Common pitfalls when building generative AI applications
Chip Huyen
Chip Huyen breaks down six mistakes teams keep making building AI apps, from using LLMs when they're overkill to skipping human review. It matters because most teams learn these the expensive way, after months of building the wrong thing.
Based on reporting by Chip Huyen — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Chip Huyen has been watching AI teams stumble into the same six potholes for a while now, and she finally wrote them down. The list reads less like a warning label and more like a confession booth for anyone who has shipped a genAI product in the last two years.
The first trap is reaching for a language model when a spreadsheet formula would do. Huyen describes a team that fed household energy schedules into an LLM to cut electricity bills by 30 percent, framing it as a breakthrough. Her question was simple: how does that compare to just running the dryer after 10pm when rates drop? The team never answered. They shut the project down soon after. Linear programming has existed for decades and doesn't hallucinate; sometimes the boring tool wins.
The opposite mistake is blaming the model when the product is actually broken. Huyen points to Intuit's tax chatbot, which flopped until the team realized users were staring at a blank text box with no idea what to type. Adding a handful of clickable suggested questions fixed the trust problem instantly. LinkedIn hit something similar building a skills-assessment bot: users didn't want the technically correct answer (
My take — AI-written commentary, not fact-checked reporting
, she found that saying
Read more about this at: Chip Huyen