The audit cycle for the FTC's 2023 Safeguards Rule amendments is now active. If your institution deployed a voice AI vendor before those rules clarified, you may have signed a contract that doesn't satisfy your current third-party oversight obligations. This guide gives you the four operational criteria

GLBA puts on your desk, the specific failure mode of AI architected for deflection rather than resolution, and three questions to put to any vendor in an RFP.

What is GLBA, and what does it require for contact center AI?

The Gramm-Leach-Bliley Act requires financial institutions to protect the security and confidentiality of customer financial information. For contact centers, the operative regulation is the FTC's Safeguards Rule, which mandates a written information security program and, since its June 2023 amendments, substantially expanded obligations around service provider oversight.

What changed in 2023: financial institutions must now select service providers that maintain appropriate safeguards, require those safeguards contractually, and oversee service provider compliance on an ongoing basis. For any institution running voice AI in its contact center, that third obligation — ongoing oversight — is the active compliance pressure.

The audit cycle has arrived. Institutions that deployed voice AI before the amended rule existed typically did so under contracts that predate the expanded service provider provisions. Those contracts may not satisfy what the FTC now requires your vendor relationship to look like — and the vendors who were not built with auditability as a first-order concern are now scrambling to retrofit it.

What does GLBA compliance require from a voice AI vendor?

A voice AI vendor cannot itself be "GLBA compliant" — GLBA compliance is the institution's obligation. What a vendor can do is support your compliance program. Four operational requirements define whether a vendor is built to do that or isn't.

Data minimization and access controls. Your information security program must govern what customer data your service providers can access, store, and transmit. For a voice AI deployment, that means your vendor must be able to specify, precisely, what data leaves the call, where it goes, who can query it, and how long it lives. "We comply with SOC 2" does not answer this question. A data flow diagram with retention policies and access tier definitions does.

Authentication integrity. Voice AI handling account inquiries will encounter callers who need identity verification before accessing sensitive information. GLBA doesn't specify the authentication method — that's your institution's decision — but it requires the method be applied consistently and that your AI vendor's system interoperates with your existing identity verification layer without creating gaps. The failure mode is an AI that passes callers to agents after a partial authentication check, creating a handoff where authentication status is ambiguous.

Service provider contractual requirements. Your contract with a voice AI vendor must include provisions requiring the vendor to maintain appropriate safeguards and notify you of any security events affecting customer data. If your current vendor contract doesn't include these provisions, you have a compliance gap independent of how well the technology works.

Escalation and audit trail requirements. When your AI encounters a caller it can't authenticate or a request it can't fulfill, the handoff to a human agent must be traceable. Your information security program implies that your contact center, including the AI layer, has defined escalation paths that are documented and auditable. An AI that drops difficult calls into a general queue without passing authentication state or call context to the receiving agent creates both a customer experience failure and a compliance exposure.

Why deflection-optimized AI fails GLBA audits

Most voice AI in financial services contact centers was built on one premise: reduce call volume by keeping customers away from human agents. The IVR systems that preceded these deployments shared the same goal. Large language models bolted onto that deflection architecture inherit its priorities. The measure of success is containment rate. Auditability, authentication chain integrity, and structured data handling are afterthoughts.

The failure scenario is specific. A caller claims an account using the last four digits of a card number compromised in a data breach — a credential that is, in the attacker's hands, not meaningless. A deflection-optimized AI passes a partial authentication check, moves the caller into the account inquiry flow, and either transfers them to an agent with ambiguous authentication state or resolves the inquiry without a transfer at all. The AI logged a successful containment. The institution has a fraud exposure and a Safeguards Rule gap.

Resolution-optimized architecture addresses this differently. PolyAI's deployment at Simply Health handles 30 to 40 percent of inbound calls and reduced claims processing time from two days to under 21 hours — not by keeping callers away from humans, but by completing the interactions that don't require human involvement while escalating the ones that do, with full context intact. That clean escalation path, with call data and authentication state passed to the receiving agent, is what makes the handoff auditable. A contained call that deflects rather than resolves leaves nothing behind for the audit trail.

How to evaluate a voice AI vendor for GLBA compliance support

Three questions to put to any vendor in an RFP or technical evaluation.

Where does call data live, who can access it, and what is the retention policy?

The answer should include geographic data residency, which roles within the vendor organization have query access to call records, and the default retention period with options to configure it. If the vendor cannot answer this in a single meeting without escalating to their security team, that is your answer.

How does your authentication model interact with our existing identity verification layer, and who owns liability when the AI passes a caller it shouldn't?

You need to know whether your existing authentication logic can be called before the AI enters the account inquiry flow, and whether the result is passed as a structured signal the AI acts on — not something the AI infers from the conversation. The liability question belongs in your contract.

What does the audit trail look like for a call that escalated to a human, and can you produce it in a format a regulator would accept?

Ask for a sample. A call that went from AI to agent should produce a record including: call timestamp, authentication status at handoff, reason for escalation, and agent ID. If the vendor's audit trail is a call recording and a containment flag, that is deflection architecture making itself legible.

Frequently asked questions

Does PolyAI support GLBA compliance?

PolyAI is built to support your institution's GLBA compliance program. GLBA compliance is the institution's obligation — a vendor supports it by providing auditable call records, structured escalation paths, configurable data retention, and contractual safeguard provisions. PolyAI's architecture is designed around resolution rather than deflection, which means escalation paths exist by design rather than as an edge case.

What is the difference between GLBA compliance and SOC 2 certification?

SOC 2 certifies that a vendor has adequate internal controls over security, availability, and confidentiality of its own systems. GLBA compliance requires your institution to govern what your vendors do with customer financial information specifically. A SOC 2 report tells you something about a vendor's internal security posture; it does not tell you whether your contract with that vendor satisfies the Safeguards Rule's service provider provisions.

What should a GLBA-compliant voice AI contract include?

At minimum: provisions requiring the vendor to maintain appropriate safeguards for customer financial information, notification requirements for security events, data residency specifications, retention and deletion obligations, and defined escalation path requirements for calls involving account access. If your current vendor contract predates the June 2023 Safeguards Rule amendments, it likely needs review.

What GLBA risks are specific to voice AI that don't apply to other contact center technology?

Voice AI introduces three risks that static IVR does not: AI-driven authentication decisions (where the model's judgment replaces a deterministic rule), unstructured call data that may be difficult to govern under a retention policy, and escalation handoffs where the AI does not pass call context cleanly to the receiving agent. Each of these has a direct Safeguards Rule implication.

Still curious? Read this guide to learn why domain-specialist LLMs are more secure than publicly available generalist models like those from Anthropic, OpenAI, and Google.