AI debt collection automation in Mexico: case study and results

Numerous contrasts mark the Mexican microloan market. On the one hand, tens of millions of people who lack access to traditional banking services regularly use payday loan services. On the other hand, this very segment shows some of the highest delinquency and default rates — a challenge that makes AI debt collection automation especially relevant.

According to the International Monetary Fund’s Financial Access Survey, the volume of outstanding loans in Mexico’s microfinance institutions reached about 42 billion pesos in 2023 — roughly over $2 billion. In some companies, overdue payment rates can reach 20% or higher. At the same time, analysts forecast that the microfinance market in Mexico will grow by more than 12% between 2024 and 2028. Despite this scale, the average loan amount remains modest, at approximately $294.

Dozens of providers operate in this market, competing not only for customers but also for efficiency in their business models. Among the most prominent players are:

• Kueski — over 1 million customers and about 6 million loans issued; the market leader in BNPL and microloans.
• Konfío focuses on loans for small and medium-sized businesses, having already supported over 85,000 companies.
• Creditea — part of the international IPF Group; serves about 800,000 customers in Mexico and has issued over 700,000 loans.
• Baubap — a mobile app for microloans; already has 2 million customers and over 8 million loans, with plans to reach 6 million users by 2025.

For lenders, this creates a unique contradiction: loan amounts are small, but the customer base is massive. Large-scale call operations with human agents can become prohibitively expensive, and every poorly designed process can result in millions of dollars in costs.

Against this backdrop, debt collection automation is not just an experiment but a strategic step. AI agents can handle the most repetitive tasks, such as mass calls, payment reminders, and confirmations of promises to pay. This eases employee workload, reduces costs, and, in practice, enhances customer interactions.

It’s worth noting that major Mexican players, such as Kueski, Konfío, Creditea, Doopla, and others, are not Apifonica clients. However, this case demonstrates the real-world outcomes of automation and can serve as a benchmark for companies seeking more effective debt recovery methods.

Project Goals and Choosing Apifonica

The task was to test whether automation could work under the most challenging conditions — debts overdue by more than 90 days. This segment has the lowest effectiveness of live operators, making repayment unlikely.

The key question was simple: could a voice AI agent come close to human operator results — or even outperform them?

Finding the right partner was not a straightforward process. The market offered only fragmented solutions: some companies covered telecom, others dialog logic, and others analytics. To tie it all together, the client would have needed in-house expertise, which they didn’t have.

Apifonica provided an end-to-end solution: telecom infrastructure, Smart Dialer, dialog scenarios, and integration into debt collection workflows. Removed several barriers at once: faster launch, lower risks, and a focus on the main goal — testing the economic viability of automation.

The rollout happened in stages. First came technical integration, followed by a commercial launch to a limited customer base. The team paid special attention to the nuances of language by testing various Spanish dialects and adapting the speech specifically for the Mexican variant. While Apifonica is not a linguistic research institute, its specialists managed to make the AI agent sound natural and acceptable to customers.

Implementation

After integrating the infrastructure, the team proceeded to build scenarios. The idea was to test the effectiveness of automation at different collection stages. The AI voice agent was trained to handle several key tasks:

• Payment reminders — soft scripts for customers with recent delinquencies.
• Promise to Pay (PTP) confirmations — the agent reminded customers of their debts and secured explicit payment commitments.
• Deep delinquency (90+ days) — the toughest segment, where repayment chances are traditionally minimal.

Here’s an example of how a typical scenario for launching an AI agent looks inside the Apifonica interface. To make it less abstract, we’ll use a rather colorful case: a not-so-faithful wife racked up debts with the bank, and her husband ended up having to deal with the fallout:

Tone of Communication: Polite but Businesslike

One challenge was choosing the right tone. The Mexican microloan audience is sensitive. A tone that is too harsh could provoke negativity or aggression, while a tone that is too soft could seem weak.

The solution displayed a professional friendliness that was polite and free from threats or pressure, while also conveying a confident tone. Such an approach reduced emotional tension and fostered trust.

