It was a Tuesday when I shut it all down. No dramatic explosion, no final boardroom showdown. Just me, alone in our empty office, moving the last box of ethernet cables into my car. Five years of my life, my savings, and the trust of my first few employees—all gone. We had raised $2.5 million, built a team of 15, and had what I thought was a revolutionary AI-powered learning platform. And it failed. Completely.
Everyone loves to talk about AI in education as this inevitable, glowing future. But they don’t talk about the brutal reality of building it. They don’t talk about the soul-crushing grind of trying to make a machine teach when you haven’t even figured out what to teach. My first startup was a monument to that failure. Here’s what I wish I knew before I started.
The Dream of an AI Tutor for Every Child
Our idea was simple, or so we thought. We wanted to build a personalized learning platform for high school math. The pitch was solid: an AI tutor for every student, adapting in real-time to their strengths and weaknesses. We saw a future where our software could identify a student’s confusion from a single hesitant answer, then generate the perfect explanation, video, or practice problem to get them back on track. We weren’t just building an app; we were going to democratize education. I remember telling an early investor, with complete sincerity, “We’re building the cure for educational inequality.” I really believed it.
The first two years were a blur of excitement. We built our first natural language processing models to understand student-written answers. We designed algorithms that could, in theory, create a unique learning path for millions of users. We had a beta with a few hundred students, and the initial data looked promising. Engagement was high. I was on top of the world. I was so sure we were building the future. I would stay up until 3 AM watching the real-time user logs, thrilled every time our AI correctly identified a student's mistake.
But then, the numbers started to flatten. The initial excitement wore off. Students would use the platform for a week, then their usage would drop off a cliff. Teachers told us it was “interesting” but they didn’t have time to integrate it into their lesson plans. The AI, which was supposed to be our magic bullet, was becoming a black hole for our resources. We spent months and hundreds of thousands of dollars trying to improve our recommendation engine, convinced that if we could just make the AI smarter, everything else would fall into place. I was wrong.
Lesson 1: AI is a Magnifying Glass, Not a Magic Wand
This was my biggest, most expensive mistake. I believed that the technology itself was the solution. I thought if we just built a powerful enough AI, it would magically solve the deep, complex problems of student engagement and pedagogy.
What I learned is that AI is an amplifier. It takes what you’re already doing and scales it. If you have a brilliant, engaging, and effective educational method, AI can help you deliver it to more students, more efficiently. But if your core method is flawed, AI will just amplify those flaws. It will efficiently bore students. It will scale confusion. It will, with terrifying precision, show you just how broken your core assumptions are.
We spent a year and nearly a million dollars building an AI feature that could generate “hints” for students stuck on a problem. It was technically brilliant. It could analyze the student’s previous answers, the specific step they were on, and even the common misconceptions for that type of problem to generate a targeted suggestion. Our lead engineer, a PhD from Stanford, called it “pedagogical dynamite.” We thought it was our killer feature. The data showed that students who used the hints were more likely to get the right answer. We celebrated. We thought we had cracked it.
But then we looked at the qualitative feedback. The students hated it. One 15-year-old wrote in a feedback form, “The hints just make me feel dumber. It’s like a robot nagging me.” That hit me like a ton of bricks. We were so focused on the technical achievement of generating the hint that we missed the emotional impact. The core problem wasn't the lack of a hint; it was that the problems themselves were boring and the platform felt like a chore. The hints just felt like a more sophisticated way of telling them they were failing. We were trying to use a high-tech scalpel to fix a problem that needed a foundational rethink of the entire learning experience. We were polishing a cannonball.
Lesson 2: The 'Last Mile' is the Only Mile That Matters
In the world of software, we often think the hard part is the backend, the algorithm, the core logic. In EdTech, that’s maybe 20% of the battle. The other 80% is the “last mile”—the messy, human, and frustratingly analog world of the classroom.
I’m talking about the user interface. I’m talking about teacher training. I’m talking about whether a teacher, who has 30 kids to manage and a district curriculum to follow, can figure out how to use your platform in under five minutes. We had built this incredibly complex system, but the interface was clunky. It required teachers to go through a 30-minute tutorial to even get started. That’s an eternity in teacher-time.
