Konfío: How to Rebuild a $1,300M Unicorn with AI

The audio and video are in Spanish. These notes are a translation.

42% of the Mexican economy is small and medium-sized enterprises (SMEs), and those companies provide 70% of the country’s formal jobs. Mexico is the 15th largest economy in the world, but credit penetration in that group sits below comparable countries in the region.

Konfío was founded thirteen years ago to close that gap, and today it is on its way to becoming a bank. Francisco Padilla is its cofounder: he started as CTO and ended up moving through operations, people, and growth.

Before that he lived through the 2008 crisis in New York on a work visa, was the technical architect for his company’s largest account at Citibank, and worked in the AIG unit that was the epicenter of the collapse. The conversation goes from there to a concrete question: if he had to rebuild Konfío today with artificial intelligence and a blank whiteboard, where he would start.

Structure and Key Highlights

The Antenna His Dad Canceled

• At ten, his parents canceled the television signal and did not replace it with anything
• The only thing left in the house was books, in particular a couple of encyclopedias for kids
• They spent four or five years rotating through those books because there was nothing else to do
• Now that he is a dad he repeats the pattern: he tells his kids that books have no budget cap
• He grew up in León, Guanajuato, Mexico’s fifth or sixth largest city, with two parents who were systems engineers and technology as a topic at the table at lunch

The Tec and Its Silicon Valley Effect

• He did his first year at the León campus and then decided to go to Monterrey, on his parents’ condition that he get a scholarship
• The jump in level was steep: the books he studied in León were written by his professors in Monterrey
• That is where he saw the pattern he calls the Silicon Valley effect: generations of graduates went to the United States, and today they are the ones hiring the next ones
• The last course before graduating was to build your own programming language, and the program director would write a program that had to compile and produce the correct result

What It Costs to Learn How to Do Business in the United States

• Mexicans tend to be less direct: you ask through favors and you navigate social networks, what in Mexico they call mano izquierda
• In the United States the dealing is more about instructions. In his training they taught him BLUF, bottom line up front
• New York is the extreme: a competitive environment, an expensive city, a breakneck pace, nobody has time for you
• And a structural difference: in the United States a lot of the infrastructure is already built, while in Latin America you have to build it yourself in order to deliver your final service
• That forces you to create a couple of core skills if you want a differentiated product, or to settle for what is already there

Buy vs Build, and Why Switching from One to the Other Hurts

• Konfío leaned hard into build at the start, above all because there were no resources to buy
• Not everything was SaaS back then either: there were annually renewable contracts instead of subscriptions, uncomfortable for a young company
• Identity verification services worked very well in the United States and not in Mexico, so ten years ago they built their own version adapted to local conditions
• The rule he uses is Jeff Bezos’s: focus only on the things that make your beer taste better
• What they learned as they grew is that the transitions between buying and building are painful when the service is already running: there are SLAs in the middle, and you pay twice, for the current system and for the investment in the new one

New York, 2008, on a Work Visa

• He moved to the United States in 2006 and arrived in New York in early 2008. Almost as soon as he got there he met David Arana, Mexican like him, then at Deutsche Bank
• He worked at MicroStrategy, in business intelligence, whose largest client globally was Citibank
• The message consultants received in 2009 was to be cost neutral: if you generate expense for the company, you probably do not last
• He saw offers that were issued and then withdrawn, and people who were thanked and let go. On a sponsored visa, losing the job gave you thirty days to leave the country
• That pushed him to reason from first principles: always have a client, and produce at least what you cost the organization
• They would call you on a Friday so that on Monday you would be in Ohio, Oregon, or Atlanta, in an industry you did not know, because someone was paying for you. Asking for a raise under those conditions was unthinkable

From Architect at Citibank to the Epicenter of the Crisis

• Between 2010 and 2011 the senior teams rotated and he ended up as technical architect for the company’s most important account globally
• It was commercial banking and corporate cards for Fortune 500 companies, with tens of thousands of employees, across multiple countries and currencies
• He admits that at that moment he saw a lot of numbers without understanding why they mattered so much to those organizations
• His last job before Konfío was at AIG, in Financial Products, which was ground zero in 2008
• He arrived years after the bailout, and there he learned about portfolios, realized and unrealized losses, and how an insurer puts premium money to work while making sure the asset stays healthy

The LinkedIn that Broke His Mental Map

• He saw on LinkedIn that David, the most successful professional he knew, had just quit to start a company
• His reaction was literal: in my mental map this does not exist, I was taught to work at the best place and stay there my whole life
• He wrote to him, David told him he had moved back to Mexico to launch online credit for businesses, and pitched him the idea
• That conversation was thirteen years ago and, he says, it is the same spirit that pushes them today

