Conversational AI in banking is software that talks with customers to get something done: freezing a lost card, answering “did my paycheck clear?”, reminding someone a payment is due, or calling back an applicant who stalled halfway through a loan form. Banks have had chatbots for years. The CFPB found in 2023 that all ten of the largest US banks run one, and that about 37% of Americans used a bank chatbot in 2022. The phone line is still where the stressful calls land, and that is where voice agents are now going.
This guide is for the operations, risk, and engineering leads deciding which of those calls an agent should take. Short version: start with calls that have one right answer in one system, verify the caller before the agent sees anything, keep card data out of the model, let your dialer enforce the outbound rules, and always leave a door open to a person.
What conversational AI handles well in banking
The calls that work have a clear goal, a system of record the agent can query, and an obvious point where a person takes over.
| Customer need | What the agent does | What it must connect to | When it hands off |
|---|---|---|---|
| Balance or recent transaction | Reads back the balance or a specific transaction | Core banking, after verification | The caller disputes a charge |
| Lost or stolen card | Freezes the card, confirms the freeze, explains replacement | Card management | The caller reports fraudulent charges |
| Fraud alert confirmation | Calls out, asks whether the customer made a flagged purchase | Fraud case system | The customer says no |
| Payment reminder | Reminds, offers payment options, records a promise to pay | Servicing system, payment scheduler | Hardship, disputes, or a request to stop |
| Loan application follow-up | Calls a stalled applicant, finds the blocker, guides them back | Application status | Eligibility or pricing questions |
| Account opening documents | Explains which document is missing and how to send it | Onboarding workflow | Identity checks that fail |
| Branch and hours | Hours, locations, holiday schedules, what to bring | An approved knowledge source | The answer isn’t in the source |
The loan follow-up row is easy to underrate. Fundamento runs it for Hero FinCorp: when an applicant sits idle on a loan form for five minutes, often because the form is in English and they aren’t a native speaker, the agent calls them in their language, sorts out the problem, and walks them back in. 42% of those dropped-off applicants came back and finished.
Why the phone, and why it’s harder
Customers who call a bank are rarely browsing. They lost a card, saw a charge they don’t recognize, or owe money they can’t pay yet. They want the problem gone, and they want to trust the voice that’s handling it. Fundamento’s co-founder put it plainly: voice quality is “what makes people pick up the phone, engage with our agents, and give more honest answers.”
Voice also raises the bar. The agent has to hear account numbers and names correctly over phone audio, know when the caller has finished talking, and reply before the pause feels like a dropped line. A chatbot that misreads a digit shows a typo. A voice agent that mishears one reads out someone else’s transaction, or nobody’s.
Where the line goes
An agent can tell a customer their balance. It should not tell them whether to refinance, decide whether they qualify for a hardship plan, or rule on a disputed charge. Write that boundary into the agent’s instructions: the tasks it handles, the topics it never answers, and what it says instead.
The harder boundary is recognizing when a caller is invoking a right. The CFPB’s report warned that a chatbot may “fail to recognize that a consumer is invoking their federal rights”: a customer saying “that charge isn’t mine” may be starting a formal error dispute, not making small talk. Treat these phrases as transfer triggers and test them with callers who say them sideways: “I never bought that,” “this has to be a mistake,” “I want to talk to someone about my rights.”
And keep the exit open. The same report says chatbots that prevent access to live help “can lead to law violations.” An agent that loops a frustrated customer through the same three options is a compliance problem as well as a bad call. If the caller asks for a person, transfer them.
Keep card data and secrets out of the model
Banking agents touch the most sensitive data a company holds. Check five things before production:
- Verify before you reveal. The agent shouldn’t read a balance or a transaction until your identity check passes. Run that check in your systems, and pass the agent a verified flag, not the raw answers.
- Never send card numbers, PINs, or one-time codes through the language model. If a call needs a card number, collect it with keypad tones in a payment flow that stays inside your PCI scope, and pass back only a token or a success message. Tell the agent in its instructions to stop a caller who starts reading out a PIN.
- Minimum necessary access. Give the agent narrow tools, such as
get_balanceorfreeze_card, scoped to the verified customer, not a general query interface to the core. - Retention and redaction. Decide how long recordings, transcripts, and logs live. On Cartesia, enterprise customers can turn on zero data retention for speech to text and text to speech, and the Delete Call API removes a call’s transcript, audio, and logs on demand.
- Where it runs. Many banks need data to stay in-country. Cartesia offers dedicated regional deployments in the US, EU, UK, India, and Australia, plus VPC and on-premises deployment. Fundamento chose Cartesia partly because of its India data center: in a market with strict residency rules, providers without local infrastructure were non-starters.
