7 Lessons for Building AI Startups in 2026

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

Today I had three technical guests, which is the reason for this episode: I wanted a conversation about what it really means to build with artificial intelligence, not about the narrative.

Damián Schenkelman was employee number 10 at Auth0, one of the most successful Latin American startups, acquired by Okta in 2021 for more than 6 billion dollars. Today he runs his research and development lab.

José Noblecilla leads technology and artificial intelligence at Valor Capital, with more than 2 billion under management. Before that he was founder and director of AI at Wildlife Studios.

Felipe Sateler is founder and CTO of Buk, the largest human resources management platform in the region, which processes the payroll of more than 4,000 companies and is valued at more than 800 million.

Structure and Key Highlights

What Really Changed in 24 Months

• Felipe describes himself as moderate: he does not see something that changes things radically, but he does see a real transformation in internal productivity
• What they could not do before: analyze data at those volumes, categorize information, generate clusters
• His number: in about six months they raised productivity 70%, measured in pull requests merged per month per person
• Damián adds the nuance of context: at Auth0 they already had machine learning models in production for years, for anomaly detection on login
• What they did decide is to wait. There was a moment when every product had to have an AI assistant and they preferred to see whether something more interesting would appear
• What changed the most for his research team: a lot of the code they generate is disposable, so they can go faster
• His important nuance: they work on a codebase of many years and full of edge cases, so the difference between models shows a lot. A jump in model substantially changed the experience
• José, from the investor side, sees two archetypes: customer-facing AI products and back office automation
• And an asymmetry that surprises him: he sees more speed adopting AI in the product than internally, because the back office runs into governance

Startups Against Large Companies

• Damián sees the opposite in very large companies: an aversion to the risk of building a product with AI
• First it was the licensing of the code, then who you hand your code to. Companies that could run in your own cloud won because of that
• What he does watch with a lot of interest: internal MCPs, to access data with auditing
• In startups, governance and security do not matter as much, so they go much faster and they test products faster
• The data point that anticipates a trend: someone passed an audit of the new ISO standard on AI systems, and in the first year the auditors did not really know what to audit. The second year was very different
• His business prediction: just as Vanta appeared to solve SOC 2, companies are going to appear that solve compliance and even insurance for AI systems

Why MCPs Do Not Solve the Problem on Their Own

• Felipe reduces it to the usual thing: who has access to what information and who can modify it in an authorized way. The only new thing is the interface
• His critique of off-the-shelf MCPs: they are not going to have the granularity to validate permissions
• The concrete example: if you hang off Google Drive permissions, you have to give access to everything. If you do it with a service account, you are giving it access to things it should not see
• Damián agrees that using a service account to connect the MCP to the API is outright a security failure
• In fact he keeps a site where he notes down the security problems he finds in AI applications. He shows it to prospects and the product sells itself
• His clarification: that is not inherent to the protocol, it is bad programming
• What he does believe adds something: remote MCPs done well, oriented to concrete flows instead of exposing an API. He calls them composite actions
• The consensus of the three: after a year and a half it is still necessary to program your business logic, your protections, and your guardrails
• José’s conclusion, with humor: I still have to do my job. And he adds that the trail of who did what and why is increasingly relevant, because the model executes actions

Where the Business Impact Shows Up

• Buk is investing in acquisition, where generated content has been a fruitful channel, and it has an assistant for the sales process
• What has curiously cost them more is support, because it requires a very good and well-formatted knowledge base
• The dilemma that creates: you do not see the impact in the short term because there is a lot of prior work, so you postpone it, and if you postpone it too much you lose the value
• José points out the measurement problem: it is not clear which KPIs you measure when you have an agent in production
• How they are reporting it today: as a one-to-one parallel with employees. A collections agent against a collections analyst, comparing each task of the role
• His game theory warning: you lower the cost of your payroll, but if everyone uses AI the cost of user acquisition goes up for the whole market
• Damián sees the impact where the product saves time for the customer’s team. Testing that takes less time, requirements gathering that goes from months to weeks
• And his observation about where the relative impact is greater: in startups, where it matters to move fast and nothing happens if you accumulate some technical debt

