We built ChatGPT for banks before ChatGPT existed

In late 2018, I built what was basically "ChatGPT for banks", before ChatGPT existed. A self-hosted system that helped bank employees generate and validate investment and credit contracts in German, trained on several million banking documents, built by a team of maximum 17 people who were basically just enthusiasts. 😅

The numbers were solid:

errors in contracts down 70%+, creation time from 16 hours to 4.

It was successful enough that one of our banking clients acquired it for $2.5M in 2023, and that happened when ChatGPT already existed.

Shut down only in summer 2025 — that's when OpenAI finally started outperforming our solution.

Here's the thing I only understood after the fact — not even when we sold it, but later, when it was already closed and I talked to people and started analyzing.

We had a go-kart, they had a Formula 1 car. Technologically, if you compare how ChatGPT was built versus what we had, that's exactly the size of the GAP. And yet our go-kart, in exactly the place it needed to be, was better. At least for a pretty long time. 🎯

OpenAI had billions, and we had contracts worth maybe $30K.

ChatGPT is built tens of times more complex, smarter, more intelligent — but because it's spread too thin, it couldn't generate German-language banking documents and validate them with the level of quality we did. 🧠

The insight sounds obvious in retrospect. But I genuinely didn't see it while we were building it.

The presence of billion-dollar competitors is not a death sentence — if you have the combination of narrow industry focus and deep expertise, both technical and domain.

You just might not fully understand that's what you're doing until someone closes the project and you start piecing it together.

That's maybe the weirdest part of this story. 🤔