Cartesia is SOC 2 Type II, GDPR, and PCI DSS service provider compliant. Details are in the Trust Center. Your security and compliance teams still need to review the full call path, because the agent also touches your systems.
Outbound calls: collections, reminders, and fraud alerts
Outbound calls are where banks see results fastest, and where the rules are densest. In the US:
- Consent. The FCC ruled in 2024 that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act, so the same consent rules apply as for prerecorded calls.
- Frequency. For debt collection, Regulation F presumes compliance only if a collector calls a person about a particular debt no more than seven times in seven consecutive days, and not within seven days after a conversation about it. That limit belongs in your dialer or campaign logic, where it can be audited. Don’t leave it to the language model.
- Disclosure. Have the agent say who it is and that it’s automated in the first sentence, and run every collections script past counsel.
Inside those rules, tone carries a lot of weight. Regal uses Cartesia voices to bring warmth and empathy to collections and confirmation calls. That matters when the person on the line is behind on payments.
Speed, numbers, and languages
Three details decide whether a banking agent sounds trustworthy:
- Speed. Fundamento runs more than 20 million outbound calls a month on Cartesia, with a 100 ms time to first byte, at least twice as fast as every other provider it tested. On an outbound call, the first second decides whether the person stays on the line.
- Numbers read back exactly. Amounts, dates, and the last four digits of an account have to come out right. Confirm them in a form people can check: “a payment of two hundred forty dollars, due Friday, October 16th.”
- The customer’s language. Fundamento’s agents speak Hindi, English, and a growing set of regional Indian languages. Cartesia’s speech model covers more than 40 languages with one model, so the same agent can switch to the language the caller speaks.
How to build a banking voice agent on Cartesia
On Cartesia Managed Agents, you configure the agent and Cartesia runs the speech recognition, language model, speech synthesis, and phone connection.
- Write the instructions. Define the role, the tasks, the topics the agent never answers, and the exact transfer behavior for disputes, fraud, hardship, and any request for a person.
- Connect narrow tools. Webhook tools let the agent call your APIs, such as
get_balanceandfreeze_card. Gate each one on a verified session. - Add transfers. The
transfer_to_numbersystem tool routes the call to a phone number or SIP address when a condition you write is met.send_dtmfcan press keys when the agent has to navigate another phone menu. See the warm transfer guide for handing context to the person who picks up. - Teach it your vocabulary. Add product names, branch names, and card program names as
keyterms. - Personalize outbound calls. Pass the customer’s first name, amount due, and due date as dynamic variables, so the agent doesn’t need account access to open the call. Bring your own Twilio account or SIP trunk for numbers outside the US.
{
"config": {
"instructions": "You are the automated servicing assistant for Harbor Federal Credit Union. Only discuss account details after the verify_caller tool returns verified. You can read balances and recent transactions, freeze a lost or stolen card, and explain how to order a replacement. Never ask for or repeat a full card number, PIN, password, or one-time code; if the caller starts to say one, stop them. Never give financial advice. If the caller disputes a charge, reports fraud, mentions hardship, or asks for a person, transfer them.",
"initial_message": "Thanks for calling Harbor Federal. This is the automated assistant. How can I help?",
"audio": {
"input": {
"keyterms": ["Harbor Federal", "Harbor Rewards Visa", "Bayview branch"]
},
"output": { "voice_id": "e07c00bc-4134-4eae-9ea4-1a55fb45746b" }
},
"system_tools": {
"transfer_to_number": {
"pre_tool_speech": "force",
"transfers": [
{
"destination": { "type": "phone", "phone_number": "+18005550147" },
"condition": "The caller reports fraud, says they did not make a charge, or disputes a transaction."
},
{
"destination": { "type": "phone", "phone_number": "+18005550100" },
"condition": "The caller asks for a person, mentions financial hardship, or asks for advice."
}
]
}
}
}
}How to pilot it
Pick one call type, such as lost cards or payment reminders, put the agent on it with staff as the fallback, and review calls daily for the first two weeks. Track:
- Resolved without a person. Cards frozen, balances answered, promises to pay recorded, as a share of calls that asked for them.
- Transfers by reason. Disputes and fraud should transfer every time. Avoidable transfers are your backlog.
- Verification failures. How often real customers fail the check, and where they drop off.
- Complaints and opt-outs. Any rise is a signal to stop and listen to calls.
- Boundary tests. A weekly scripted set: callers who read out a PIN, ask for advice, or dispute a charge in roundabout words. The agent should handle every one correctly.
When the numbers hold, extend hours before you add call types, because nights and weekends are when unanswered card and fraud calls pile up.
To try it, create an agent in the Cartesia playground, connect a test endpoint, and call it from a real phone. The financial services page has a live demo of a loan assistance agent.