Overview
A US-based consumer finance company specialising in personal instalment loans needed a scalable way to contact existing borrowers who had fallen behind on payments. These were not collections calls in the aggressive sense. Most borrowers were existing customers who had missed a payment because of timing, oversight, or short-term cash flow pressure. The company’s goal was to contact them quickly, verify identity, explain the account status clearly, and offer a simple next step — either a secure payment link or a transfer to a senior team member. The challenge was scale. Their internal system identified newly overdue borrowers three times per day, creating roughly 300 new contacts per batch and around 900 call attempts per day. Human agents could not keep up consistently, and the nature of the conversations required a tone that was compliant, calm, and respectful. RapidCall was used to deploy an AI voice agent that could handle this outreach autonomously while following a tightly controlled script and escalation flow.The Goal
The company needed an agent that could:- call newly overdue borrowers automatically,
- verify identity before discussing account details,
- share the past-due balance clearly,
- offer a secure payment link by SMS and email,
- transfer the borrower to a senior agent when needed,
- and do all of this in a compliant, respectful, non-confrontational way.
Results After 6 Weeks
March 2026 performance
The 11.1% conversion from connected conversations to payment links sent reflects the reality of overdue-account outreach. Most attempts reached voicemail, and many borrowers who answered were not ready to engage. What mattered was performance inside the conversations that did connect.
Out of 1,360 live conversations, the agent successfully sent 151 payment links and generated 44 transfers to senior agents, with no compliance incidents, no escalation failures, and no borrower complaints.
The sentiment profile was also strong for this use case. In debt-related outreach, the target is not positive sentiment. The target is calm, controlled, non-hostile interaction. 85% neutral sentiment indicated that the agent was delivering difficult account updates without escalating tension in most calls.
Why This Worked
Several things made this deployment effective. First, the agent did not try to collect payment directly on the call. It focused on verification, clarity, and routing. That reduced friction immediately. Second, borrowers were always given a choice: receive a secure payment link, speak to a senior team member, or end the interaction. That preserved a sense of control. Third, the agent was designed to sound legitimate and respectful. This mattered because borrowers are highly sensitive to tone, hesitation, and anything that feels robotic or untrustworthy.Voice Stack and Configuration
Core stack
The call flow was structured and rules-based, so the company did not need a heavier reasoning model. GPT-4.1 was chosen because it handled the script reliably at a cost and latency level that worked well for high-volume outbound operations.
The voice needed to sound calm, warm, and credible. A steady American female voice performed well for this use case because it carried empathy without sounding overly soft or patronising.
Call-centre ambience was added intentionally. In a context like personal finance, subtle background office noise made the call feel more legitimate and reduced the sterile silence that can make AI calls feel unnatural.
Key speech settings
These settings were tuned specifically for payment-recovery conversations.
A very high response eagerness was important because hesitation in this context reduces trust. When someone answers a call about a financial account, long pauses do not feel thoughtful. They feel uncertain.
Interruption sensitivity was kept high but not extreme. Borrowers sometimes react emotionally, and the agent needed to yield when interrupted, but not get derailed by every small background sound or reactive noise.
Accuracy-optimised transcription mattered because the agent needed to verify dates of birth. Mishearing a single digit would lead to failed verification and a wasted call.
Prompt Design
The agent followed the standard five-part prompt structure used across RapidCall deployments, with an additional voicemail rule placed at the very top.Role
The agent, named Rosie, was framed as an automated customer care assistant helping the company’s team. The persona was:- calm,
- warm,
- respectful,
- clear,
- and never pushy.
Context
Each call included dynamic data pulled from the company’s internal system, including:- customer name,
- date of birth for verification,
- loan number,
- past-due balance,
- days past due,
- secure payment link,
- voicemail status,
- and state information.
Guidelines
The prompt included strict behavioural rules. The key constraints were:- verify both name and full date of birth before sharing any account details,
- never read back or hint at the stored date of birth,
- never make promises or speculate,
- ask only one question at a time,
- use the customer’s name sparingly,
- offer transfer when appropriate,
- and stay within a narrow, compliance-safe scope.
