AI Voice Agents for Debt Collection: Rules, Design and Build

Why collections teams are adopting voice AI

Collections is a volume business. Lenders, banks, utilities, telecom providers, healthcare billers and agencies contact large numbers of customers every day, often with the same messages: a payment is due, a payment was missed, here are your options. Staffing for that volume is expensive, and the work is repetitive and emotionally draining for agents.

AI voice agents can make early-stage and routine calls consistently, at any hour the rules allow, in the customer's language, and never get tired or impatient. They follow the script and the rules on every call, which is not always true of large human teams under pressure.

The opportunity is strongest in early collections, before accounts become seriously delinquent, where a timely, polite reminder and an easy way to pay or arrange a plan resolve many cases. Later-stage and complex cases still need skilled people, now with more time to spend on them.

Customers often prefer it too. Many people find it easier to discuss a missed payment with an automated assistant than with a person, and an AI agent that offers clear options and a simple way to pay removes some of the embarrassment and friction from the conversation, while a person is always available on request.

Where AI fits across collection stages

Most organisations start with pre-due reminders and early delinquency, measure results, and expand only when compliance and customer outcomes are clearly good.

Segment accounts before choosing who the AI calls. Customers with a history of disputes, recent hardship flags or legal holds should go straight to staff, while customers who simply forgot a payment are ideal for an automated reminder with an easy way to pay.

A compliant call flow

• Before dialling: check consent, calling hours for the customer's location, frequency limits per debt and any cease or do-not-call requests.
• On answer: confirm you are speaking to the right person before mentioning the debt; if not, end politely without disclosure.
• Give the required disclosures in the required wording, and say the caller is an AI assistant.
• Explain the balance and options clearly; offer only pre-approved plans and amounts.
• Take a promise to pay or route to secure payment; confirm by text or email.
• Escalate disputes, hardship, vulnerability and requests to stop calling immediately, and record them.
• If voicemail: leave only a permitted limited-content message, or no message.

Keep the flow short. Customers are more likely to stay on the line and agree a plan when the call gets to the point quickly, so disclosures and verification should be clear but efficient.

United States: FDCPA, Regulation F and TCPA

US collections calls are governed by the Fair Debt Collection Practices Act and the CFPB's Regulation F for debt collectors, plus the TCPA for automated and artificial-voice calls and state laws, some of which are stricter. Key points that shape an AI agent:

This is a planning summary, not legal advice. Collection agencies, first-party lenders and different states can face different requirements; your compliance counsel confirms the rules for your portfolio.

United Kingdom, India and the UAE

Rules change; for example, RBI has consulted on further changes to recovery conduct. Build the rules as configuration per country and portfolio, so they can be updated without rewriting the agent.

Example: a compliance gate before every dial

The safest design puts rule checks outside the AI, in code that decides whether a call may be placed at all. The AI never decides on its own to call someone.

# compliance_gate.py: simplified pre-dial checks for one account and debt
from datetime import datetime, timedelta
from zoneinfo import ZoneInfo

RULES = {
    "US": {"start": 8, "end": 21, "max_calls_7d": 7, "quiet_days_after_convo": 7},
    "IN": {"start": 8, "end": 19, "max_calls_7d": None, "quiet_days_after_convo": 0},
}

def may_call(account, now_utc: datetime) -> tuple[bool, str]:
    rules = RULES[account.country]
    local = now_utc.astimezone(ZoneInfo(account.timezone))
    if account.cease_requested or account.do_not_call:
        return False, "cease or do-not-call on file"
    if account.requires_ai_voice_consent and not account.ai_voice_consent:
        return False, "no consent for AI voice calls"
    if not (rules["start"] <= local.hour < rules["end"]):
        return False, f"outside calling hours ({local:%H:%M} local)"
    if account.state_rules and not account.state_rules.allows(local):
        return False, "stricter state rule"
    week_ago = now_utc - timedelta(days=7)
    attempts = [a for a in account.call_attempts if a.debt_id == account.debt_id and a.at >= week_ago]
    if rules["max_calls_7d"] is not None and len(attempts) >= rules["max_calls_7d"]:
        return False, "weekly call limit for this debt reached"
    last = account.last_conversation_at
    if last and now_utc - last < timedelta(days=rules["quiet_days_after_convo"]):
        return False, "inside quiet period after a conversation"
    return True, "ok"

Real implementations add per-state rules, holidays, channel preferences, bankruptcy and legal holds, and an audit log of every decision. Keeping this logic separate makes it testable and easy to show to auditors.

Tone, fairness and vulnerable customers

Collections conversations are stressful for customers. A good AI agent is calm, clear and respectful, and its consistency is an advantage: it never raises its voice or uses pressure tactics.

Review a sample of calls each week specifically for tone and fairness, not just for outcomes. Regulators increasingly look at customer outcomes and treatment, and so should you.

Language matters as much as tone. Offer the customer's preferred language at the start, avoid jargon such as “delinquency” or “charge-off”, and confirm understanding before ending the call, so agreements are clear and fewer disputes follow.

Taking payments safely

Payment is where collections calls create value, and where security matters most.

• Capture card details outside the AI with masked keypad entry or a secure payment link sent by text.
• Pause recording and transcription during payment capture, and log the pause.
• Use tokens from the payment provider; never store or show full card numbers to the AI or in logs.
• Confirm the amount and date verbally, then send a written confirmation with a reference.
• Write promises and payments back to the account system immediately to stop further calls.

Offer more than one way to pay. Some customers will pay during the call, others prefer a link they can open later, and some want to visit a branch or use their banking app. Making payment easy matters more than which channel it happens in.

Measuring results

Measure outcomes and conduct together. A collections agent that improves recovery while increasing complaints is not a success.

Compare AI-handled and human-handled accounts in the same segment and stage rather than across the whole portfolio, because the AI usually starts with the easiest accounts and a blended comparison flatters it.

What it costs

As a guide, matching our voice agent cost ranges:

Running costs are per minute for telephony, speech, model and voice. Collections calls are often short and many go unanswered, so telephony and answering-machine handling matter more than in other use cases. See our AI voice agent cost guide and predictive dialer guide.

Common mistakes in AI collections

Most problems with AI collections come from treating it as a dialing efficiency project rather than a regulated customer conversation.

Frequently asked questions