# AI for debt collection: voice agent calls and WhatsApp reminders

> How to use AI for debt collection with phone calls and WhatsApp: a step-by-step collection sequence, a respectful script, what Brazilian law requires, when to hand off and what to measure.

- **HTML version:** https://tolky.to/en/blog/ai-for-debt-collection
- **Author:** [Marlos Carmo](https://tolky.to/en/author/marlos), Co-founder of Tolky
- **Published:** November 20, 2026
- **Reviewed:** November 20, 2026
- **Language:** en

## Summary

**AI for debt collection** is an agent that sends payment reminders, calls customers who are past due and negotiates within the company's rules, by voice and by WhatsApp, handing anything outside those rules to a person. It works best as a sequence: a reminder before the due date, a notice on the day, a call when the account is late and a negotiation, always without pressure or threats, as article 42 of Brazil's Consumer Protection Code requires.

Collections is the conversation nobody wants to have, and someone has to have it with hundreds or thousands of people every month. The team calls, half don't pick up, and whether a deal gets made depends on who was on shift.

An AI for debt collection changes the scale of that conversation: it reminds, calls, listens to why the payment is late and offers what the company has approved. But collections has its own rules, and a badly configured collections robocall leads to complaints, lawsuits and a blocked number. This guide covers the sequence, the script and the legal care, using Brazil as the reference market. Rules vary by country, so check your local law.

*Updated in November 2026.*

## What is AI for debt collection?

**AI for debt collection is a software agent that contacts customers with a payment coming due or past due, by phone or message, understands the reply in natural language and steers the conversation toward a goal: confirm the payment, record a promise to pay or close an agreement within the company's rules.**

The difference from the automated collection calls of the past is that the agent holds a conversation. Instead of playing a recorded message and hanging up, it asks whether the person has already paid, hears "I get paid on the 10th" and records the promise with the date. When the customer asks for something outside the rules, such as waiving the late fee, the agent passes the request to a person.

For the technical basics (transcription, language model and speech synthesis), see the guide to the [AI voice agent for customer service](https://tolky.to/en/blog/ai-voice-agent-customer-service). Here the focus is collections.

## How do you build a collection sequence with AI?

**A collection sequence, sometimes called a dunning sequence, is the series of contacts that follows an account from the due date to negotiation, with a different goal at each step.** The AI runs the sequence; the company decides the steps, the timing and what can be offered.

| Step | When (illustrative) | Suggested channel | Goal of the conversation |
| :--- | :--- | :--- | :--- |
| Reminder | 3 days before the due date | WhatsApp | State the amount and date, send the payment method |
| Due date notice | On the due date | WhatsApp | Remind with the payment code or link |
| Early delinquency | 1 to 5 days late | AI call | Understand the reason and record a promise to pay |
| Negotiation | 15 to 30 days late | AI call and WhatsApp | Offer installments or a new date within the rules |
| Handoff | Any step | Team member | Exceptions, disputes, discounts outside the rules |

The timing in the table is an example, not general advice. A school, a [vehicle protection association](https://tolky.to/en/blog/conversational-ai-vehicle-protection) and a lender run different sequences, and your own payment history shows which day works best.

Stop the sequence as soon as the payment lands, record every promise with a date and always give the customer a clear way to reach a person.

## Voice, WhatsApp or both: which channel should you use?

**Use WhatsApp to remind and voice to talk: a message works before the due date, when informing is enough, and a call works once the account is late, when you need to hear the reason and negotiate.** In most operations, the best results come from using both.

### When WhatsApp is enough

Reminders before the due date, duplicate invoices, payment codes and agreement confirmations. The message stays on record, the customer pays when they can and nobody is interrupted. The post on [bulk WhatsApp messaging](https://tolky.to/en/blog/whatsapp-bulk-messaging) explains the sending rules for the official API.

### When it pays to call

Once an account is late, messages often go unanswered. A call creates a conversation: the agent asks, hears "I lost my job" or "I thought I had paid" and responds accordingly. For WhatsApp calls, the contact has to grant permission first; over regular phone lines, the AI can reach any number. See how [WhatsApp Business calls](https://tolky.to/en/blog/whatsapp-business-calls) work and why a menu-based [outbound IVR](https://tolky.to/en/blog/outbound-ivr), still common in collections, cannot negotiate.

### Both together

The most common design is to call and then send a message with a summary and the payment method, so people who miss the call still get it and people who answer get the agreement in writing. And anyone who calls back needs to be answered, not dropped into an IVR that has no idea what the call is about.

