One of the most unpopular technical decisions I made was during the early days of an AI conversational platform company I previously worked at. At the time, there were three competing approaches emerging in the chatbot space: traditional machine learning models, language-based AI systems, and knowledge graph models focused on word relationships and semantic mapping. This was before the advent of today’s LLM-based conversational models.

Each engineering team strongly believed its approach was the best and wanted the company to fully commit to its technology. Instead of choosing a single winner, I pushed for a different strategy of combining all three models into a unified analysis platform.

The decision was unpopular because it increased complexity and required the teams to collaborate rather than compete. We built an analysis feature that routed every user query through all three models simultaneously, then compared and scored the outputs. This required six months of alignment work to establish common evaluation criteria, scoring techniques, and transparency into why one model’s response was ultimately selected.

I handled the pushback by reframing the conversation away from “which team wins” and toward “how do we deliver the best and most explainable outcome for customers.” I also made the evaluation process data-driven so decisions were based on measurable performance rather than opinions.

In the end, it proved to be the right call. The combined approach gave customers better accuracy, greater transparency, and the flexibility to evolve as AI technologies matured, instead of locking the company into a single approach too early.