The Only Two Things That Still Matter: Intent and Taste
I joined Rocky Khan on the Masters of Data, BI & AI podcast to talk about what I’ve learned building data teams across Alibaba, HelloFresh, and Mistplay, and what’s changing now that I’m on the consultancy side at PyMC Labs. Here’s the episode, and below it a summary of the ground we covered.
Listen on Buzzsprout → | Listen on Spotify →
Pricing Is the Most Powerful Data Problem
Looking back at every company I’ve worked with, the projects that moved the needle most were almost always about pricing. At Alibaba’s Lazada in Singapore, we built a route optimization system for grocery delivery that cut delivery times by 15%. But the real win came from optimizing which delivery slots to make premium. Understanding which users would pay more for a specific time window generated additional revenue while also improving on-time rates and customer satisfaction.
At HelloFresh, one of my analysts noticed that new customers would get a steep discount on their first box, then see the full price on their second box and immediately churn. The price jump was too sharp. So we started splitting the discount across multiple boxes to ease people into the service at a consistent price. Retention improved by a massive amount.
Then we stumbled onto something I still think about. Telling people “$40 discount” and telling them “10 free meals” is the exact same offer in dollar terms. But “10 free meals” converted significantly better. People don’t love discounts. They love free things. We ran hundreds of experiments per year on different combinations of framing and segmentation, and that insight alone moved the needle on customer acquisition costs.
Mistplay was the same story. We replaced a fixed pricing engine with a dynamic one based on machine learning. Revenue tripled in three years.
Whoever understands price elasticity with data, wins. And today, with A/B testing, Bayesian bandits, and synthetic consumer panels, you can do in days what used to take millions of dollars and weeks of focus groups.
AI Is Eating Software. What’s Left?
Rocky asked about the old Andreessen Horowitz line, “software is eating the world,” and the Databricks follow-up, “AI is eating software.” I’m 100% on board with this.
The way we write software has completely changed. At PyMC Labs, we ship dozens of features a day with agentic coding tools like Claude Code, with minimal human intervention. The developer’s job isn’t writing code anymore. It’s architecting the guardrails and plans that AI implements.
So what’s the competitive moat in software now? I think there are only two left. The first is exclusive data access: if you have credit card transaction data, government contracts, or proprietary user behavior data that nobody else can get, that’s real. The second is distribution. If you’re Microsoft, pre-installed on 60% of laptops worldwide, nobody can copy that by prompting an LLM.
Everything else (features, UX, security compliance) is getting easier to replicate by the month. A startup building a SaaS product today needs to understand that their idea can be reverse-engineered and rebuilt in weeks. SOC 2 compliance, which used to be a barrier, is now something startups like Vanta help you set up quickly. It’s table stakes, not a moat. It buys you months, not years.
What VPs of Data Should Actually Be Asking
When enterprise buyers evaluate AI solutions today, they typically ask: “How trustworthy are your agents?” That’s the right question, and at PyMC Labs we invest heavily in evaluation frameworks and even publish academic papers validating our approaches.
But the question they’re not asking matters more: “How can we rethink entire business processes around AI?”
Most companies come to us wanting to automate a piece of an existing workflow. Stick an AI agent in the call center. That’s fine, it delivers real productivity gains. But what if you redesigned the entire product experience? Instead of a standard e-commerce checkout, what if a shopping agent interacted with consumers directly to mediate the transaction? A few companies are starting to think this way, and the gap between them and everyone else is widening fast.
Even how teams collaborate is changing. We’re building an AI data scientist that lives in Slack. Picture this: a CFO and CMO are in a thread, they realize they need an analysis, they tag the agent, and get a forecast right there in the conversation. You can imagine inviting these agents into a phone call next. The way teams work together looks completely different when an AI colleague is always available.
What People Are Getting Wrong
I keep changing my mind on this, honestly, almost on a weekly basis. But here’s where I land.
Most people underestimate the pace. Exponential growth is nearly impossible for the human brain to process intuitively. Most people still think of ChatGPT circa GPT-4, and the experience with those older models was good but not great. Every few weeks now, a new model release brings a massive jump in capability. My rule of thumb: if a problem seems unsolvable today, wait three weeks. Some new model will probably crack it.
At the same time, people overestimate how fast organizations will actually adopt this stuff. The technology moves exponentially, but humans adopt it linearly. There are always champions who move fast and capture outsized value, but on average, organizations are slow. This mismatch is why hundreds of billions are being poured into AI infrastructure right now, and why some of those bets may fail. Not because the technology was wrong, but because the timing was off. Companies like OpenAI won’t go away because they made the wrong bet. They might stumble because they were too early.
Intent and Taste
Rocky asked what jobs on a data team will disappear. Writing SQL will go. Building models will go. Everything that’s execution will be automated. People will use plain English, or just talk to their phone, to describe the problem, the constraints, their domain knowledge. The AI will handle the implementation.
So what’s left for humans? Two things.
Intent: knowing what you actually want. What are you optimizing for? What are the boundaries? “Hey, price can’t be more than $10” or “these two variables have nothing to do with each other.” That kind of judgment.
And taste: knowing whether what comes out is good or is just slop. Can you look at a model output and tell if it’s quality work or garbage? That’s a skill no amount of compute replaces.
Those are the only two competitive advantages left. Everything in between is just execution.
You can find me on LinkedIn or reach out through my website. The book I recommended on the podcast is The Hard Thing About Hard Things by Ben Horowitz.