What's the key to making an AI project actually work?
My take
I’ve been an early adopter of generative AI since I started using ChatGPT in February 2023. Since then I’ve used AI for idea generation, then coding help, and now process automation and AI-powered workflows. The tools have changed enormously. The lesson hasn’t.
An example
We needed to work out where residents were in the delinquency cycle based on the content of their messages. The tempting move was to write a prompt and let a model sort it out.
Instead, an intern and I read through sample messages ourselves and identified the patterns by hand. Only after we understood what each stage sounded like did we use Snowflake’s Cortex Code to classify messages at scale.
Why this keeps being true
- A model can only classify, summarize, or automate what you can define. If you can’t explain the process, the AI can’t either.
- The slow, manual steps worth automating are usually invisible until you sit with the people doing them.
- The most valuable skill in AI projects is translation: turning how a team actually works into something a system can act on.
That’s been my job in every role, long before AI. It turns out to be the job AI needs most.