AI Debt Collection in the Philippines: Voice Agents

VoctivUpdated July 202613 min read

AI debt collection has quietly crossed from experiment to production, and the first call with a lender in Makati always opens the same way: “What collection rate do your AI agents actually get?” It is the right question, and the honest answer is that it depends – on the product, on how far the account has aged, and on whether you are a bank supervised by the Bangko Sentral ng Pilipinas (BSP) or a financing or lending company registered with the Securities and Exchange Commission (SEC), because the collection rules are not the same for the two. Underneath the “it depends” sits a pattern we have watched hold across dozens of lenders on four continents, the Philippines among them: a well-built AI debt collection agent collects more than your best human team, at a fraction of the cost, in the language the borrower actually answers in.

Calculate your savings

The first question every lender asks

Ask us for a single collection rate and we will not give you one, because it would be a lie. Collection performance moves with the product, the bucket and the borrower.

A pre-due reminder before the auto-debit arrangement (ADA) hits on payday behaves nothing like an account that is 90 days past due and already with a field collector. A “buy now, pay later” balance collects differently from a small-ticket digital loan, a salary loan, a credit card, or the monthly amortization on a motorcycle. Aging changes the job as much as the product does: current and first-cycle buckets are a reminder, a payment link and a nudge to re-fund the ADA or replace a post-dated check (PDC) that bounced, while late buckets are a negotiation that ends in skip tracing, a field visit, a demand letter, or an account endorsed to legal or remedial. Any vendor who quotes you one flat number is quoting a slide, not your portfolio.

What we can tell you is the pattern we see once an agent is live. Across the books our clients run on the platform, our AI agents recover roughly 10 to 20% more than the human team that worked the same campaign, and they do it for a small fraction of the cost per peso collected. Three things drive that.

• We train on your champions, not your script. Before an agent goes live, we study your real call recordings: the approved script on paper and, more importantly, how your best collectors actually talk. Top collectors improvise inside the rules. They read the borrower, switch between English and Taglish mid-sentence, and use the moves that work but never made it into the script. We distill the best of all of them into one agent, so every call runs to the standard of your best collectors, without the variance a floor has across a shift.
• It never has an off day. If the playbook says work an objection three times, a tired collector often tries once and wraps up. The agent follows the rule on every call, all day, including the 15th and the end of the month, when your queue spikes against a floor that cannot grow for two days. It also finishes the job: it confirms the amount, takes a dated promise to pay (PTP), and points the borrower to a channel they will actually use – re-funding the auto-debit arrangement (ADA) before the next cycle, GCash, Maya, or over-the-counter (OTC) at Bayad Center, 7-Eleven or Cebuana Lhuillier. Those small, consistent executions are exactly the kept PTPs a human floor leaves on the table.
• It reaches more people. Philippine unsecured books do not die from poor talk time, they die from contactability: numbers churn, borrowers switch SIMs, and a large share of the file becomes uncontactable within a couple of cycles. Smart dialing, disciplined retries and optimal-hour detection – tuned on our call data across many countries – lift connect rates to often two or three times what a standard call center sees, which is what feeds right party contact, and every attempt stays inside the window that applies to your entity type and that bucket. More conversations mean more pesos collected.

The result is an agent that costs about three times less per connected minute and per peso collected on the defaults in the calculator below, and recovers more. Cheaper and better, not cheaper or better.

Figures reflect Voctiv’s own results across live deployments, not audited external benchmarks; performance varies by product, bucket and market. See method.

Start with your cost to collect

Before we talk recovery, look at the cost side. Enter your own numbers – Quick for a fast estimate, Advanced for full control – and see how many times more it costs to collect a peso with people than with AI voice agents, down to the cost of a single minute of conversation. One honest warning before you press the button: a Philippine seat is among the cheapest anywhere, so the gap here is narrower than the one a US vendor will show you. It counts cost only, so the recovery gain sits on top of whatever it shows.

"We tried AI collections, and it didn't work"

If you piloted AI debt collection a couple of years ago and swore off it, we understand. That generation earned its reputation: stiff, scripted, and obviously a robot the moment the borrower said something off-script or answered in Cebuano.

That technology is already obsolete, and we would know - we lived through it. Voctiv runs the latest large language models, tuned on our own platform specifically for collections and hardened for real conversations rather than demos. Because the whole stack is optimized end to end, the agent answers with almost no delay, at the rhythm of natural speech, carrying intonation, accent and emotion.

On live campaigns the voice is natural enough that the call stops being a robocall and becomes a conversation. Borrowers explain why they missed the due date, argue about the amount, ask what a restructuring would look like - and the agent handles it the way a good collector would. They stay on the line instead of cutting the call in the first five seconds, and that is where the recovery actually comes from.

