Voice AI for arrears recovery, at scale and with empathy.
AI Voice Agent for a US Consumer Finance Platform
A polite, structured outbound voice agent that calls customers about overdue balances, agrees payment plans on the call, schedules callbacks, and routes disputes to human agents. Bidirectional, so the same agent handles inbound callbacks. In production, capturing thousands of structured outcomes per week.
Book a Case Walkthrough
A US consumer finance platform engaged Softblues to build a voice AI agent that could handle the bulk of routine arrears communication: payment reminders, plan offers, hardship routing, and dispute escalation. The brief was to scale outbound contact while improving consistency, lowering cost per contact, and keeping the conversation polite and compliant. The agent now operates in production, handling both outbound campaigns and inbound customer callbacks through the same core.
The Challenge
Scaling outbound arrears contact while keeping tone consistent, communication compliant, and every outcome structured.
- Manual outbound calling is expensive at scale; cost per contact for human agents is the binding constraint on coverage
- Human agent quality is inconsistent across shifts, particularly on tone, as patient, empathetic communication is hard to deliver after the hundredth call of the day
- Regulated outbound work requires strict adherence to mandatory disclosures, recording rules, and right-party verification on every call
- Customers who cannot pay immediately need a sensible alternative path (payment plan, callback, hardship route) rather than escalation
- Inbound callbacks need to land somewhere intelligent rather than into a generic IVR queue
- Outcomes must be captured in a structured way so collections operations can act on them
Our Solution
Softblues built a real-time outbound voice agent with empathetic tone engineered into the conversation graph, offering payment plans by default, and capturing a structured outcome on every call. The same agent core handles inbound callbacks.
- A real-time outbound voice agent built on Softblues' production voice stack (LangGraph orchestration, ElevenLabs voice, LiveKit transport)
- Structured conversation script with a compliant opening, right-party verification, and statement of purpose, followed by adaptive branching based on customer response
- Four standard outcome paths (will pay now, needs a plan, cannot pay now, disputes balance) with appropriate sub-actions for each
- Empathetic conversational tone by design, patient, polite, never argumentative, and always offers an alternative path before escalating
- Bidirectional architecture: the same agent core handles inbound customer callbacks as well as outbound campaigns
- Full integration layer: every outcome writes back to CRM, an audit log, and compliance recording storage; payment intent hands off to the existing payment processor
Built with Enterprise-Grade Technology
Goals and Objectives
The client came to us with clear objectives to transform their operations.
Scale Outbound Contact
Free human agents from routine reminder calls. The voice agent handles the high-volume, structured part of the workflow; human agents focus on disputes, hardships, and complex cases.
Empathetic at the Hundredth Call
The agent's tone stays patient, polite, and helpful regardless of call number. No fatigue, no frustration, no shortcuts. Customers receive consistent, respectful communication every time.
Compliance Built-In
Every call opens with a compliant disclosure, includes right-party verification, and is recorded with a full audit trail. The agent never deviates from required script elements.
Structured Outcomes Every Time
Every call ends with a captured outcome that operations can act on: paid in full, payment plan agreed, callback scheduled, escalated to human, or no answer. No call ends in ambiguity.
Bidirectional from Day One
The same agent core handles outbound campaigns and inbound customer callbacks. Customers who return a call land in the right place, and the agent picks up the conversation context.
See the Platform in Action
From intake to completion, explore how the solution transforms operations.
Structured Script with Adaptive Branching
Every outbound call follows the same opening sequence: a compliant introduction with company, purpose, and recording disclosure; right-party verification to confirm the agent is speaking with the account holder; and a clear statement of the outstanding balance. From that point, the conversation branches based on what the customer says. If they will pay now, the agent captures payment intent and hands off to the payment processor. If they need a plan, the agent offers options and agrees terms on the call. If they cannot pay now, the agent captures the reason and offers a hardship route or schedules a callback. If they dispute the balance, the agent stops pursuit and escalates to a human agent. Every path converges on a captured outcome, written to CRM, audit log, and compliance recording.
Bidirectional Operations, Single Agent Core
Outbound campaigns and inbound callbacks flow through the same voice agent core. The campaign list provides scheduled batches with retry logic for outbound work; inbound calls route directly to the same agent for customers calling back about an earlier conversation, in-flight queries, or replies to scheduled callbacks. The agent core uses LangGraph to orchestrate intent detection, response branching, and follow-up logic, with ElevenLabs handling voice and LiveKit handling transport. Every conversation ends with one of four structured outcomes, each wired into the platform's CRM, audit log, compliance recording storage, and payment processor handoff.
How It All Works Together
Voice Transport Layer
LiveKit handles real-time media transport, room orchestration, and concurrent call capacity. The same transport layer serves outbound campaigns and inbound callbacks with no separate infrastructure.
Speech Layer
ElevenLabs provides speech-to-text and text-to-speech with consistent voice fidelity at scale. Voice activity detection manages turn-taking so the agent knows when to speak and when to listen, even on noisy mobile lines.
Conversation Decision Layer
A graph of specialised LangGraph nodes handles each conversational decision in its own narrow prompt: opening compliance, right-party verification, intent recognition, branch selection, plan offer logic, and escalation triggers. The graph approach keeps accuracy high under variable customer responses where a single mega-prompt would crack.
Operations and Compliance Layer
Every outcome writes back to the platform's CRM. Every call is recorded with a compliance-grade audit log capturing the disclosure, verification, and outcome. Payment intents hand off to the payment processor; hardship routes hand off to human agents; callbacks schedule into the operations calendar. Designed to support compliance with US consumer protection rules covering automated outbound contact and debt collection practices.
Value and Impact Delivered
Measurable improvements across every dimension of operations.
Both Call Directions, One Agent Core
The same agent core handles outbound campaigns and inbound customer callbacks, with no separate infrastructure or duplicated conversation logic.
Structured Outcome on Every Call
Every conversation ends with a captured outcome (paid, plan agreed, callback scheduled, escalated, or no answer) written to CRM, audit log, and compliance recording. No call ends in ambiguity.
Empathetic Tone Maintained at Scale
The agent's tone is patient, polite, and helpful on every call, regardless of volume. Customers receive the same considered conversation on the thousandth call of the week as on the first.
Coverage Beyond Office Hours
The agent operates outside human contact-centre hours, expanding the window for productive customer contact and reducing the queue of accounts awaiting first-touch.
Compliance-Ready by Default
Every call includes a compliant opening with company, purpose, and recording disclosure. Every recording is stored with an audit log linking the disclosure to the outcome captured. US consumer protection rules built into the conversation graph.
Sub-2s Response Latency
End-to-end round-trip response latency keeps the conversation natural, with no awkward pauses between customer speech and the agent's next turn, even over mobile lines.
Frequently Asked Questions
Ready to Transform Your Financial Services Operations?
See how AI can help your organisation reduce errors, speed up processing, and improve outcomes. Let's discuss your specific challenges.
Book Discovery CallExplore Other Projects
Discover more AI solutions delivering measurable results across industries