Domu makes collections calls for the Americas' biggest banks
Every bank has a floor nobody brags about. It's the one where hundreds of agents sit with headsets, dialing people who are behind on a loan, reading from scripts written by compliance lawyers, and absorbing more hostility before lunch than most employees face in a year. The work is expensive, the turnover is brutal, and a single agent who improvises the wrong sentence can hand a regulator a lawsuit.
Banks can't skip it. Lending is the business, and money that doesn't come back isn't lending, it's a gift. So the industry keeps feeding one of its worst jobs: by Domu's framing, collections is a $20 billion industry built largely on people in chairs making phone calls.
Domu, a San Francisco company out of Y Combinator's Summer 2024 batch, builds AI agents that do this work instead. Its agents make and take collections and loan-servicing calls, and carry the same conversations over email, SMS, and WhatsApp, in Spanish, Portuguese, and English. According to its YC profile, Domu now automates debt collection conversations for 8 of the 20 largest banks and insurance companies in the Americas. Two years ago the company didn't exist.
The job nobody wants
It helps to be specific about why collections is such a miserable function to run, because the misery is exactly what Domu is selling against.
Start with cost. A delinquent portfolio needs volume: thousands of outbound calls a day, timed to the hours regulators allow, repeated across weeks. Human call centers scale that linearly: more delinquency means more hiring. And collections hiring is a treadmill, since burned-out agents leave fast and new ones start out bad at a job where being bad is costly.
Then there's compliance, which in collections isn't a department so much as the weather. In the US, the FDCPA constrains what a collector can say and when; the TCPA constrains how you're allowed to dial; the CFPB watches for anything that looks unfair or abusive. Every large lender in Latin America operates under its own version of this regime. A human agent having a bad day is a compliance incident waiting for a recording. Banks spend heavily on QA teams that sample calls after the fact, which means most calls are never reviewed at all.
And underneath both problems sits an awkward truth: the person on the other end usually isn't a deadbeat. They're someone who lost a job or mistimed a paycheck, and who screens calls precisely because the calls are unpleasant. An angry collector doesn't collect more. A consistent, calm, infinitely patient one often does.
That last point is where software stops being a cost story and becomes a performance story.
What an AI collections agent actually does
Domu's product is an AI agent that handles the delinquent-account conversation end to end: it reaches out on the channels the borrower actually answers, explains where the account stands, negotiates within the rules the bank sets, and books the payment or the promise to pay. Voice is the hard part and the core of the pitch, since collections has always been a phone business, but the same agent persists across email, SMS, and WhatsApp, which matters in Latin America where WhatsApp is where financial conversations actually happen.
The company is candid that this is not a plug-and-play voice bot. Deployments are integration projects: the agent has to sit on top of the bank's core systems, know the real account state, follow the bank's negotiation policies, and hand off to humans where the bank draws that line. That honesty reads as a limitation until you remember who the customer is. Banks don't want a widget; they want a system of record they can defend in an audit.
Compliance is where the automation argument gets interesting, because an AI agent flips the QA math. A bank sampling human calls reviews a few percent of them. Domu's pitch is that every interaction can be analyzed against rules like the FDCPA and TCPA, every time, with the transcript already structured for review. The script never gets improvised in the heat of an argument. The agent makes the disclosure every single call. Domu has leaned all the way into this positioning: the company describes itself as the first fully licensed AI collections agency in the United States, taking on the regulatory status that collection agencies themselves carry rather than hiding behind "we just make software."
Proving it at Nubank scale
Enterprise AI pitches are cheap. Production traffic at a major bank is not, and this is where Domu's story separates from the crowd of voice-AI startups.
The flagship is Nubank, which Domu's case study calls Latin America's largest bank. Domu's agents went into production there in 2025, handling customer outreach in Portuguese and Spanish, and the company counts more than 1.6 million live AI-handled conversations across Brazil and Colombia. Across all of its clients, Domu says its agents have processed over 73 million calls. Vendor-reported numbers deserve some skepticism, but volume like that is well past the pilot stage. It describes a system a very large bank tried and then decided to keep.
Banco BV, one of Brazil's largest privately held banks, is the more revealing case. The collections agent it built with Domu became the bank's first major agentic AI initiative, and the rollout was careful enough that Banco BV's Responsible AI function grew directly out of the project. Flaviana Sallowicz, the bank's head of collections, has said the AI significantly reduces operational risk, and that it removes most of the cost base in traditional collection processes. Collections, the unloved back office, turned out to be the door through which agentic AI entered the whole bank.
Then in June 2026 came a different kind of validation: Alorica, one of the world's largest customer-experience outsourcers, announced a partnership with Domu for AI-driven loan servicing, reporting strong early outcomes. BPOs run the human call centers that AI agents are supposed to replace. When one of the biggest decides it would rather resell the technology than compete with it, the direction of travel is hard to miss.
The sales motion is the moat
Plenty of startups can demo a voice agent that sounds human. Almost none of them can get that agent certified, integrated, and live inside a top-20 bank. Domu's actual edge is that it figured out how to do the second thing repeatedly.
Part of the answer was starting in Latin America rather than treating it as an afterthought. The region has enormous consumer-lending volume, high delinquency-management costs, and banks that have proven willing to adopt AI faster than their US counterparts. It also has a language requirement, native-quality Portuguese and Spanish on the phone, that US-focused competitors mostly can't meet. Domu built for that market first and let the logos compound: Nubank and Banco BV make a better sales deck in São Paulo and Mexico City than any benchmark could.
The early economics suggest how well the motion worked. By the time CEO Nico Díaz sat down with the Startupeable podcast in 2025, he described a company selling about $1 million a year with a team of five. Five people, selling AI into banks, one of the slowest-buying customer categories on earth. That number says less about revenue than about repeatability: the deals were closing faster than the headcount grew.
And collections is a wedge, not a destination. The same agent that can negotiate a past-due credit card payment can handle loan servicing questions, renewal outreach, and the broader customer-operations work Domu already lists in its product. Once a bank has done the hard part, the integration and the compliance review, every additional conversation type is an expansion, not a new sale.
Two founders who picked the unglamorous problem
Domu was founded by Nico Díaz and Camila Zancanella, who met while Zancanella was finishing her MBA at HEC Paris and Díaz was running his previous AI company. The pairing maps onto what the company has to be good at: Díaz had already built with generative AI before it was the default choice, and the company's whole posture, licensing, QA, careful enterprise rollouts, reflects an operator's respect for how regulated institutions actually buy.
What stands out most, though, is the choice of problem. In 2024, most AI founders with voice technology chased restaurants, receptionists, and sales calls. Domu picked the category where the customer is a bank, the regulator is always in the room, and the end user is a person having one of their worse weeks. The company's own writing frames the mission around trust and dignity in an industry that historically offered neither, and its deployments back the framing: the agent that never gets frustrated and never goes off script is a better experience for the borrower, not only a cheaper one for the bank.
Hard markets repel competitors for the same reasons they repel everyone else. Collections is adversarial and reputationally radioactive, and every conversation happens under a regulator's gaze. That is precisely why the startup that mastered it is now sitting inside some of the largest financial institutions in the hemisphere, answering the phone.