The Biggest Design Change: API First and Agent Experience

• For Damián the most important thing is to assume that an AI is going to consume your product and to redesign from there
• His guiding questions: if I had to use a CRM from a chat, what would it be like? If I had to buy, what would it be like?
• And the other thing that changed from one day to the next: showing up in the answers of the models
• Felipe confirms it with data: the opportunities that arrive through paid marketing fell sharply and the ones that arrive through agents rose a lot
• The threat he sees for niche verticals built on top of another platform: they are the easiest to replicate
• His frame: what is your center of gravity, what makes it hard for your customer to switch and easy to hook processes on top
• His example: Shopify. You set up the store in three clicks and then your whole commercial life lives there
• And his own case: at Buk the center of gravity is the company’s employees. Every hire, exit, or role change triggers processes, and they are the authoritative source of that information

The Agent Builds on the SaaS, It Does Not Replace It

• Felipe’s position: people are going to operate on top of the SaaS, because the authoritative and well-labeled data has to live somewhere
• An Excel does not work, because it gets messy in 30 seconds
• His two examples: the purchasing agent does not replace Shopify, it builds on Shopify. The sales agent integrates with your CRM
• José adds the building blocks he sees emerging: boards where the AI creates the tasks the human has to complete and guides them step by step
• The second: how you give the agent feedback. Today it is thumbs up or thumbs down and some text, and he finds it clearly insufficient
• The third: operational metrics per agent. How much it collected, how long it was active, what it delegated to the human
• The frustration Damián describes as a user: he uses a product that sends him things that do not match his filter, he corrects it several times and it does not improve
• What is missing there, in his opinion, is a debugging view for the end user. The thumbs down is too opaque
• And a question he has open: what is an agent’s inbox. How it communicates with the outside world and with other agents

Building Products with AI Is a Different Discipline

• Felipe separates two ideas. One is old: keeping something in production was always much harder than making the prototype
• The other is new, and there he strongly agrees: the key word is reliability
• Traditional software is predictable. If it worked well, it works well always. With probabilistic models that breaks
• And the space of inputs is infinite. What you can write to a chat is infinitely more than what fits in a form with a numeric field
• The economic consequence: running the product is more expensive, in the execution and in the monitoring, and you need people dedicated to watching how it behaves
• His conclusion: the gross margins of a full AI company should be much lower
• What José points to as expensive on the inside and a commodity on the outside: keeping the logic of a human in the loop alive, verifying actions and measuring whether the human did them
• And where the opportunity appears: understanding at what moment it makes sense to do fine tuning of an open model to lower costs and gain a moat
• His day zero recommendation: log inputs, outputs, and what you expected. That history is what later lets you fine-tune or understand the user better

Which Technique Is Worth It and Which Can Wait

• Damián describes it as a new hat. Before you put on the reliability one and the security one. Now also the data one
• The concrete work: see where the system fails and lower the number of failures per typical execution
• One of the techniques he mentions is prompt optimization, even automated, deciding which model to send each thing to
• But where he sees the true cost in B2B is in data engineering: cataloging, labeling, and structuring the internal information of each company
• His formulation: you have to solve the context for the specific domain, and you have to have that data no matter how the models evolve
• The second thing that can change everything is the interface. His example is Granola: the best part is that it is not invasive, it takes the notes and summarizes. The non-interface is what counts
• And Enzo’s own anecdote: the biggest improvement they saw in their content did not come from optimizing prompts, it came from labeling their data well

