Before Diego Avellaneda asked artificial intelligence to negotiate a debt, he asked electronics to survive a hotel. Hadron Lab, the hardware company he started with Daniel Martínez more than nine years before Callbook AI, installed electronic access systems at Casa Fundadores in Villa de Leyva and in attractions at Bogotá's Mundo Aventura. The work had a useful intolerance for abstraction. A hardware bug was never merely a ticket. It was a car ride, a panel removed from a wall, a client counting downtime, and two founders discovering that a tiny fault could acquire mileage.
Avellaneda later remembered the lesson as “commitment to real operations.” It is an unfashionably solid phrase. Software encourages the fantasy that every problem can be patched from somewhere with decent coffee. Hardware knows the customer is standing beside the broken thing. Callbook, his current company, sells a very different machine, but it still operates where errors become personal. Its AI agents locate borrowers, contact them through calls, texts, and email, and negotiate what happens next.
Collections is an awkward theatre for artificial intelligence. There is money, embarrassment, legal process, institutional pressure, and a human being on the other end who may have an explanation. The voice cannot merely sound fluent. It has to choose timing, tone, context, and the moment to hand the conversation to a person. Avellaneda has spent the better part of his career collecting the odd assortment of skills this requires.
A soldering iron, a lunch app, a research lab
The apprenticeship was not arranged in a straight line
After Hadron Lab came Almuerza.co, a delivery product for neighborhood restaurants. Avellaneda and Helman Andrés Bohórquez approached it with the engineer's familiar temptation: make everything correct before it meets the world. His later summary of the lesson had no ornament at all - “sales, sales, sales.” The difference between a clever system and a company, he found, is that someone must urgently want the clever system.
He also worked as a senior software engineer at CI&T, where his Y Combinator biography says he helped build Audi's e-commerce car-purchasing platform. At Pontificia Universidad Javeriana, he pursued a master's degree in the Internet of Things and applied TinyML to real-time truck telemetry. TinyML is the craft of running machine learning on small, constrained devices. It rewards thrift: less memory, less power, less delay, no room for a model that behaves like an aristocrat traveling with twelve trunks.
In 2023, Avellaneda co-authored a paper on indoor asset tracking that placed deep learning near the sensors doing the tracking. He was credited across the work, from conceptualization and software to data curation and the original draft. A second peer-reviewed paper credited him for software in a compact, low-cost system that measured soil permittivity on site. One project located assets. The other made field measurement portable. Both asked computation to leave the comfortable center and become useful at the edge.
That sequence became a personal motto. The phrase is not anti-AI. It is a rebuke to the habit of putting a model before the problem. Avellaneda's maker profile offers the charming early evidence: a ring that changes color as a meeting approaches its end, an Alexa-connected exercise timer, Wi-Fi light switches, a game that assigns colors to players. These are small inventions with an obvious user and a physical consequence. The whimsy helps. Engineers are easier to trust when they have occasionally built something simply because it made an evening more fun.
The company found its accent before it found its category
Callbook began after Avellaneda and Martínez searched online forums for recurring operational complaints. Lead qualification and appointment scheduling appeared as stubborn bottlenecks, so they built voice automation around them. The early company ranged widely. Its published notes mention campaigns in education, ecommerce, clinics, churches, dropshipping, amusement parks, and collections across more than 15 countries.
This variety became a laboratory for the opening seconds of a call. Across more than 50,000 calls in Latin America and Spain, Callbook reported that generic greetings could produce immediate rejection rates as high as 48 percent. Adapting the greeting to the country and cultural setting reduced first-five-second drop-off by as much as 35 percent. The differences fit inside a breath. “¿Bueno?” for Mexico. “¡Aló!” for Colombia. “Hola, ¿qué tal?” for Spain.
The first five seconds
In a collections and post-registration example published by Callbook, localized greetings reduced immediate rejection from 42 percent to 27 percent. The 48 percent figure is the high observed for generic greetings across a broader outbound set.
A mispronounced bank or retailer can puncture the illusion even faster. Callbook's analysis found that mangling an important word in the opening could lift the immediate hang-up rate by as much as 22 percent. So the company built custom pronunciation and tuned accent, pace, pauses, and emotion by use case. A collections voice should be measured. A sales voice can carry more lift. A lead-qualification voice ought to sound curious. “Human-like” is not a voice in a dropdown. It is a series of local decisions.
The million-call field notebook
In April 2025, Avellaneda published three lessons from more than one million calls. They are wonderfully short on futurism. First, redial with commercial intelligence. Second, transfer to a human with context. Third, screen early so the next attempt is better informed. The recommendations read less like a manifesto than notes taped above an operations desk.
Return intelligently
Space attempts, rotate thoughtfully, and call when a person is more likely to be available.
Keep the thread
Give a human agent the conversation history so the customer does not have to begin again.
Use every answer
Classify the response, adapt the next message, and let an unsuccessful call improve the campaign.
The first attempt in its outbound data contacted about 30 percent of a list. Five well-distributed follow-ups could bring cumulative contact to 50 or 60 percent. The extra attempts only worked when they respected the shape of a day. Avellaneda highlighted 11 a.m., 2 p.m., and 5 to 6 p.m. as strong windows - the small pauses around lunch and the end of work when humans briefly become available to institutions.
The handoff matters just as much. A transfer without context makes the customer repeat a story and leaves the agent guessing. Callbook sent the prior exchange into a CRM, WhatsApp, or SMS and kept the transition inside the same call. In campaigns with medium-to-high transaction values, the company reported a 40 percent increase in close rates and more than 60 percent less human handling time. Automation earned its place by preparing the conversation, not by pretending people were unnecessary.
By 2026, Callbook's website reported the scale above. The company had also become more specific. It now presented itself as an AI collections agency, using calls, SMS, and email as one continuous conversation. Each account could be segmented by ability and willingness to pay. The next action might be a payment plan, product support, a debt review, or a route to a human. The purpose was no longer merely to make calls cheaply. It was to decide what a particular account needed next.
A local problem walks into a global accelerator
In June 2026, Callbook joined Y Combinator's Summer batch. Avellaneda framed the acceptance through geography: a collections problem rooted in Colombia and Latin America had global relevance. The company's San Francisco chapter did not erase the earlier one. Callbook's regional experience was the product knowledge - how Spanish changes by border, how institutional language can sound cold, how debt conversations require both persistence and room for explanation.
The YC version of the company said it was negotiating with more than 100,000 people each week. Avellaneda and Martínez arrived with complementary histories. Avellaneda brought embedded machine learning, research, software, and the scar tissue of two earlier companies. Martínez, a former product leader at Treble, brought years of B2B product work. They had already been partners when a hardware failure meant getting in the car. Now the breakdowns are more subtle, but the principle survives: stay near the failure until you understand it.
There is a temptation to describe voice AI by how uncannily it speaks. Avellaneda's work is more interesting when measured by how carefully it listens. A restaurant employee says the manager returns at four. A borrower disputes the balance. A lead asks to be called tomorrow. Those responses are not obstacles between the system and its goal. They are the data that should alter the goal's next step.
“Every call has a purpose: sell or learn,” Avellaneda wrote before Callbook's tighter collections focus. In debt, the first verb may change, but the second remains. Learn why a person has not paid. Learn which channel works. Learn whether the product itself failed. Learn when a human should enter. It is a builder's answer to a social problem, appropriately modest and demanding: the machine gets another conversation only if it paid attention to the last one.