programme with administrative, technical and physical safeguards, and oversee service providers that handle customer information.
- False pretences. Obtaining customer information by deception, such as impersonating the customer, is prohibited, and institutions are expected to guard against it.
The detail is worked through in GLBA and AI agents handling customer financial data.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What is different when the system is an agent?
Most systems handle data along fixed paths. An agent decides at run time what to read and where to send it. Three consequences follow.
Sharing can happen by action. An agent with an email tool or an outbound API connection can disclose customer data to a third party in the course of "helping", for example by forwarding a statement to an address in a message thread. Each outbound tool is a potential sharing channel that your privacy notice and opt-outs must account for.
The vendor chain is longer. Behind an agent there is often a platform, a model provider and connected data sources. Each that receives customer information is a service provider you must oversee.
Pretexting gets a new target. Fraudsters who once called a contact centre can now try to talk an agent into revealing account details, or plant instructions in a document the agent reads. GLBA's safeguards duty extends to this.
What may the agent do, and what stays with a person?
- Agent may: read customer records needed for the task, answer an authenticated customer about their own account, and draft documents.
- Agent may, within limits: send information to the verified customer through approved channels.
- Stays with a person: any disclosure to a third party not on an approved list, changes to contact details used for verification, responses to requests from people claiming authority over an account, and decisions on suspected pretexting.
An illustrative example
Take a US mortgage servicer that uses an agent to answer borrower emails. A message arrives from an address that matches the borrower's name but not the address on file. It asks for the latest payoff statement "for my attorney" and gives a new email address for delivery.
An unbounded agent would comply: it has the document and a send tool. The servicer's agent is bounded differently. It may only send documents to contact details on file, and any request to change the destination is held for a person. The request is held, the borrower is contacted through the number on file, and the attempt is logged as suspected pretexting.
Which controls meet GLBA for agents?
| GLBA duty | Agent control | Evidence |
|---|---|---|
| Privacy notices and opt-outs | Outbound tools mapped to sharing categories; opt-out flags checked before sharing | Tool inventory and check results |
| Safeguards programme | Least-privilege data access per agent; encryption; access reviews | Access logs and reviews |
| Service provider oversight | Due diligence and contract terms for platform and model provider | Vendor files |
| False pretences | Authentication before disclosure; destination changes held for a person | Holds and verification records |
Privacy laws elsewhere follow different logic: Canada's approach rests on consent, described in PIPEDA and AI agents handling customer data, and Brazil's is covered in whether the LGPD applies to AI agents. The cross-regulation map is in AI agent regulation, mapped to controls.
Questions people ask
Is our model provider a third party under GLBA?
If it processes customer information to provide a service to you, it is usually treated as a service provider. You still need to oversee it and contractually require it to protect the information.
Does GLBA ban sending customer data to an AI model?
No. It requires the sharing to fit your notices and exceptions, and the recipient to be overseen and bound to protect the data. Using the data to train a provider's general model would need careful review.
Do state privacy laws also apply?
They can. Many state privacy laws exempt data covered by GLBA but not everything an agent touches. Check each data type against both.