Early last year, Silverman added an AI director to focus on those two buckets: guest- and employee-facing possibilities.
To start, Silverman and his team wanted to create a guest-facing AI that was “insulated from a lot of the problems” that exist today with large language models (LLMs). “They can make mistakes, and particularly when you’re allowing guests to put in pure text into a chatbot,” he explained. Plus, they wanted to avoid bad actors “jailbreaking” (bypassing the security guardrails) of the AI model.
The team intentionally avoided using a traditional text-input chatbot. “If you output raw text from AI, you don’t have tight control over what it says. We were looking at doing it in a more controlled way but still using AI under the hood to power decisions based off the menu. We control what values (guests) could select. We looked at what flavors people are most interested in. It’s more a templated approach; we’re taking in what they’re saying but using AI under the hood to match their flavor preferences with our entire menu offering and every ingredient we offer.”
