GLBA requires US financial institutions to protect customers' nonpublic personal information and to train the staff who handle it. Once AI agents read account data, draft responses and act on records, that training must cover the agents too: what data each agent may see, how pretexting targets agents, and what staff must check before customer data leaves the firm.
What does GLBA ask of staff who work with agents?
The Gramm-Leach-Bliley Act governs financial privacy in the United States. In substance it does three things that matter here. It requires financial institutions to tell customers how they collect and share personal financial information, and to honour opt-outs from certain sharing with unaffiliated third parties. It requires them to safeguard that information through a written information security programme, which includes risk assessment, access controls, oversight of service providers, and security awareness training for staff. And it prohibits pretexting: obtaining a customer's financial information by false pretences.
None of this mentions AI. All of it applies when an agent touches customer data. An agent that reads statements to answer a question, drafts a letter with account details, or passes data to a model provider is processing nonpublic personal information. The people who brief, supervise and approve its work are handling that information through it. The systems view is in GLBA and AI agents handling customer financial data.
Who does it apply to?
GLBA applies to financial institutions in a broad sense: banks, credit unions, lenders, brokers, insurers, investment advisers, payment and money transmission firms, and other businesses significantly engaged in financial activities. Oversight is split by sector. The OCC, the Federal Reserve and the FDIC cover the banks they supervise, the CFPB has authority over privacy rules for many consumer financial firms, the SEC covers securities firms, and state insurance regulators cover insurers. State regulators such as NYDFS add their own cybersecurity expectations.
Inside a firm, the staff whose training matters include contact centre and servicing teams who supervise customer-facing agents, operations staff who approve agent-drafted correspondence, the people who brief agents and choose their data sources, IT and security staff who manage agent access, and vendor managers who oversee model and platform providers. Whether the Act reaches the agent itself is covered in does GLBA apply to AI agents handling customer data. The short answer: the firm's duties do.
What competence does GLBA imply when agents act?
Security awareness training written for email and passwords does not cover agents. Staff need five additional skills.
Knowing what each agent may read. Every agent has a set of sources. A servicing agent may read balances and recent transactions; it should not read full account numbers or another customer's file. Staff who brief agents must apply least access. Staff who supervise them must notice when an answer contains data the agent should not have had.
Recognising pretexting aimed at the agent. A caller who cannot pass identity checks with a person may try the chat agent instead, using a partner's name, an urgent story or a forged document. Staff must know how the agent verifies identity, that a declared claim is not verification, and that an emailed document is external evidence that cannot authorise disclosure.
Checking what leaves the firm. Agent-drafted letters and emails can include more data than needed. At depth 2 the agent drafts and a person sends, which gives that person a chance to strip unnecessary detail. Where the agent sends routine messages at depth 3, staff must sample them.
Handling sharing preferences. Customers who opted out of sharing with unaffiliated third parties must not have their data passed to such parties by an agent. Staff configuring agent connections must check opt-out status flows through.
Overseeing service providers. A model provider or agent platform is a service provider under the safeguarding duty. Vendor managers need to know what data reaches it, where it is stored, and whether it is used to train models.
How do you meet the duty and show it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Security awareness training | Agent-specific module for staff who brief, supervise or approve agents | Training records by role |
| Access controls | Agents limited to the sources their task needs; staff trained to spot excess | Agent source lists; access reviews |
| Pretexting protection | Agents refuse disclosure without verified identity; staff trained on agent-targeted pretexting | Refusal records; seeded test results |
| Minimum data in outbound messages | Drafts held for review or sampled at depth 3 | Review and sample logs |
| Honour opt-outs | Opt-out status checked before any agent sharing | Configuration records; test results |
| Service provider oversight | Due diligence and contract terms for model and platform providers | Vendor assessments; contracts |
| Record of agent activity | Every agent verdict filed with the data sources used | Append-only decision record |
The pretexting row deserves a seeded test. Place a few calls or chats built to pressure the agent into disclosure and see whether it refuses and whether staff spot the attempt in the queue.
What will an examiner or auditor ask?
- Which agents access nonpublic personal information, and which sources does each read?
- How are staff trained on agents, beyond general security awareness?
- How does an agent verify identity before disclosing account data?
- Show us agent-drafted letters: who checked what data they contained?
- How do opt-outs reach agent configurations?
- What do your model and platform providers do with customer data?
What does GLBA not cover?
It does not regulate automated decisions or fairness; ECOA and Regulation B govern credit denials, including the need for specific reasons when a model is used. It does not cover health information, which HIPAA governs. It does not displace state privacy laws, such as California's CCPA as amended by the CPRA, which apply to some data GLBA does not reach. And it does not set an AI literacy standard as such. Firms outside the US face similar duties under other laws; the Kenya Data Protection Act and staff training is one comparison. The guide to AI literacy and certification sets out the general model.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does pasting customer data into a general AI tool breach GLBA?
It can, if the tool's provider is not overseen as a service provider and the data is not protected. Staff should use only approved agents with known data handling.
Do agents count as staff for GLBA training?
No. Training duties apply to people. Agents are governed through access controls, bounds and the safeguarding programme.
Who should own agent-related GLBA training?
The person who oversees the information security programme, working with the teams that run the agents.