He Left Investment Banking to Start a Company, and Chose Not to Build a Workplace AI Assistant

Doing AI at an investment bank, most of his products died in compliance review. Zac quit to start his own company, but steered clear of the workplace AI assistants everyone else is building: “That market may end up swallowed by OpenAI and Anthropic.” He chose construction instead, where the data is messy and scattered, and the barrier to entry is the moat.

The most expensive moment on a construction site is when it stops.

Builders always order extra bricks to leave room for waste. But once the waste exceeds the buffer, they have to place a rush reorder, and reorders come with lead times. During the days spent waiting for bricks to arrive, the schedule stalls, workers still clock in, equipment rentals still run, and money keeps burning.

The person who told me this example is named Zac. This is the opportunity he saw for his startup.

At the investment bank, Zac led a team building AI applications. They built a lot of products, but most got stuck in compliance and never rolled out. It was a good job. But as he put it: “Watching everything develop at high speed while you spin in place, it’s painful.”

That was when the idea of starting a company took root.

For direction, he didn’t choose white-collar work: document writing, customer service, coding assistants. “That market is hard to capture. In the end it may all get swallowed by OpenAI and Anthropic.”

He chose a traditional industry: construction.

Before investment banking, Zac built optimization models for oil companies. Oil fields inject water to push the oil out, and if the oil-water mixture coming up contains sand, it damages the nozzles. In serious cases it can write off an entire pipeline worth over a billion dollars, or even threaten the safety of the well. So sensors have to monitor everything in real time and raise alerts at any moment.

“Applying AI to this field, the moat is very high,” he said.

Construction is the same. The big model companies build general-purpose products; they won’t collect the fragmented, dirty data of one specific industry. For a startup, that fragmented data is both the barrier to entry and the moat.

Using operations research for supply chains is nothing new. Linear programming and production scheduling systems have existed for decades, and large companies in manufacturing and retail have used them all along. Traditional industries never managed to adopt them, for two main reasons. First, the data is scattered across paper, Excel, phone calls, and chat logs. Second, an industry like construction is project-based: every new site means a fresh set of suppliers, subcontractor crews, and site conditions, so data never accumulates, and building a dedicated system for a single project isn’t worth the cost. So site managers rely mostly on experience and judgment. Once there are too many people and too many moving parts, one person simply can’t keep track of it all.

What Zac wants to do is gather up this fragmented data, turn it into inputs a model can use, and help managers make decisions.

Procurement inquiries are another example. In the past, getting quotes meant a purchaser calling and emailing suppliers one by one, and comparing prices meant compiling spreadsheets by hand. Because it was so time-consuming, they usually only asked the two or three suppliers they knew well. But building materials are commodities: prices float, order volumes are large, and a tiny difference in unit price adds up to a big difference in the total bill. AI can do the inquiry step much better: consolidating quotes and information from all suppliers, drafting the inquiry emails, and sending them out after a human reviews them.

I asked Zac whether giving up his comfortable compensation carried too much risk.

“Even if it completely fails, I can come back and find a job in six months,” he said. “The risk is still manageable.”

And now he can see a glimmer of success. Some clients are willing to grow alongside their product.