Main script
The live call flow had seven core steps:
The script was linear by design. The goal was not to create a highly flexible conversation. The goal was to move borrowers safely and efficiently toward one of a small number of valid outcomes.
Handler Design
The agent included dedicated handlers for the most common edge cases.
These handlers mattered because overdue-account conversations rarely stay on the ideal script. The agent needed specific, non-generic responses for resistance, skepticism, and emotional reactions.
One particularly important design choice was proactive AI disclosure. Rather than waiting for the borrower to ask, Rosie disclosed early that she was an automated assistant helping the company team. In this use case, that increased trust rather than hurting it.
Workflow Architecture
The agent sat inside a fully automated outbound workflow.Data source
The company’s loan management system pushed overdue contact records into a Google Sheet three times per day. Each push contained the full set of variables required for the call. The sheet acted as a lightweight CRM and operations board, allowing managers to monitor status without touching the production database.Call cadence
Calls were distributed across three daily windows:- morning,
- midday,
- and afternoon.
Voicemail logic
Voicemail handling was one of the most important design decisions. On the first unanswered attempt, the agent left a short voicemail. On later unanswered attempts, it hung up silently rather than leaving repeated messages. This prevented voicemail flooding, helped protect caller ID reputation, and kept the complaint rate at zero.Outcome logging
After every call, the workflow wrote outcomes back into the Google Sheet, including:- call result,
- category,
- timestamp,
- sentiment,
- summary,
- recording URL,
- and message delivery status.
State-based filtering
Because the company operated across multiple US states, the workflow filtered contacts by state before dispatching calls. This ensured that outreach only happened in authorised jurisdictions and created a cleaner audit trail.Payment-link delivery
When a borrower chose the self-service path, the agent triggered a function that sent a unique payment link by SMS and email. The link pointed directly to the company’s own payment portal. No sensitive payment data passed through the agent.What We Learned
1. Voicemail strategy matters more than most teams expect
Without controlled voicemail logic, the agent would have left multiple voicemails per borrower per day. That is how numbers get flagged as spam and complaint risk rises. Leaving one voicemail, then going silent on later attempts, worked far better.2. Proactive AI disclosure increased trust
The initial concern was that disclosure would reduce engagement. In practice, it did not. By disclosing early and then immediately moving into a legitimate verification flow, the agent sounded transparent rather than deceptive.3. Neutral sentiment is the right target
In overdue-balance outreach, positive sentiment is not the objective. Neutral is. The goal is to deliver the message clearly, respectfully, and without triggering unnecessary hostility.4. Choice reduces friction
Giving borrowers two clear paths — secure payment link or live transfer — made the interaction feel less confrontational. Most borrowers chose the lower-pressure self-service option.5. Accuracy matters more than raw speed in verification-heavy flows
Switching to accuracy-optimised transcription improved date-of-birth verification noticeably. The slight latency trade-off was worth it.6. A Google Sheet was enough
For this deployment, a simple Google Sheet worked very well as an operational CRM. It gave the team real-time visibility, filtering, and auditability without added complexity.7. Multiple daily windows improved reach
Spreading calls across morning, midday, and afternoon windows produced meaningfully better reach than a single daily pass.8. The dispute handler prevented escalation
When borrowers challenged the balance, the agent did not defend it. It either offered written details or a transfer to a senior team member. That was the correct design. Arguing would have damaged trust immediately.When to Use This Pattern
This playbook is a strong fit for any business that needs to contact existing customers about:- overdue balances,
- missed payments,
- account status changes,
- billing reminders,
- or resolution pathways that must follow a controlled script.
- call volume is high,
- compliance rules are strict,
- the process is repetitive,
- and the human team should focus on judgment-heavy escalations rather than first-touch outreach.
The human team handles the exceptions, disputes, and higher-stakes conversations.
PROMPT TEMPLATE: