AI phone agents for financial services

Financial services has the same phone problem as everyone else and a much shorter leash. The calls are repetitive, the volumes are high, and the regulator has opinions about how you handle them.

AI phone agents work well here, but only if compliance is designed in from the first conversation rather than bolted on before launch.

TCPA: consent before you dial

In February 2024 the FCC confirmed that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act. That is the single most important fact for anyone planning outbound calling.

In practice it means prior express consent before an AI agent calls a US mobile number, clear disclosure that the caller is speaking to an AI system, an opt-out that actually works, and records that prove all three.

The exposure is not theoretical. TCPA penalties run from 500 to 1,500 dollars per call, and there is no aggregate cap. A campaign that dials without consent can generate a very large number very quickly.

Inbound is a different matter. Someone who calls you has initiated the contact, which is why most first deployments in this sector are inbound only.

FINRA treats your script as a communication

If you are a FINRA member firm, what the agent says is not just product copy. A script that describes your services or your investment approach is a communication with the public, and falls under FINRA Rule 2210 like any other marketing material.

That means your compliance officer reviews the script before it goes live, and reviews it again when it changes. Treat prompt changes with the same discipline you would apply to a brochure, because that is effectively what they are.

Every call is a record

Recorded calls and transcripts are potentially books and records under your firm's retention obligations. Before launch, decide where recordings live, how long they are kept, who can retrieve them and how quickly you could produce them if asked.

This is worth settling early because it constrains vendor choice. A platform that cannot export transcripts on demand, or that retains data somewhere you cannot account for, is not usable regardless of how good the voice sounds.

Where the boundary sits

An agent can comfortably handle scheduling, answering factual questions about your process and hours, confirming appointments, collecting basic information and routing calls.

It should not give advice, make recommendations, discuss specific holdings or performance, or attempt to handle a complaint. Any caller disclosure that amounts to a regulated complaint should route to a qualified human, and that rule should be built in rather than left to the model's judgement.

Where to start

Inbound, during business hours, on a narrow set of call types. Read the transcripts weekly with your compliance officer for the first month. Expand only into call types you have seen handled correctly.

It is a slower rollout than other sectors justify. In this one it is the whole point.