It was never 700 engineers

The Builder.ai collapse was never about 700 Indian engineers pretending to be a chatbot. On May 22, Bloomberg reported that the company had given its lenders a 2024 revenue figure of about $220 million, when the real number was about $50 million. The fraud, if a court ends up calling it that, was in the spreadsheet.

The version that spread was a headline instead: "The company whose 'AI' was actually 700 humans in India." It fit a joke that was already going around, the one where AI stands for "Actually Indians." I'm an Indian engineer, a year out of a master's and writing software in Chicago. I get the joke.

It's also wrong.

Gergely Orosz of The Pragmatic Engineer talked to former Builder.ai engineers and wrote up what they told him earlier this month. The company had a code generator called Natasha, built on Claude and other large language models, and it separately ran an "external development network" of roughly 500 to 1,000 outsourced developers who did custom work for clients. Those are two different businesses. Nobody in the second one sat behind the first, typing its answers.

The meme didn't come from nowhere. In August 2019, The Wall Street Journal reported that Engineer.ai, as the company was then called, used human engineers rather than AI for most of its coding work. If the Journal had it right, that's AI-washing. Selling AI while humans do most of the work is lying in your pitch, and the company earned that story.

But this year's story is about money.

Builder.ai raised more than $445 million from backers that included Microsoft, SoftBank and the Qatar Investment Authority, it was valued at $1.5 billion, and on May 20 TechCrunch reported that it was running out of money. Staff were told it would go into insolvency. Bloomberg also alleged "round-tripping" with VerSe, an Indian company, in which the two billed each other to inflate sales. VerSe called the allegations "absolutely baseless and false." The company has now filed for bankruptcy in the US.

On May 26, Bloomberg reported that federal prosecutors in New York had asked for the company's financial statements, accounting policies and customer list.

The code isn't on it.

The internet picked the 700 engineers because that version was the comfortable fraud. If you sell AI, it tells you that Builder.ai was never real AI, so its collapse says nothing about your product. If you think AI is a bubble, it tells you there are people inside the bubble, doing the typing. Both camps get to keep what they already believed. The revenue story needed a Bloomberg subscription. The meme needed a thumb.

The comfortable fraud wanted someone to laugh at, and it picked us. About 1,000 people, roughly 80% of the staff, lost their jobs in this collapse, and the part that went viral was a joke about Indians. The joke only works if an Indian engineer is what you get when the real thing runs out. The developers in that outsourced network did custom work for clients. That's a job. It's the same kind of work I do, and there's nothing fake about it.

Natasha was built on the models the rest of us paste code into. A model drafts, an engineer decides. That's the normal way to use an LLM, and it's how I use one. The "fake AI" was ordinary AI with ordinary limits, sold with the usual exaggeration.

The comfortable fraud let both camps skip the balance sheet and feel clever about it.

Everyone argued about the engine. Nobody checked the odometer.