## How do you write a respectful collection script?

**A respectful collection script identifies the company, confirms it is speaking with the account holder before mentioning any debt, states the amount without judgment and offers options.** The tone is a reminder, not a threat.

A short, illustrative script for the early delinquency step:

1. **Opening.** "Hi, this is the virtual assistant from Company X. This call is transcribed. Am I speaking with Ana Souza?"
2. **Identity check.** If it is not Ana, the agent thanks the person and ends the call without saying why it called.
3. **Reason.** "I'm calling about your October payment of R$ 189.90, due on the 10th. It still shows as open on our side."
4. **Listening.** "Did something come up? I can help you find a date that works."
5. **Options.** A new date, installments within the rules or the payment code sent on WhatsApp.
6. **Close.** Repeat what was agreed, record the promise and explain how to reach a person.

Avoid talk of ruined credit, threats of steps the company will not take, pushing after "I can't talk right now", calling several times a day and promising a discount the AI cannot grant.

## What does Brazilian law say about collection calls?

**In Brazil, the Consumer Protection Code bans exposing debtors to ridicule, pressure or threats when collecting, and makes it a crime to collect in a way that interferes with the consumer's work, rest or leisure.** These limits apply to people and to collection robots alike. Rules vary by country, so review the law where your customers live.

- **CDC, article 42.** When collecting debts, "the defaulting consumer shall not be exposed to ridicule, nor subjected to any kind of embarrassment or threat." The sole paragraph grants a refund of double any amount wrongly charged, unless the error was justifiable ([Law 8,078/1990](https://www.planalto.gov.br/ccivil_03/leis/l8078compilado.htm), checked on November 20, 2026).
- **CDC, article 71.** Using threats, coercion, embarrassment, false statements or any procedure that exposes the consumer to ridicule or "interferes with their work, rest or leisure" in debt collection is a crime, punishable by three months to one year of detention and a fine (same source).
- **Calling hours.** The CDC does not set a time window, but article 71 makes it clear that calling at night, early in the morning or repeatedly on the same day is a risk. Set business-hour windows and check state laws and your industry's rules.
- **Third parties.** Discussing the debt with whoever answers in place of the account holder exposes the consumer. Confirm identity first and, if it is not the account holder, end the call without stating the reason.
- **LGPD.** Brazil's data protection law, [Law 13,709/2018](https://www.planalto.gov.br/ccivil_03/_ato2015-2018/2018/lei/l13709.htm), allows processing data to protect credit (art. 7, X) and to perform a contract (art. 7, V), but requires a stated purpose, the minimum data needed and transparency (art. 6). See the [LGPD](https://tolky.to/en/glossary/lgpd-in-customer-service) glossary entry.
- **Do-not-call list and short calls.** Brazil's [Não Me Perturbe](https://www.gov.br/anatel/pt-br/consumidor/chamadas-abusivas) registry blocks telemarketing from carriers and payroll-loan banks, but does not apply to collection calls, according to [Procon-RS](https://www.procon.rs.gov.br/passo-a-passo-do-bloqueio). Even so, the regulator Anatel blocks callers that place a high volume of short calls for 15 days, a rule that changed on November 12, 2026 and that we explain in the post on [voice broadcasting](https://tolky.to/en/blog/voice-broadcasting) (checked on November 20, 2026).

This summary is not legal advice. Have your legal team review the script before the first campaign.

## When should the AI hand collections to a person?

**The AI should hand the conversation to a person whenever the customer disputes the debt, asks for something outside the rules, shows a sensitive situation or simply asks to talk to someone.** The handoff carries the reason and the transcript, so nobody has to repeat their story.

Common [handoff](https://tolky.to/en/glossary/handoff) signals:

- "I don't recognize this charge" or "I already paid": a dispute needs someone who can look up the case.
- A request for a discount or installments beyond the approved limit.
- Illness, a death in the family, job loss or obvious distress in the voice.
- An aggressive tone, a threat to sue or a mention of a lawyer.
- The direct request: "I want to talk to a person."

The guide to [AI to human handoff](https://tolky.to/en/blog/ai-to-human-handoff-7-rules) has ready-made rules for each case.

