I’ve seen a lot of AI projects die. I’m talking about projects with brilliant teams, millions in funding, and world-changing ambitions. And most of them died for the same stupid reason: their data was a mess.
Everyone wants to talk about the sexy side of AI. The models that write poetry or create art. As an investor in companies like OpenAI, Anthropic, and Scale AI, I get a front-row seat to the magic. But I’m here to talk about the part nobody wants to talk about. The part that isn’t glamorous. The part that’s a tangled, ugly, and expensive nightmare.
I’m talking about the data.
For every successful exit I’ve had, like RemoteTeam getting acquired by Gusto, there were a dozen data-fueled dumpster fires I had to put out. I’ve spent more nights than I can count wrestling with corrupt databases and broken data pipelines. It’s a brutal, thankless job. But surviving it taught me a few things most founders never learn until it’s too late.
Here are the five brutal truths I learned from wrestling AI data across 10 different startups.
Truth #1: Your Data Isn’t Just Messy. It’s a Ticking Time Bomb.
Let’s be honest. Your data is a disaster. You might think it’s okay, but you’re probably wrong. I’ve looked under the hood at over 200 startups as an angel investor, and I’ve seen the same horror show every time.
You’re pulling data from everywhere. User clicks, sales logs, social media, third-party APIs. It’s all getting dumped into a data lake that’s quickly turning into a data swamp. It’s a chaotic mess of inconsistent, incomplete, and just plain wrong data. And it’s not just messy. It’s a liability.
I remember a fintech startup I advised. They were building a fraud detection model and were so proud of their massive transaction dataset. But nobody had bothered to normalize the currency field. They had dollars, euros, and yen all mixed together. The model was completely useless. It was flagging legitimate transactions and letting fraudsters walk away with millions. It almost killed the company.
Another one, a health-tech company, had a patient dataset riddled with duplicates. One patient was in the system 15 times under slightly different names. The predictive model they built was a joke, and they burned through a $2 million seed round before they even realized the problem.
This isn’t a small problem. It’s the default state of data. And ignoring it is like trying to build a skyscraper on a foundation of quicksand.
Truth #2: Stop Chasing “Big Data.” You Need “Right Data.”
The tech world has a dangerous obsession with “big data.” Founders love to brag about their petabytes of data. But here’s the secret: most of it is garbage. More data just means more noise.
I’ve seen startups with more data than God that couldn’t build a model to predict what a user would have for breakfast. And I’ve seen startups with a tiny, curated dataset build models that felt like they could predict the future.
The difference? The first group was chasing a vanity metric. The second group was focused on what actually matters: the right data.
“Right data” is clean, relevant, and ruthlessly curated for the problem you’re solving. It’s data you can actually trust.
At RemoteTeam, we didn’t have a massive dataset. We tracked user behavior, but we were maniacal about data quality. We spent more time on data cleaning and enrichment than on model building. The result? We built a churn prediction model that was 95% accurate. It literally saved us millions of dollars and was a huge reason Gusto acquired us. We had the right data, not just big data.
So forget the petabytes. Forget the hype. Find the data that actually matters and treat it like gold.
Truth #3: Your Data Scientists Are Not Janitors.
This one drives me crazy. You hire a PhD in machine learning, pay them a Silicon Valley salary, and then you have them spend 80% of their time cleaning up your messy data. It’s the most expensive and inefficient way to run a data team.
Data scientists are not data janitors. They’re supposed to be scientists. They should be running experiments, testing hypotheses, and building models that create value. They should not be writing regex scripts to fix your broken CSV files.
If your data scientists are spending their days doing data cleanup, you’re failing them. And they will leave. I’ve seen it happen over and over. The best ones always do.
The solution is simple: hire data engineers. Hire the specialists who live and breathe data pipelines, warehousing, and quality. Let them build the infrastructure that turns your data swamp into a clean, reliable resource. Let them do the dirty work.
At MovieLaLa, which was acquired by Gfycat, we had a small but mighty team of data engineers. They built a pipeline that could process billions of user events without breaking a sweat. Our data scientists were free to focus on building a recommendation engine that was so good, it got us featured in TechCrunch. That was only possible because we didn’t treat our data scientists like janitors.
Truth #4: Data is a Team Sport, Not a Black Box.
Too many startups treat their data team like a mysterious oracle. They’re a group of wizards in a dark room who occasionally emerge with a new model. The rest of the company has no idea what they do or how they do it.
This is a recipe for disaster. You end up with models that are technically brilliant but commercially useless. You get a recommendation engine that only recommends things you’re out of stock on. You get a churn model that can’t be explained to the sales team.
Data isn’t a solo mission. It has to be a collaboration between your data team, your product managers, your engineers, and your business leaders. Everyone needs to be in the same boat, rowing in the same direction.
We had a simple rule at my last company: no data project gets started without a product manager and a business stakeholder in the room. The data team had to explain what they were building and why it mattered. And the business team had to explain how they would use it. It forced a level of communication that was uncomfortable at first, but incredibly valuable.
You need a culture where data is everyone’s responsibility. It’s not just something the “data people” do.
Truth #5: Your Data Will Never Be Perfect. Build for It.
This is the hardest truth to accept. After all that work—the cleaning, the structuring, the collaboration—your data will still have flaws. There will be biases. There will be outliers. There will be mistakes.
And that’s okay. Perfection is impossible. The goal is not to have perfect data. The goal is to understand the imperfections in your data and build models that are resilient to them.
I’ve seen founders who were so obsessed with data purity that they never shipped anything. And I’ve seen founders who were so naive about their data quality that they shipped models that blew up in their faces.
The key is to be a realist. Understand the limitations of your data. Be transparent about the accuracy of your models. Set realistic expectations with your customers and your investors.
Your AI is only as good as your data. And your data is a messy, imperfect reflection of a messy, imperfect world. The sooner you accept that, the sooner you can start building AI that actually works.
It All Comes Down to This
Building a successful AI company isn’t about having the fanciest algorithm or the biggest dataset. It’s about mastering the unglamorous, frustrating, and absolutely essential work of managing data. It’s about respecting the data. If you can do that, you have a fighting chance. If you can’t, you’re just another startup waiting to die.
Frequently Asked Questions
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.
Can I implement all of these at once?
I'd strongly recommend against it. Pick the 2-3 items that resonate most with your current situation and focus there. Trying to do everything simultaneously is a recipe for doing nothing well.
How do I know which items apply to my situation?
Start by honestly assessing where your biggest bottleneck is right now. The items that address that specific constraint will give you the highest return on your time and energy.
Which item on this list has the highest impact?
It depends on your stage and context, but in my experience, the items near the top of the list tend to have the broadest applicability. That said, sometimes the less obvious items create the biggest breakthroughs for specific situations.