Smart Dialer and Extra Tools

To maximize reach, the Smart Dialer system was used. It adapted to customer behavior by:

• Changing call times.
• Adjusting call frequency.
• Using different geographic numbers.

If a customer ignored or rejected a call, the system attempted to call again later or used a different number more familiar to that region. The Smart Dialer also analyzed operator statuses: if a number was flagged or blocked, it was replaced automatically.

Built-in algorithms identified voicemail and used special bypass scripts. Plans include adding SMS with payment links during calls — simplifying the process so customers can settle debts immediately.

Early Findings

Even at the pilot stage, it became clear that the AI agent wasn’t just replicating operator work — it was reshaping the process. It handled large volumes in parallel, never tired, and didn’t waste time on empty calls. Most importantly, it didn’t trigger emotional resistance from customers.

Customer Reactions

A key concern was how customers would respond to AI calls. Mexican borrowers were used to human interactions, and a voice agent was something new.

Reactions were calmer than expected. Many customers even felt more comfortable: the agent didn’t interrupt, raise voice, or display emotions. Conversations averaged about 40 seconds — enough to convey the main message and get a clear answer.

Requests for Human Transfer

Some customers asked to speak with a live agent, but such cases remained rare. Out of 16,300 productive conversations, AI agents needed to transfer calls only 730 times, which accounts for 4.5% of the total number of calls. For comparison, the industry average stands around 10%.

This showed that the AI agent’s scripts were persuasive enough to keep most customers engaged in automated dialogue.

Handling Conflict and Emotion

One clear advantage was how the AI agent managed conflict. Customers who might interrupt or swear at human agents behaved differently: the AI didn’t react to aggression, didn’t lose focus, and calmly stayed on-topic. This often led customers to shift into a more constructive mode, reducing emotional strain.

Psychological Barrier

A fascinating insight: some customers hesitate to reveal their motives to a human operator — out of fear of judgment or subjective assessment. With an AI agent, these barriers were lower. Customers were more willing to admit, for example, “I can’t pay right now, but I can in two days.”

This not only simplified interaction but also gave lenders valuable information about the customer’s real situation — insights that human operators often couldn’t obtain.

Results of AI Debt Collection Automation

While customer stories matter, business results are ultimately about numbers. Comparing live operators with automation on the same workload highlighted key outcomes.

Call Volume and Processing Speed

The AI agent handled 80,000 calls. Of these, 16,300 (36%) were productive; the rest were voicemails, invalid numbers, or short connections.

For a live team, this would have been overwhelming. Even an empty call takes at least a minute, while a real conversation lasts three minutes. In total: 112,000 minutes, or nearly 1,900 hours. For 15 operators, that’s about a month and a half of nonstop work.

The AI agent completed the same workload in 2.5 business days, running batches of 1,000 calls every 15 minutes and working in parallel without human time constraints.

Time and Cost Savings

Time savings translated directly into money:

• A 15-agent team costs about $22,500 per month.
• Handling 80,000 calls would take 1.5 months, or about $34,500.
• The automated solution cost just $3,300 (project-based pricing).

That’s over $31,000 saved — plus freed-up staff resources for higher-value tasks.

Redials

In practice, customers rarely answer on the first try. Live agents must repeat calls, stretching campaigns to nearly two months. For the AI agent, redials were trivial: they added just one extra day, with no additional cost.

Conversion and Debt Recovery

The critical question: could the AI agent match operator results? It didn’t just match — it exceeded them.

The Promise to Pay (PTP) rate was 0.29%. At first glance, this seems modest. However, remember that these were debts that were over 90 days overdue, typically considered nearly uncollectible. By comparison, human operators achieved only 0.14%.

As a result, the AI campaign successfully recovered around $10,000 in payments.

Customer Behavior Insights

Automation provided not just efficiency gains but also new insights about customers — behavioral patterns that operators often missed.

