10% Faster or 10x Better?

What we talked about.

In this episode of the Sunny Ray Show, host Sunny Ray talks with Jorge Alcantara, founder of Zentrik, about the shifting economics of building with AI. Jorge brings over a decade of enterprise AI experience, including early chatbot work for Microsoft, Samsung, and Airbnb long before transformer models existed. He traces how AI selling and managing has flipped: once hard to sell but easy to control, now easy to sell but much harder to manage as models grow more stochastic. Jorge shares lessons from building human-in-the-loop training pipelines that powered Samsung support chatbots, and warns that both students and practitioners often rush into building without understanding the underlying problem space. The conversation centers on his framing of 10 percent versus 10x thinking, arguing that true transformation comes from rethinking workflows entirely rather than automating small pieces of them. Jorge also explains why the harness around a model, not the model itself, is the durable asset as providers and pricing shift. He closes by describing how his small, fast moving team at Zentrik wins enterprise deals against much larger incumbents like Atlassian.

A twelve year AI veteran on why chasing 10x impact beats chipping away 10%, and why the harness outlasts the model.

The questions, and the answers.

What is Zentrik and what are you building?

We help large companies decide what to build and how to build it, since development itself is now commoditized by LLMs. The hard problem isn't writing code anymore, it's choosing what to bring into your product so it delivers the most value to customers, while capturing the most value back for the business.

You've been doing enterprise AI since 2015. What were some of the key patterns back then, and what's changed since?

Back in 2016 selling chatbots to companies like T-Mobile, they worried a bot might harm the brand, but models were deterministic, so risk was near zero. Today models are stochastic and chatbots can say almost anything, yet everyone still rushes to deploy them. It used to be easier to manage AI but harder to sell it. Now it's easier to sell but much harder to manage.

At your previous company, you built human-in-the-loop training pipelines. What did you learn about where the human actually has to sit?

We built a layer on top of power users answering support questions, like a crowd for Samsung, and trained a chatbot from those answers, rewarding people for training the system. The real lesson wasn't technical, it was learning how to manage teams of human experts so they actually improve the model instead of gaming it for rewards.

You teach generative AI as well as building it. What do students get wrong that maybe practitioners do not?

Students and even some practitioners jump too fast into the application layer without understanding the problem space underneath. If you skip the basics you end up iterating blindly. Developers should learn the fundamentals, and practitioners should understand their users before building. Today's tools let you start building without truly understanding what you're doing, and that gap is the real risk.

One of your talks is called Correct Code, Wrong Product. What was the moment that title came from?

I gave that talk at an AWS Loft about a year ago to a room of engineers. The point was that code has zero value on its own. Software only matters once it becomes a product that solves someone's problem. Engineers need to think about the value they create for users, not just write technically correct code, because syntax and code are increasingly solved.

Why is the harness the durable asset and not the model?

If you build only on top of one model provider, you're stuck when pricing or access changes, and some plans effectively subsidize you until they don't. The harness, the layer that understands your users and your context, is what survives as models get swapped in and out. The model comes and goes, but the harness keeps working for you.

What separates a team that gets 10% back from one that gets 10x better?

The 10% path is taking your existing human workflow and automating pieces of it, chipping away at tasks. The 10x path means rethinking the funnel completely, like turning it into something entirely new. If you only chip away, you end up with none of your day being yours and it feels soulless. I push people to think about multiplying impact, not just trimming tasks.

You're a very small team selling into enterprise against large incumbents. How do you win a deal you maybe shouldn't win?

Tools like Jira, Aha, and Productboard are hard to wrangle and slow to react. We built our product AI native so it works cleanly with Claude and ChatGPT, and we're fast, sometimes turning a customer call into a shipped fix the same night. That speed is something a team like Atlassian can't match, which is how a two person, two year old company has displaced them at billion dollar accounts.