I just got off the phone with a founder. It was an early Sunday morning call, and they were distraught. The company had launched with a breakout AI feature. That one worked. It delivered. But every new release since then? Nothing’s sticking. The team is moving fast. They’re adding features. The roadmap looks full. But adoption is flat. Internal momentum is fading. Users are trying things once, then never again. No one’s saying it out loud, but the trust is gone. This is how AI features fail. Because they teach the user a quiet lesson: don’t rely on this. The damage isn’t logged. It’s not visible in dashboards. But it shows up everywhere. In how slowly people engage. In how quickly they stop. In how support teams start hedging every answer with “It should work.” Once belief slips, no amount of capability wins it back. What makes this worse is how often teams move on. A new demo. A new integration. A new pitch. But the scar tissue remains. Users carry it forward. They stop expecting the product to help them. And eventually, they stop expecting anything at all. This is the hidden cost of broken AI. Beyond failing to deliver, it inevitably also subtracts confidence. And that subtraction compounds. You’re shaping expectation, whether you know it or not. Every moment it works, belief grows. Every moment it doesn’t, belief drains out. That’s the real game. The teams that win build trust. They ship carefully. They instrument for confidence. They treat the user’s first interaction like a reputation test, because it is. And they fix the smallest failures fast. Because even one broken output can define the entire relationship. Here’s the upside: very few teams are doing this. Most are still chasing the next “AI-powered” moment. They’re selling potential instead of building reliability. If you get this right, you become the product people defend in meetings. You become the platform they route their workflow through. You become hard to replace. Trust compounds. And when it does, it turns belief into lock-in.

RAG is dead posts are annoying as F "R" is retrieval and "AG" is the LLM. This means you think retrieval is dead. Seriously, you think retrieval is dead? Keyword search, metadata filtering (dates, users), grep, and other filtering are retrieval. Good luck without retrieval

I’ll be talking about the most common mistakes people make with domain specific evals at this event lu.ma/multi-modal?utm_source…