Rail Governance

GLBA Pretexting Rule Obligations for AI-Assisted Customer Interactions

GLBA's pretexting rules already cover AI agents, so document your controls.

· 9 min read
Cover illustration for “GLBA Pretexting Rule Obligations for AI-Assisted Customer Interactions”
SOC 2 in Financial Services · November 11, 2026 · 9 min read · 1,951 words

AI-assisted customer interactions bring GLBA’s existing pretexting obligations into a new channel. With no AI-specific interpretation, financial institutions must work through how the statute applies to the way their agents verify identity, disclose financial information, and represent themselves. For compliance and operations leaders, those obligations require documented controls even before rules written expressly for AI arrive.

What GLBA's Pretexting Provisions actually prohibit, and why AI channels make the question urgent

GLBA’s Pretexting Provisions, 15 U.S.C. §§ 6821–6827, prohibit obtaining or disclosing customer financial information through false, fictitious, or fraudulent pretenses. Together with the Privacy Rule and the Safeguards Rule, they form the law’s baseline data protection framework. Their classic target was impersonation by information brokers. In the late 1990s, those brokers posed as consumers to get account data.

The Southern District of New York held in FTC v. RCG Advances, LLC in 2023 that “the plain text of the statutory provision controls,” rejecting an impersonation-only reading. Under that interpretation, any misrepresentation coupled with the collection of financial information may constitute a violation. An AI channel thus raises questions about both its ability to spot an impostor and whether what it says creates false pretenses for collecting customer data.

How the absence of AI-specific GLBA guidance creates a reasoning obligation, not a compliance holiday

A new customer interface does not suspend existing obligations. AI-assisted channels have no GLBA-specific carve-out, even though no AI-specific rule or formal interpretation has been issued under the law. The institution must reason through how the statute applies to its channel. That requires distinguishing established duties from questions about how particular AI behavior might be treated.

The federal guidance gap is explicit in a related area. The OCC, Federal Reserve, and FDIC published revised interagency model risk management guidance on April 17, 2026, through OCC Bulletin 2026-13. It states that “generative AI and agentic AI models are novel and rapidly evolving” and excludes them from its scope. That exclusion is not permission to leave those systems ungoverned. The NCUA has likewise issued no AI-specific rules, but says that existing technology-neutral rules apply and that credit unions must carry out appropriate due diligence when they use vendor AI.

What does that leave compliance leaders to do? They must connect each existing obligation to the actual customer interaction, rather than treat the absence of tailored guidance as an exemption. They should trace where deception can enter the exchange and where financial information can leave it. That produces three specific exposure points, which are identity verification, information disclosure, and conversational authenticity.

Identity verification: how AI channels inherit the pretexting rule's toughest choke point

Generative AI tools allow an attacker to clone a customer’s voice and use it against an AI-assisted channel. The attacker may try to pass voice authentication or induce credential resets, higher transfer limits, or the addition of a beneficiary. FinCEN has reported an increase in suspicious-activity reports describing suspected deepfake media and warns that deepfake audio can be used to impersonate people. The deception is familiar, even if the instrument is new.

Pretexting turns on whether the channel becomes a means of obtaining customer financial information under false pretenses, a narrower issue than whether the institution can stop every fraudulent transaction. Voice biometrics alone is no longer a reliable control for a voice-based AI channel. A system’s confidence in a voice match does not resolve the underlying question of who is requesting access.

For institutions addressing that exposure, the recommended architecture combines callback verification against a registered device with multi-factor verification for high-risk actions. Those actions include large transfers, new payees, and contact-information updates. Friction belongs at the point of heightened risk. Applying it uniformly would obscure the distinction between an ordinary interaction and a request that changes access or payment authority. The institution must also distinguish a channel deceived by an attacker from one misconfigured to expose data, because those failures call for different controls.

Information disclosure: when an AI agent's conversational output becomes a pretexting exposure

The risk changes direction when the agent supplies information. An AI agent handling customer financial prompts processes nonpublic personal information, or NPI, through unstructured text. Conventional data-loss-prevention tooling built for structured data does not adequately inspect that conversational outflow. Authentication is therefore only part of the protection. The institution also needs to govern what the agent can retrieve and what it can release.

When a misconfigured agent reveals one customer’s information to another, that is a security failure, but the mistake alone does not establish false pretenses. By contrast, a social-engineering prompt that tricks an agent into revealing NPI places deception directly in the path to disclosure. Both belong in the institution’s pretexting risk assessment, without treating them as legally identical. Access restrictions address the reach of a misconfigured agent, while controls over unstructured text egress address information leaving through conversational output.

Broad access makes either failure more consequential. An agent that can read customer records across accounts or products creates a Safeguards Rule gap where least-privileged access and segregation of duties should apply. The written information security program must cover role-based access controls and audit logging in the AI pipeline. Separately, the RCG Advances interpretation raises potential exposure when the agent itself misrepresents what it is or what it can share in connection with a transfer of financial data.

Conversational authenticity: the obligation to govern what an AI agent represents itself to be

If a customer supplies financial information believing the agent is a human banker, and the institution designed or permitted that confusion, the false-pretenses standard could apply to the institution’s system. This remains an application of existing law in the absence of a settled AI-specific ruling. Institutions must assess whether the representation helps obtain financial information through deception.

