One thing that surprised us while building AI products

Hey everyone 👋

One thing that surprised us while building AI systems:

most reliability problems are not model problems.
They’re system design problems.

A lot of teams assume better prompts or bigger models will fix issues like:

• hallucinations
• inconsistent answers
• high API costs
• poor user trust

But in practice, the biggest improvements usually come from:

→ better retrieval pipelines
→ cleaner context handling
→ caching + async workflows
→ narrowing the AI’s scope early

We’ve seen smaller models outperform larger ones simply because the surrounding system was designed better.

Curious if others building in AI have noticed the same thing—or if model quality has been the bigger factor for you?