Which Bank Calls Should an AI Voice Agent Answer?

A bank’s first AI voice agent should handle a narrow set of approved information requests and routing tasks. Public branch information and appointment requests are reasonable candidates to test. Account lookups need verified identity and a reliable data connection. Fraud, disputes, hardship and other urgent conversations need a prompt route to trained staff.

The boundary is the action and information involved, not how ordinary the caller’s opening sentence sounds. “What is my balance?” can become “I did not make that withdrawal.” The agent must recognize the change and preserve the report, even if the original lookup worked.

Glia’s banking page markets routine voice self-service and routing to bank specialists. Treat that as a capability to verify against your systems. It does not tell you which calls your bank should automate or what share will resolve successfully.

Use the open scorecard to evaluate a voice AI proposal.

Define four call permissions

The call matrix below is a proposed deployment boundary for a first evaluation, not a legal safe harbor or a claim that every platform supports it.

Freeze-card actions, replacement cards, address changes, payments and fee adjustments deserve a separate authorization design. Do not include them merely because a vendor can expose a button or API. For the initial scope, route requests involving these actions to the established service process.

Likewise, a published rate is not a personalized offer. If rates are in scope, bind the answer to an approved product, effective date and relevant conditions. When the source is stale or the caller’s question requires interpretation, transfer the call instead of improvising.

Design the transfer before the greeting

Specify where each urgent or unsupported request goes during business hours and after hours. A generic callback promise is inadequate if nobody is responsible for receiving the task. Agree what the agent says when a live queue is unavailable and how the bank’s existing urgent reporting route remains reachable.

The transfer record should carry the caller’s stated problem, time received, verified identity status, relevant account reference and actions already completed. Limit sensitive information to the systems and staff authorized to receive it. A short summary helps, but staff need access to the original report if the summary is inaccurate.

The CFPB’s 2023 chatbot spotlight estimated that over 98 million users, about 37% of the U.S. population, engaged with a bank's chatbot in 2022, and it identifies inaccurate answers, failure to recognize disputes and barriers to human support as risks. It is not a voice-product certification. Have the bank’s compliance team translate applicable complaint, error-resolution, disclosure and recordkeeping obligations into the actual workflow.

Voice adds two risks that chat does not

Treat the caller's voice as unverified, and treat outbound AI calls as regulated calls.

First, authentication. FinCEN's November 2024 alert on deepfake fraud notes that criminals may use tools that generate synthetic audio and video to respond to live verification prompts, and that scammers use deepfake voices to impersonate trusted people. A voice agent should never accept how a caller sounds as proof of identity. Use the bank's approved authentication factors, and treat a request to change contact details, add a payee or move money as higher risk that belongs with staff.

Second, outbound calling. The Federal Communications Commission ruled in February 2024 that the Telephone Consumer Protection Act's restrictions on an "artificial or prerecorded voice" cover current AI technologies that generate human voices, so calls using them require the called party's prior express consent absent an emergency purpose or an exemption. If the same agent will place appointment, payment or collection calls, have compliance confirm consent records and every other rule that applies before it dials.

Use difficult calls in the acceptance test

Give the implementation team a bank-approved set of test conversations. Include callers who change topics, speak over the agent, use an unsupported language, provide incomplete details or request a person immediately. Include silence, background noise and failed authentication. Use authorized test data rather than exposing customer account details in a demonstration.

• Public answer: a holiday changes the branch schedule. The agent must use the current approved answer or acknowledge it cannot verify the hours.
• Account lookup: authentication fails. No account-specific information should be revealed while the caller is offered an appropriate recovery route.
• Urgent change of topic: a balance caller reports an unfamiliar transfer. The agent stops the routine flow and routes the report.
• System outage: the core cannot answer. The agent must not convert an old cached balance into a claim about the current balance.
• Incomplete transfer: no staff member accepts the handoff. The task remains visible and follows the bank’s agreed fallback.

Review transcripts and outcomes by call type. Count wrong answers, repeat calls, failed transfers and time to a qualified person, as well as completed self-service requests. A call ending without a live agent is not automatically a resolved call.

Estimate the workload after the pilot

Use observed call mix, handling time, repeat-contact rate and residual staff work to assess economics. Include telephony, software, integration, testing, monitoring and knowledge maintenance. Avoid applying another institution’s deflection percentage to your queue.

Expand one permission at a time after the relevant tests pass and staff can recover failed interactions. If the next candidate is loan-status service, establish which status the loan origination workflow can reliably provide. The broader finance operations plan should connect the phone answer to the work that actually resolves the customer’s request.