I’ll never forget a visit to a pilot school in East Palo Alto. We were so excited to show off our platform to a veteran math teacher, a woman who had been teaching for 25 years. We sat her down in the school library, and I started my polished pitch. Her eyes glazed over. She interrupted me after two minutes. “Son,” she said, “I have 32 students in my next class. Four of them have special needs, and I think six of them didn’t eat breakfast. I have about three minutes between classes to prepare. Can your thing help me with that?”
My heart sank. I showed her our dashboard, with its dozens of data points and customizable reports. She just shook her head. “I don’t have time for this.” She wasn’t being rude; she was being honest. We, the tech guys, had designed a platform for ourselves, not for our actual users. We had a powerful engine, but we’d forgotten to install a steering wheel that anyone could actually use. We were so proud of our car's top speed, we didn't realize it couldn't get out of the driveway.
The real product in education isn’t the AI; it’s the experience. It’s the confidence a student feels when they finally solve a problem. It’s the relief a teacher feels when they find a tool that actually saves them time instead of creating more work. We failed because we fell in love with our technology and forgot about the people it was supposed to serve.
Lesson 3: Data is a Trap if You Don't Know the Question
We had so much data. Terabytes of it. Every click, every answer, every second a student spent on the platform was logged. We had dashboards that could track over 100 different metrics. And it was all useless.
We were drowning in data but starved for insights. We would sit in meetings looking at charts showing “engagement time” and “problem completion rates,” and we’d argue about what they meant. Did a student spending 10 minutes on a problem mean they were engaged, or just hopelessly stuck? Did a high completion rate mean the problems were effective, or just too easy? Our dashboard showed a “Grit Score” for each student, a metric we had invented that combined time on task, number of attempts, and hint usage. We thought it was brilliant, but it was meaningless. It didn't correlate with actual learning outcomes or student satisfaction.
Our mistake was thinking that collecting data was the same as learning. We were obsessed with the what and never stopped to define the why. We never established a clear pedagogical framework to even begin to interpret the data. A/B testing a button color is easy. You can measure clicks and conversions. But how do you A/B test a teaching method? How do you measure the long-term impact of a particular explanation on a student’s conceptual understanding? These are hard, messy, human questions, and we tried to solve them with clean, simple data.
If I could do it all again, I would have started with a single, simple question. Something like: “Can we reduce the time it takes for a student to master the concept of fractions by 10%, while also increasing their self-reported confidence?” That’s it. A clear, measurable, education-focused goal. Then, and only then, would I have started thinking about what data to collect and what technology to build. We did it completely backward. We built a data-gathering machine first and hoped the questions—and the answers—would just magically appear.
The Phoenix from the Ashes
Shutting down the company was one of the hardest things I’ve ever done. It felt like a personal failure on every level. I spent a few months just decompressing, angel investing in other founders’ dreams, and trying to process what went wrong. But after the dust settled, those three lessons became the foundation for everything I did next. My next venture, RemoteTeam (which was acquired by Gusto), wasn't in EdTech, but the lessons applied just the same. We started with the user's problem—the administrative nightmare of managing a remote team—and built the simplest possible solution.
I did eventually return to EdTech, but as an investor, not a founder. I backed a company that was also using AI, but this time, they didn’t lead with the tech. They started with the teachers. They spent six months in classrooms, not pitching, just listening. They found that teachers' biggest pain point wasn't a lack of personalized learning, but the sheer amount of time they spent grading homework.
So the company built an AI that did one thing really well: it automated the grading of repetitive homework, saving teachers hours every week. It wasn't as sexy as a futuristic AI tutor. It didn't promise to revolutionize pedagogy. But it solved a real, tangible problem. The teachers loved it. The students were happier because they got faster feedback. And the company is now one of the fastest-growing startups in the space.
It wasn’t as sexy as a futuristic AI tutor. But it worked. And it’s a lesson I’ll carry with me forever. Don’t fall in love with the technology. Fall in love with the problem you’re trying to solve for the human on the other side of the screen.
Frequently Asked Questions
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.