The $14B Business of Boring Work

💡Alexandr Wang’s Insight: Seeing the Hidden Market

Alexandr started Scale AI in 2016. He realized early that clean, organized, labeled data can be a treasure trove. Back then, it was a real bottleneck for AI.

He didn’t see labeling as a means to an end; he saw a service business. He could become the supplier of a much-needed resource: datasets.

Today, his company, Scale AI, employs 100,000 labelers and services the top AI firms, such as Meta, Apple, and OpenAI.

Wang himself calls his business “mundane and unsexy.” But that’s often where the opportunity lies.

It’s a service business that employs gig workers at $8 an hour. The only tech here is the tools, which must be easy to onboard and use and ensure high quality and productivity.

That’s all the tech that was needed to create this rare resource. But there were more hurdles to overcome.

Why Some People Spot Big Opportunities Early

I saw this work as just a task. Our company was striving for bigger goals but focused on model development—I could not see the service business opportunity right in front of me.

Entrepreneurial minds see beyond a product's sexy part. Sometimes, the “boring” infrastructure is the real win.

Wang clearly saw a pain point. But what he also saw was a scalable solution to fix it.

The math can be simple: pay $x an hour to a labeler that creates a usable dataset of images that can be sold for $y. Then build the tools to scale this up - tools have the goals to hire fast, make them productive fast, ensure quality, and then go and sell sell sell.

Many of these smaller startups were coming up thinking precisely that way. But Wang succeeded because he was able to fix the most significant bottleneck for his success: hiring labelers fast and cheap. He couldn’t crank out new datasets fast enough - the hurdle to scale was still the manual work. But he found a way to fix that.

For this purpose, in 2017, he created Remotasks, a web service that focused on recruiting cheap labor. With the site, they targeted cheap labor countries and found the right communities to recruit in, such as internet cafes.

Timing the Market

Wang launched Scale AI during the self-driving car boom. This was when the entire industry — including the startup I was at — needed massive amounts of labeled data. Vehicles had to recognize objects, traffic lights, and everything else around them, and Wang capitalized on that demand.

He signed up Tesla and Cruise as early customers. By 2019, Scale AI hit a $1 billion valuation.

Then, through his ties to an early Y-Combinator startup, he ended up with a pandemic roommate — none other than Sam Altman, the CEO of OpenAI.

As the self-driving car hype started to slow down, generative AI started to ramp. Wang saw the shift early and jumped on his first generative AI contract with OpenAI.

Was luck, timing, and the right connections involved? Definitely yes. But as one of Scale AI’s board members put it, Wang “is very good at seeing around the corner. He has multiple entrepreneurial and visionary skills blended into a single person.” He sees the next big wave before it hits.

Lessons for Product Thinkers

• Find the bottleneck and own it. What is slowing progress in your industry? What is the solution that removes the friction?

• Follow the money, not the hype. Find what companies are willing to pay for, not just the hype they get excited about.

• Unscalable work can lead to a scalable business. Manual labor is not sexy tech, but tackling it and layering automation on top can lead to massive scale.

• Services can still win. When automation isn’t ready, people-powered businesses can thrive—even in AI.

• Tech’s most significant pain points aren’t always complex — Organizing human labelers was a real challenge. Operational pain points exist everywhere.

• Timing and execution beat genius.

TL;DR Find what’s boring but crucial.

What Other “Boring but Crucial” Opportunities Exist?

You might be wondering: “Sebastian, what other “boring but crucial” opportunities are sitting in plain sight that you haven’t jumped on?”

Well, there are a few:

• Data cleaning - AI training data is messy. If it’s low quality input it needs fixing. AI models are only good as their input. But I don’t think data cleaning has been figured out at scale.

• Data origin tracking - We all know that AI models are trained on copyrighted content. And more regulation will come. So companies need to prove where their data originates from. No matter the exact regulation - there will need to be manual work.

• AI model testing & validation - We can only rely on AI if we trust it. We talked about this in our agentic AI deep dive and found a big hurdle to success. And the more we rely on AI the more guardrails we need. Few companies offer scalable ways to validate AI at scale.

If you’re looking for your own “grunt work goldmine”, one of these might be it. If you decide to jump on one, let me know. I’m clearly not going to. I’d rather spend my free time writing this newsletter for you. 🙂