Engineering Teams

• José’s concern: what is going to happen to whoever is studying computer science today. He sees more and more fellowships between companies and universities
• The profile he sees emerging: a hybrid between software engineering and someone from data, because the work of context is essentially data engineering
• Felipe’s stance is the opposite on junior hiring: they ran an internship program with about 45 or 50 people and they kept around 35
• His argument: the software engineer career is poorly specified, it is as broad as saying analyst or manager
• Where the true value of a developer sits in his business: understanding the customer better than the customer understands themselves, and modeling the situation in an implementable way
• What did disappear: the role of someone who is handed the ticket and only develops it. If I am going to tell you exactly what to do, I might as well tell the model
• Damián is suspicious of the panic: in 2020 and 2021 they hired a huge number of people that nobody needed, so the comparison is anchored to an inflated number
• And his read of the discourse: a lot of the message about replacing programmers was marketing at a moment when capital had to be raised
• The quote they bring up: not hiring juniors is one of the worst ideas, because they are the ones who use these tools best
• What Damián does believe is going to happen: collapsing roles. Less separation between product manager, engineer, designer, and project manager, with the hat worn by fewer people
• His reference is The Mythical Man-Month: the team of seven people that book describes can today be assembled with fewer
• The advantage: a small team executes faster because it has all the context, and that compounds exponentially
• Felipe’s nuance: he agrees with smaller teams, but he likes keeping a product manager per team, as the CEO of that squad

Prototypes, Hackathons, and Financing

• An investor told Felipe that the value of a prototype today is zero
• His correction: it is not that it is worth zero, it is that it loses value as a signal that you can take this to production
• But for whoever is raising capital it makes it easier, because you arrive at the investor with something much better to show
• His anecdote: in a 36-hour hackathon the demos were more polished than his own product, with effects and animations
• What José adds: there are teams with designers who now program, and that raises the aesthetic bar
• The financing pattern he sees: bootstrapped teams, and teams that do their seed round in Latin America and go straight to the United States for the series A
• And a warning from Damián: if you did not raise capital and others did and on top of that they use these tools, you are competing against people who go faster
• A change that Enzo highlights as positive: before, in Latin America, there was a bias toward the founder with a business profile. Today the technical founder gets funded more

Where the Opportunities Are

• Damián starts with the general: there is a mandate at large companies to use AI. If you show that it produces more efficiency or less cost, you have a product
• And he points to an open field: internal B2B productivity. It is boring for a lot of people, and for that very reason he likes it
• His warning for whoever builds from the region: do not think only about Latin America, because you restrict yourself
• But if you do, there is more informality and a much higher proportion of small businesses, so profitable lifestyle businesses are going to appear without venture capital
• The trend José sees: bundling into a single product what used to be two or three companies. Taxes and accounting as features of the same thing
• His read of the market: it is fragmented at two levels, the old one and that of the traditional SaaS of the 2010s
• On talent, Damián points to a real arbitrage: in the region there is very good talent capable of building this and still few large companies competing for it
• The disadvantage he acknowledges: not being in the valley costs more than before, not so much because of business but because of innovation and informal conversation
• José confirms it with his experience at Wildlife: in 2014 they competed for talent with Google and Facebook, and in 2017 with the large startups of the region
• And his optimistic close: Wildlife was built global from Brazil, with Brazilian talent, with nobody in Silicon Valley at the start
• Felipe’s reminder: software is worth much more in the United States, and talking about Latin America as a single market is misleading, because it is highly atomized

Silicon Valley Glossary

People, Companies and References

People Mentioned

• Damián Schenkelman: employee number 10 at Auth0, today he runs his research and development lab, and he invests as an angel
• José Noblecilla: leads technology and AI at Valor Capital, previously founder and director of AI at Wildlife Studios
• Felipe Sateler: founder and CTO of Buk
• Enzo Cavalie: host of the podcast

Companies and References

• Auth0: acquired by Okta in 2021 for more than 6 billion dollars
• Valor Capital: more than 2 billion under management
• Buk: payroll of more than 4,000 companies in five countries, valued at more than 800 million
• Wildlife Studios: built global from Brazil with Brazilian talent
• Shopify: the recurring example of center of gravity and of agent experience
• Vanta: the precedent of a company that is born to solve a standard
• Granola: his example that the best interface can be none
• The Mythical Man-Month: the book from the seventies whose ideal team today fits in fewer people