1/10 – Future B2B software unicorn. Everything changed when its founder believed that an AI prototype created in 2 days could become an industrial product. The story of a fatal illusion.
We are leaving the Paris commercial court. A company which, just eighteen months ago, was presented as the “future French unicorn of B2B software” has just been liquidated. Its product: a project and business data management tool very popular with mid-sized companies and large groups.
Here’s what happened.
Jordan, the founder and CEO, is an excellent product designer. He developed the first version of the software himself with a small team of experienced developers. The product had found its market, gained customer loyalty, and was experiencing profitable growth. The company was healthy, even if the technical team struggled to keep up with the pace set by Jordan.
Then came the wave of powerful LLMs.
Jordan, like many other executives, saw these artificial intelligence models as an opportunity to revolutionize his product. He locked himself away for two days with his computer and two LLMs and came out with a complete prototype of a new “Enterprise” version. The interface was beautiful, the features expected by large accounts were present, and the internal demonstrations produced impressive results. The prototype seemed to understand the business better than some historical collaborators.
This is where the first great illusion operated: “If the AI made the prototype in two days, industrialization will only be a formality.”
Convinced that he had a rare nugget in his hands, Jordan did what many founders do: he multiplied the announcements. Interviews, LinkedIn, customer teasers. “In less than a month, the version that will change the game for large groups.” Investors were euphoric. A few sold their shares at record valuations. Most kept on, convinced they were sitting on the next unicorn.
Then it was time to move from prototype to production.
The technical team, led by Anaïs (very competent technical director), quickly saw the problems: securing customer data, rights management, traceability, robustness, integration with existing IS, monitoring, fallback, non-regression tests, etc. The prototype worked well under controlled conditions, but it was extremely fragile and dependent on responses from the LLM.
Jordan was getting impatient. He didn’t understand why industrialization was taking so long. He was beginning to wonder if his team was “stranded in the old world” or destabilized by the AI. Why refuse the generalization of AI agents to speed up work? He requested daily follow-up meetings, sometimes bypassed Anaïs, and pushed to go faster.
After two months, some of the code had been generated or aided by AI, but the quality was poor. Anaïs and two tech leads resigned, believing that Jordan had lost confidence in traditional engineering and was taking reckless risks.
Three months after the initial announcement, an “almost complete” version was put on the market despite opposition from Zoé, the quality director. Jordan had decided: “We have to show that we are moving forward.”
The first customer feedback was catastrophic. Instability, inconsistent responses, strange bugs, degraded performance. Many demanded to return to the old version. But the company, by focusing on the new, had neglected the maintenance of the old.
Then the final blow arrived: the LLM provider deployed a new major version. The behavior of the model has changed enough to make the next-generation software unusable by customers. The product, very dependent on calls to AI, collapsed.
After the departure of the main technical managers, the remaining team failed to stabilize either the new version or quickly restore the old one. The customers, furious, sent formal notices. Churns have multiplied. The cash flow dried up in a few weeks.
At an extraordinary general meeting, investors voted to dismiss Jordan and dissolve the company. The teams were reclassified as best they could. End of story.
The false beliefs that killed the business:
- Believing that an impressive prototype made with an LLM is equivalent to an industrial product.
- Underestimate all the engineering work (reliability, security, observability, integration) and quality improvement, which represents the real value.
- Thinking that AI replaces senior developers rather than making them more productive as long as they maintain control.
- Confusing speed of prototyping with speed of bringing critical software to market.
- Overestimate one’s own ability to assess technical complexity once outside one’s area of expertise.
AI is an extraordinary tool. But like all powerful tools, it severely punishes those who believe it abolishes the fundamental laws of engineering and management.