That question creates a genuine operational tension. An agent that repeatedly hedges about its identity can make a simple customer exchange cumbersome. Yet a disclaimer buried in terms of service does not govern what the agent says during the conversation. Institutions should define and document what their agents represent themselves to be, build that representation into the interaction, and test it under the written information security program’s monitoring requirements.

Third-party AI vendors and the Safeguards Rule: why the institution owns the exposure it outsources

Most financial institutions acquire AI. Customer NPI then passes through the vendor’s staff and systems. Outsourcing the interface does not transfer the institution’s Safeguards Rule responsibility for how that information is handled. Vendor oversight is the mechanism that connects the institution’s obligations to systems outside its direct operation.

That connection must exist in the contract. AI vendors handling customer information must be contractually bound to equivalent safeguards, with terms that permit institution oversight and require incident notification. A missing contractual obligation is a compliance gap even if the vendor’s actual practices are sound. Institutions also need clear answers about how NPI is segregated, retained, and deleted. Without those answers, they cannot adequately assess whether the outsourced channel exposes information to deceptive acquisition or disclosure.

The NCUA’s position brings credit unions to the same practical task. Its required due diligence covers how the vendor’s AI works, the risks it adds, and the safeguards the vendor uses. A vendor’s description of the product is not enough. The institution needs to connect those answers to the customer information and interactions entrusted to it, rather than assume that acquiring a service also acquires a defensible compliance posture.

Enforcement terrain: the penalties and precedents that make delayed compliance costly

GLBA pretexting violations can bring civil and criminal penalties. Criminal violators can face fines under federal law, imprisonment for up to five years, or both. In aggravated cases, the prison term can be as long as ten years. These are the statute’s stakes, not a prediction that an AI error will produce a criminal case. For boards and senior management, the useful distinction is between the consequences the law permits and the facts needed to establish a violation.

The FTC’s enforcement trajectory also extends beyond traditional impersonation cases. The FTC alleged a GLBA pretexting violation under 15 U.S.C. §6821 against Greystar, which agreed to a substantial settlement in December 2025, while the Colorado Attorney General’s claims were brought under the Colorado Consumer Protection Act. The case demonstrates the breadth of institutions exposed to the FTC’s interpretation. At a May 2026 FTC workshop, Bureau of Consumer Protection Director Chris Mufarrige reaffirmed pretexting as an active tool for misrepresentation combined with financial-data collection and separately noted AI’s use in fraud.

The agency has a further reason to use that tool. The Supreme Court’s 2021 AMG Capital decision curtailed the FTC’s ability to seek equitable monetary relief under Section 13(b). The FTC has since leaned on GLBA pretexting as an independent basis for consumer redress. Leaders seeking funds within their institutions can use this to make the case that AI governance is a compliance need now, even before future AI rules arrive.

Audit trails and explainability: the governance infrastructure pretexting compliance requires in AI pipelines

Agent workflows often lack the detailed audit trails that examiners and internal auditors expect. Without a consistent trail, an institution cannot reconstruct why an agent accessed information, disclosed it, or took an action. That weakens incident response and the evidence available to demonstrate governance. Explainability in this setting means being able to answer a concrete question about the handling of customer financial information.

The IMF’s 2026 note How Agentic AI Will Reshape Payments identifies an architecture relevant to that task. It includes mandate-based authorization, separation of decision-making from execution, agent identity frameworks, programmable payment controls, audit trails, and tiered human involvement. These mechanisms map to Safeguards Rule requirements without creating a separate AI-only compliance regime. Separation of duties remains a useful design constraint. An agent’s ability to make a decision need not imply unrestricted authority to execute it.

For institutions applying those controls, the written information security program must explicitly cover AI risk assessments, access controls, encryption, monitoring, and incident response. Keeping a program written before deployment on file is not enough to meet that duty. Existing banking rails can provide a structural advantage because transactions executed on them inherit audit and control mechanisms already in operation. That gives institutions more evidence of the governance they intend to put in place, though they must still address conversational access and disclosure.

A practical compliance mapping: applying each pretexting obligation to AI-assisted customer interactions

Institutions should list every AI system that touches NPI, record what it can access, and name an owner who is accountable for it and has the power to stop deployment. The mapping must reflect the channel. A voice agent, a text chat agent, and a transaction-initiating workflow need different controls even though the law places the same obligations on them.

For identity verification, institutions should set out the verification architecture in the written information security program, because a verbal policy alone is not enough. High-risk voice interactions need callback verification against a registered device and multi-factor verification. For information disclosure, each agent should be able to access only the records it needs for that interaction, with egress controls for unstructured text. For conversational authenticity, the documented identity representation should be part of interaction design and regular testing.

Vendor arrangements must support the same mapping. Institutions should require clear accounts of data segregation, retention, and deletion, and verify that contracts were updated after the 2023 Safeguards Rule amendments to impose equivalent safeguards, oversight, and incident notification. A documented review cadence should then reopen the assessment after any material change to the model, data, or use case. The work does not end at deployment. Until AI-specific GLBA guidance exists, reasoned application of existing law is the governing task, and each change to the channel must remain answerable to it.

More in SOC 2 in Financial Services