Why Konfío Exists

• 42% of the Mexican economy is SMEs, and they provide 70% of formal jobs
• Mexico is the 15th largest economy in the world, and credit penetration in that segment is below comparable Latin American economies
• The segment is so varied in industries, regions, and products that it is impractical to serve it without a process that is one hundred percent technological
• From the MVP they decided it would be a parametric process driven by artificial intelligence
• The split was clear: David set up the company and the data decision model, and he built the amortization engine, the back office, the database, and the analytics
• Nine months in build mode before going live

The Double Life, and the Family Against It

• Two years of Wall Street by day and a startup by night, with sixty- to seventy-hour weeks as the standard
• In photos from that period he has extra weight and dark circles under his eyes, but he says he was happy and motivated
• Neither his wife nor his parents agreed with leaving that job. He was already a dad, and his visa did not give his wife permission to work, so two people depended on him
• The family’s argument: you made it to a place a lot of people would dream of reaching, and you are willing to burn it
• His counterargument: I look at the older people in the office and I know exactly where I am going to be and at what point in my life. If I jump this way, I do not know where I will be in five years

What AI Made Trivial

• Things that used to require an OCR, a data scientist, or four dedicated engineers became trivial in a striking way
• If you know how to be wise with your tokens and your platforms, you build in a short time things that five years ago were disruptive
• His simplest example: he and David lived on Stack Overflow, and Stack Overflow became irrelevant
• Now you have a developer who knows the correct syntax and who, if they get it wrong, self-corrects

Hallucinations Are Real, and His Example Is Precise

• The models are spectacular, but their retention is short: if you build fast, you run out of context
• When that happens, the model compresses what it had into free text, and that is where the numbers degrade
• The concrete case: a 0.3 turns into “a little bit,” and in the next step “a little bit” turns into 0.25, so the result changes
• The lesson is to put guardrails beyond the technical ones or the security ones, so the process is repeatable
• That is why they mix pieces of deterministic technology with powerful LLMs instead of leaving everything to the model

If He Had to Rebuild Konfío Today

• He would start with the decision models: building them got faster and became much more precise
• Not only credit, also propensity, and anything that can be predicted
• In the 2021 wave the directive was to protect data scientists as much as possible. They are still valuable, especially those who know how to work with AI, but the barrier to entry fell
• That opens questions people did not ask before: who can be my customer, who can I sell more to, what signals tell me that someone needs me
• His conclusion is that the intelligence of the organization is what is crucial, because it gets you closer to what the customer is looking for

Hiring Engineers Is a Different Thing Now, and So Is the Interview

• They had a timed exam of twenty-four hours, with an extension up to seventy-two: build this, with these outputs, and you figure out how yourself
• It was for a profile so rare that searching for it on LinkedIn returned three results in Mexico. Out of a hundred interviews they kept five or six, and that was the core team for years
• Today the recruiting, the tests, and even the product those people used to build are different
• The full stack engineer was a unicorn, and now it is easy to be one, but in exchange you have to know architecture
• The debate about whether to hire QAs or not is shifting: a well-fed QA process is repetitive and can become an agent. The same applies in sales, customer service, and collections
• The new question in interviews is how well you command AI, and it works like “do you speak English?” in Mexico: everyone says yes until you switch the conversation. So he asks for examples and for the most different thing they have built

Where They Have Seen the Return

• In data models they have seen five times more productivity
• Things that used to sit in the backlog for several sprints or quarters now come out in batches of weeks
• That changes the ambition: if you can make predictions at that level, you consider parts of the business you used to rule out for lack of inputs
• Sales is the hard part, both with humans and with AI. The opportunity he sees is building better tools for the team
• There is information that used to be hidden and today is obtained in a scalable way, which empowers the salesperson. The concept of a sales qualified lead mutated, and the bar is different

From Credit to a One Stop Shop, and to a Bank

• The core is working capital credit with no collateral, in amounts of up to ten million pesos
• An IDB study measured that their customers grow 20% in sales after a year, and 40% when they are women
• The problem with scaling only that is that the customer eventually graduates to another bank, so you have to surround them with products in order to be a destination and not a stop along the way
• They have already launched a credit card and a payments solution, and they see needs in accounts receivable and accounts payable
• They serve three segments: the independent professional, the SME of ten to thirty people that he calls an awkward size and that is their core, and the mid-market of fifty employees and up
• Deposits and investments are the natural next step after ten years in the market, with customers who have been with them for eight or ten years
• Selling to that segment is not enterprise, where five contracts make your year, nor consumer, where you shoot and iterate. It sits in the middle, it has stakeholders even if you deny it, and if you throw humans at it you lose the economics of the model

The Big Blocks: Cloud, Networks, Mobile, Now AI

• The building blocks of the last twenty years all had a before and an after in the way business gets done
• He tells of an audit where they told him he was not going to be able to do something in the cloud, and his answer was that that process only exists in the cloud
• He expects the same to happen with LLMs: you ask me for this without LLMs and it is not going to work
• His underlying bet: in every industry there are mountains of unstructured data, and the dominant fintechs of the future will be the ones that learn how to turn that into value for the customer