Barriers with Humans

Some customers avoided explaining non-payment to human agents due to fear of judgment, criticism, or emotional responses. They would dodge the conversation, stay silent, or give one-word answers.

With AI, these barriers were reduced. Customers knew they were speaking with a neutral system, free of criticism or emotion. As a result, they were more likely to say, “I can’t pay now, call me in two days.” This was valuable information that companies previously lacked.

Tone and Dialogue Rhythm

Another takeaway: effectiveness depended less on script length and more on communication style. Short, simple, and polite scripts were more effective than lengthy explanations of contract terms. Customers engaged more quickly, remained focused, and made decisions more readily.

Emotions and Constructiveness

The AI agent had another key advantage: conflict resolution. If a customer was rude or emotional, the AI redirected the dialogue back to the topic. Eventually, customers adjusted to a more constructive conversation.

Let’s take a look at an example from another case. It’s in German, but that doesn’t matter — you can still clearly see that the customer is very dissatisfied and aggressive, yet the AI agent handles it successfully:

A New Level of Analytics

Every AI call is recorded and analyzed, creating a rich dataset. From this, companies can identify which phrases work best, when customers are most likely to respond, and which arguments increase the chances of repayment. Such analytics are hard to achieve with human-only operations.

The bottom line: automation not only cuts costs but also generates customer insights. These learnings help refine scripts and improve the entire debt recovery system.

Share of Automation and Future Plans

In this project, the AI agent handled 100% of calls. This “pure” test confirmed the technology could manage maximum workload. Results proved that AI can fully replace operators in the early stages of debt collection, delivering predictable outcomes.

However, full automation is not always practical. The optimal share is around 40–60%. This balance reduces operator workload while keeping live agents available for complex cases or customers needing personal attention.

The participating company set a strategic goal: reach 40% automation in year one. Internal analysis showed this is the level where automation delivers maximum economic impact:

• Mass calling costs shrink.
• Operators are freed for “warm” leads.
• Conversations are shorter, more precise, and more effective.

This is not about replacing humans with machines but about building a hybrid model: AI filters and handles routine tasks, while operators focus on empathy, flexibility, and individual solutions.

Thus, the project is shifting from experiment to strategy. Automation is no longer a pilot but a systematic tool for managing debt portfolios.

Conclusion: Automation Brings Order

• Scale and speed — What took a 15-agent team 1.5–2 months, the AI agent completed in 2–3 days, including redials.
• Cost savings — Human handling of 80,000 calls would cost about $34,500. Automation did the same for $3,300 — a savings of over $31,000.
• Customer insights — Automation reveals how tone and rhythm affect engagement, reduces psychological barriers, and uncovers hidden insights.
• Human + Conversational AI — Automation doesn’t replace people but amplifies them. The AI filters and manages routine volume, while operators focus on high-value interactions.

This hybrid model is the future of debt collection in the financial sector.

FAQ

How does AI debt collection improve recovery rates compared to traditional human-only methods?

AI agents can handle massive call volumes in parallel, work without breaks, and maintain consistent scripts. This scalability doubles recovery rates in hard-to-collect segments, where humans alone usually fail.

What impact does AI automation have on the customer experience?

Borrowers often perceive AI calls as less judgmental and more neutral. This lowers emotional resistance, encourages honesty, and improves the quality of information lenders receive.

Can AI debt collection fully replace human agents?

No. AI is best suited for repetitive, large-scale tasks. The optimal model is hybrid: AI filters and processes routine cases, while human agents focus on empathy-driven conversations and complex negotiations.

How does automation affect operational costs and scalability?

By replacing weeks of human work with days of AI-driven calls, companies cut costs by tens of thousands of dollars per campaign. At the same time, automation provides the flexibility to scale debt collection instantly, without expanding headcount.

What are the ethical considerations of using AI in debt collection?

Ethical AI deployment requires careful script design and tone calibration to avoid customer stress. Transparency, fairness, and data privacy are critical. Done right, AI reduces pressure on borrowers and makes the process less confrontational.