## How do you measure an AI collection campaign?

**Measure AI collections by the money recovered and by the cost of each goal reached, not by the number of calls placed.** Lots of calls with little conversation just burn credits and wear out the number.

| Metric | How to calculate | What to watch |
| :--- | :--- | :--- |
| Right-party contact | Calls answered by the account holder ÷ calls placed | A sudden drop may mean the number was flagged as spam |
| Promise to pay | Promises recorded ÷ right-party contacts | Promises without a date are worth little |
| Promises kept | Payments on the date ÷ promises | Shows whether the negotiation was realistic |
| Recovery | Amount received ÷ past-due amount in the campaign | The number finance tracks |
| Cost per goal | Campaign cost ÷ goals reached | Compare with the cost of your team doing the same |
| Handoffs | Calls passed to people and why | Repeated reasons call for a script change |

Listen to a sample of calls every week to see where the script sounds harsh.

## Where Tolky fits in voice collections

**Tolky is a conversational AI CRM in which the AI calls your contact list over WhatsApp, regular phone lines or both, follows a script and pursues a goal, with the conversation saved to the contact's history.** For collections, that is the [Voice Campaigns](https://tolky.to/en/solutions/voice-campaigns) solution, supported by the [collections](https://tolky.to/en/departments/collections) area page.

- **Goal and script.** You write the goal in plain text or point to a flow node. The AI generates the script (opening, permission, discovery, offer and close), which starts with a sales structure: for collections, you adjust the wording and rehearse the call playing the customer.
- **Hours and retries.** Calling windows, retries with an interval and a voicemail message.
- **Channel.** In combined mode, Tolky tries WhatsApp and, without permission, calls over the phone network.
- **Number protection.** Gradual calling, skipping people who were called recently and a do-not-call list that applies to every campaign, with an alert when the number may have been flagged as spam.
- **Transparency.** Every call opens with a transcription notice that cannot be turned off.
- **People and cost.** Live transfer during the call, scheduled callbacks, a credit estimate before you start, a daily spending cap and cost per goal reached on the real-time dashboard.

Customers who call back reach [Voice Service](https://tolky.to/en/solutions/voice-customer-service), available on activation: the AI answers on the company's number, knows what the call is about and transfers to the team when needed. In [financial services](https://tolky.to/en/industries/financial-services), that combination covers the whole cycle.

![Real screenshot of the Tolky Voice Campaigns screen, showing the period summary (people in audience, calls, answer rate and goals reached), the Do-not-call list button and the campaign list with Draft, Completed and Cancelled statuses](https://tolky.to/screenshots/campanhas-voz-en.png)

**What Tolky does not do:** it is not a portfolio management system or a credit bureau, and it does not decide on its own what discount to offer: the AI follows the rules your company writes.

## Frequently asked questions

### Can AI make debt collection calls?

Yes. A voice agent can call, confirm identity, state the amount and record a promise to pay or an agreement within the company's rules. In Brazil, collections remain subject to the Consumer Protection Code, with no pressure or threats.

### Are collection robocalls allowed in Brazil?

There is no ban on automated systems in collections, but the Consumer Protection Code limits apply all the same: no threats, no exposure to ridicule and no calls that disrupt work, rest or leisure. A high volume of short calls can also get the line blocked by Anatel.

### Can I discuss the debt with whoever answers the phone?

Only with the account holder. Confirm identity before mentioning any amount and, if someone else answered, end the call without saying why you called. Exposing the debt to third parties can amount to embarrassment under Brazilian law.

### Does Brazil's do-not-call registry block collection calls?

No. According to Anatel and Procon-RS, Não Me Perturbe blocks telemarketing from carriers and payroll-loan banks and does not apply to collections. Even so, anyone who asks not to be called should go on your own do-not-call list.

### Which channel works better for collections: WhatsApp or calls?

Both have a role. WhatsApp works for reminders and sending the payment method; calls work once the account is late, when you need to understand the reason and negotiate. Combining calls and messages usually reaches more customers.

### How do you know if AI collections are working?

Track right-party contact, promises to pay, promises kept, amount recovered and cost per goal reached. The number of calls placed tells you nothing on its own.

---

## Conclusion

AI for debt collection works when it follows a clear sequence, talks with respect and passes anything outside the rules to a person. Start with the early delinquency step, write a script that confirms identity before mentioning the debt, set calling hours and a do-not-call list and measure by the amount recovered.

[See Voice Campaigns](https://tolky.to/en/solutions/voice-campaigns), [explore Voice Service](https://tolky.to/en/solutions/voice-customer-service), [see the plans](https://tolky.to/en/pricing) or [talk to the team →](https://tolky.to/en/contact). Let's talk?

---

This is the Markdown version of the page. The full version, with images and navigation, is at https://tolky.to/en/blog/ai-for-debt-collection

Tolky: https://tolky.to