Growing at Double Digits Without Raising Spend

• It has been profitable for two years, so OPEX is not taken lightly, and that discipline also helps as they face the banking license
• The question he asks area by area is not whether the area gets eliminated, but what percentage can be done another way: customer service, collections, marketing content
• There is no single metric for everyone. For engineers it is harder to measure than in other areas
• His analogy: nobody asks for a single recipe for a cloud stack. Each company designs its own in order to scale. With AI it is the same. Each company defines what success is for itself, and then maintains it, improves it, and moves on to the next challenge

How AI Gets Pushed Inside the Organization

• They run hackathons
• They review license use in both directions: you open one and nobody uses it, or you have to shut it down because someone is consuming too much credit
• They have education forums and screen sharing sessions where someone shows what they built and what benefit it brought
• In those sessions they teach the concrete stuff: how the prompt is written, whether it runs offline or on demand, and at what alert level. That is what makes people lose their fear
• The models come in tiers: the basic version is open, and for the more expensive ones you have to bring a justification, and afterward they review what you did with them

Tokens Are Going to Be a Line on the P&L

• It already happened with the cloud: one or two years on the startup free tier, and today it is one of the organization’s largest vendors by spend
• With tokens the difference is that the spend adds up fast and gets expensive fast, so you have to be more judicious
• His internal line: you do not want to use a bazooka to kill an ant. Do not put the most powerful model on giving you the monthly update
• Managing models is starting to look like managing a team: you do not hire or compensate the whole organization as if it were C-level
• Organizations with a good AI architecture are going to have agents in different layers. When a new model comes out you gain a gear, but that gear can cost double or triple, so you have to watch who uses the latest generation and for what

The Banking License and the Fear of Becoming Slow

• The license brings trust, more funding, and more products, and on the other side regulation, compliance, and reporting
• Building with the customer in mind steers them toward platforms that are not necessarily the standard the market’s players use today
• Those bets force a rearchitecture of the platform, and that is a commitment measured in months and sometimes in years
• His anecdote about agility: a person who came from supervising banking institutions arrived and built agents that do her own job
• Before, those decisions got stuck in meetings to define whether something was inside compliance. Now she does the work with the tool and only escalates when the parameters break

What He Would Tell People About to Graduate in León

• If someone wants to leave and they pull it off, welcome. But his alternative at the time was to repeat the pattern of his parents’ generation, and he decided not to settle
• The barriers to entry fell: when it was his turn, YouTube did not exist, nor Spotify, nor podcasts
• Today someone can build remarkable things from home at a completely different cost
• The big digital bets of his era were the consultancies that served corporates, and that was fine because it was what existed
• Now any country has its unicorn or its publicly listed company, and that means the experience can be lived locally

Who He Thanks

• Karla, his wife, twenty years together. She was with him in the good times and the bad, and when he found his limits she took care of him too
• David, his partner, whom he admires for his judgment, the way he acts, and his principles
• And his parents, who left him principles, an education, and the initial resources, and gave him the chance to dream
• His close: the book is still open and we are still writing some pages

Silicon Valley Glossary

People, Companies and References

People Mentioned

• Francisco Padilla: cofounder of Konfío. He started as CTO and moved through operations, people, and growth
• David Arana: his partner and cofounder, whom he saw quit on LinkedIn
• Karla: his wife, twenty years together, the first person he thanks
• Jeff Bezos: the line about focusing only on what makes your beer taste better
• Enzo Cavalie: host of the podcast

Companies and Organizations

• Konfío: credit for Mexican SMEs, now on its way to becoming a bank
• MicroStrategy: where he worked in business intelligence when he arrived in New York
• Citibank: that company’s largest global client, and the account where he ended up as technical architect
• AIG: his last job before Konfío, in Financial Products
• Deutsche Bank: where David was when they met
• El Tec: first the León campus, then Monterrey
• IDB: the Inter-American Development Bank measured the sales growth of its customers
• Stack Overflow: what stopped being relevant when the models arrived

Data and References

• 42% and 70%: how much of the Mexican economy is SMEs, and how much of formal employment they provide
• 15th largest economy in the world: the size of Mexico set against its credit penetration in SMEs
• 10 million pesos: the ceiling on its working capital credit
• 20% and 40%: how much their customers grow in sales after a year, and how much when they are women
• 5 times: the productivity they gained in data models
• 3 LinkedIn results: how rare, in Mexico, the engineer profile they were looking for was
• 100 interviews, 5 or 6 hires: the result of the twenty-four to seventy-two hour exam
• 9 months: how long they spent building before launch
• 13 years: how long Konfío has been going since that first conversation
• 2 years: how long they have been profitable