GLBA follows the data, not the tool. Once an AI agent reads, writes or passes on a customer's nonpublic personal information, every GLBA duty applies to that processing: privacy notices, limits on sharing with outside parties, a written security programme and protection against pretexting. The platform and model provider become service providers you must contract with and oversee.
What does GLBA require?
The Gramm-Leach-Bliley Act sets financial privacy and security duties for US financial institutions. In substance it has three parts.
Privacy. Institutions must give customers a privacy notice describing what nonpublic personal information they collect and share. Before sharing it with non-affiliated third parties, they must, in most cases, give consumers the chance to opt out. There are exceptions, including sharing with service providers that perform functions for the institution under a contract limiting their use of the data, and sharing needed to process transactions the customer asked for. Information received from another institution may only be reused and passed on within limits.
Safeguards. Institutions must maintain a written information security programme with administrative, technical and physical safeguards to protect customer information. Which version of the rules applies depends on who supervises you: the federal banking agencies for banks, and a separate federal rule set for many non-bank financial firms.
Pretexting. It is unlawful to obtain customer financial information by false pretences, for example by impersonating the customer. Institutions are expected to have controls that make pretexting hard.
Who does it apply to?
Financial institutions in a broad sense: banks, credit unions, lenders, brokers, insurers, payment firms and others significantly engaged in financial activities. Rulemaking and supervision are split among the OCC, Federal Reserve, FDIC, CFPB, SEC, state insurance regulators and others, which is why the article on who regulates AI agents in financial services matters for scoping. Whether an agent deployment is caught at all is covered in does GLBA apply to AI agents handling customer data.
What does GLBA mean when the system is an agent that acts?
Four things stand out.
Every hop is a disclosure. When an agent sends a customer's account details to a model provider to draft a reply, that is a disclosure of nonpublic personal information to a non-affiliated party. It fits the service provider exception only if a contract limits the provider's use to performing the service for you. If the provider can use the data to train its own models, the exception is in doubt.
Reuse limits travel with the data. An agent that pulls in account data received from another institution, such as an aggregator, may only use it for the purpose it was received for. An agent that freely combines sources can breach this without anyone deciding to.
Agents are pretexting targets. A fraudster who cannot talk a trained person into revealing a balance may try the agent instead. An agent that discloses account information on the strength of a convincing story is a pretexting control failure. Authentication must happen before the agent sees the customer's data, not inside the conversation.
Vulnerable customers raise the stakes. Disclosures to someone claiming to act for an elderly or unwell customer are a common route for abuse. Agents need the same rules for third-party authority as your staff; our vulnerable customer handling checklist sets them out.
How do you meet GLBA when agents handle the data?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Accurate privacy notice | Notice covers agent processing and service providers involved | Notice versions; mapping of agents to notice sections |
| Opt-out before third-party sharing | Agents cannot send data to any party outside the service provider and transaction exceptions | Approved destination list; refused transmissions |
| Service provider contracts | Platform and model provider contracts limit use to your service and bar training on your data | Signed contracts; vendor attestations |
| Reuse and redisclosure limits | Data from other institutions tagged with permitted use; agents read it only for that use | Source tags; access rules |
| Written security programme | Agents and their access included in the programme's risk assessment | Programme document; risk assessment |
| Pretexting protection | Customer authenticated before the agent can read account data; third-party authority verified | Authentication logs; refused disclosures |
| Oversight of service providers | Initial and periodic review of the agent platform's controls | Due diligence file; review records |
The detail of access controls, which is the heart of the safeguards duty, is covered in GLBA safeguards and AI agent access controls.
What will an examiner ask?
- Which agents handle customer nonpublic personal information, and where does it go?
- Is each recipient covered by an exception, and where is the contract?
- Does any provider retain or use your customers' data for its own purposes?
- How does the agent know it is speaking to the customer before it discloses anything?
- Is the agent platform in your information security programme and your vendor risk register?
- Show us how the agent handles a caller claiming to act for an account holder.
What does GLBA not cover?
It does not cover the quality or fairness of agent decisions; credit decisions are governed by ECOA and Regulation B, which require specific reasons for denials. It does not cover health information, which falls under HIPAA. It does not pre-empt every state law: California's CCPA, as amended by the CPRA, exempts data covered by GLBA but can still reach other personal data the same firm holds. Outside the US, other laws apply; compare PIPEDA and AI agents handling customer data and POPIA and AI agents handling customer data in South Africa. For the full set of controls, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is sending data to a model provider a sale of customer information?
Not if the provider acts as a service provider under a contract limiting its use to your service. If it may use the data for its own purposes, you have a disclosure that may need an opt-out, and possibly a sale under state law.
Do we need to tell customers in the privacy notice that an agent is used?
GLBA notices describe categories of information collected and shared and with whom. They do not usually name tools. But if agent use changes what you share or with whom, the notice must reflect that.
Can an agent verify identity by asking security questions itself?
It can collect answers, but the check should be run by your authentication system, not judged by the agent. The agent should receive a verified status, not decide whether the answers were good enough.