That realism is never used to hide who is calling. The Manual of Regulations for Banks (MORB) requires a collector, in-house or third party, to disclose full name and true identity to the borrower, and SEC Memorandum Circular No. 18, Series of 2019 (August 19, 2019) puts the same duty on financing and lending companies and their third party service providers in its Sec. 4, on top of making false representation or any deceptive means an unfair collection practice. So the agent opens by identifying itself as an automated voice agent calling for the creditor, in wording your compliance team approves. Recording works the same way: RA 4200, the Anti-Wiretapping Act, requires the consent of every party to the call, so consent to record is taken at the head of the call and logged as a timestamped event against the account before recording starts, separately from anything you rely on to process the borrower's data. Everything above is measured with those disclosures in place, not despite them - and because the models keep improving, we keep the platform on the current generation, so you are never collecting with last year's technology.

Hear a live example in Taglish or Cebuano

Where we run AI debt collection, and for whom

Debt collection is one of the most in-demand use cases on the Voctiv platform - and one of the oldest. We have been building conversational AI for collections since 2019, living through several generations of the technology from the inside: from the early scripted voice bots that gave AI collections its stiff reputation, to today's large language models that hold a real conversation. That long run is why collections is not a demo for us - it is a core part of what we build for lenders and fintechs.

We have launched live collection and reminder campaigns for dozens of lenders across the Philippines, Malaysia, Indonesia, Vietnam, India, Kazakhstan, Mexico, the United States, the European Union, the Middle East and Africa - each in the local language, not a translated approximation. Our clients run from classic, heavily regulated banks to the most modern fintechs building at the edge of the market. In the Philippines, agents we built and deployed for fintech lenders have covered the full lifecycle: identity verification and onboarding at loan origination, pre-due and due-date reminders, early collection from the first missed payment, and late-stage, deep-overdue collection.

We build these agents for any language and any country our clients lend in, not only this one. This page covers Philippine rules and Philippine economics; if you also run books in other markets, the global version - with the wider regulatory picture and the full method behind the numbers - is in our main article on AI debt collection.

• Local language, done properly. Native collection conversations in Taglish for Metro Manila and Cebuano (Bisaya) for Visayas and Mindanao books, plus other local languages where a portfolio needs them - alongside the languages US-built tools cannot serve well across Southeast Asia, the CIS, Latin America, the Middle East and Africa.
• On-premise or cloud, your call. Deploy inside your own perimeter when your security team or your Data Protection Officer requires it, or run from our cloud for speed and a faster go-to-market - here is how we handle platform security and reliability. Either way you remain the personal information controller under the Data Privacy Act of 2012 (RA 10173) and we operate as a personal information processor, on your instruction and under a data processing agreement. NPC Circular 20-01 as amended by NPC Circular 2022-02 sets the limits on how borrower data may be processed for collection, and the agent is configured to them: it dials only the numbers you supply, it never derives a number from the borrower's phone contact list, and for collection only a named guarantor is called. A co-maker is a party to the loan, not a contact-list entry. Your Data Protection Officer should confirm what registration, if any, our role attracts.
• Compliance built into the product. The contact window here is 6:00 a.m. to 10:00 p.m., and there is no single national exception to it. For a BSP-supervised bank the window may be lifted on an ordinary loan more than sixty (60) days past due (MORB Sec. 301); for an SEC-registered financing or lending company the threshold is fifteen (15) days (SEC MC 18); and for a bank credit card there is no past-due exception at all (MORB Sec. 312). The platform enforces the window per entity type and per bucket, not one setting for the whole book, together with your approved disclosures, prohibited-conduct guardrails, escalation to a human, and a tamper-evident audit log retained for the period you set. No Philippine rule sets a number of calls - a numeric cap is a United States rule under Regulation F, not a Philippine one - but repeated or oppressive calling is itself actionable as an abusive collection practice, and no call volume is a safe harbor. So contact caps are yours to set and ours to enforce.
• Owned telephony. We run our own carrier layer rather than reselling a generic international carrier, which is how we lift answer rates and keep caller ID consistent, so borrowers see the same number your institution uses for collections instead of whatever a wholesale route happens to present.

In-house, agency, or Voctiv AI: the honest comparison

How to find out if it fits

The next step is not a contract. It is a conversation.

Our sales team talks with qualified lenders and, on that call, does two honest things. We share the real cases and the actual metrics - what a portfolio looked like before us and what we moved it to - openly, and in the numbers you are judged on: right party contact, kept PTP rate, roll rate, collection efficiency and cost to collect. And we tell you straight whether we can take you on: whether your volume, your product and your bucket mix are the right fit for what we do. We would rather say "not yet" than sell you a pilot that will not land.

1. Leave your details in the form below - a few questions about your book.
2. We review your product, your buckets and your volume before we call.
3. We get on a call, show you what we have done for lenders like you, and check the fit both ways - including which contact window applies to your entity type.
4. If it fits, we scope a low-risk pilot on a single aging bucket, set up the way your team already tests vendors: champion/challenger against one accredited agency on the same bucket. And when AI works the early buckets well enough, fewer accounts reach endorsement at all, along with the agency commission that comes with it. We agree the comparison period up front, because December flatters everybody's numbers on 13th month pay.

Frequently asked questions

Figures on this page are for general guidance, not financial or legal advice. Adjust the calculator to your own numbers, and clear any collection process with your own compliance counsel.