The art and science of menu design @ Blue Apron

I recently participated in an “Ask Me Anything” with First Round Capital, and one of the questions that I had the most fun answering was:

“How do you collaborate with other teams that may not have the same decision making process?”

Here’s my answer.

In general, Blue Apron is an extremely data-driven business that is hungry for quantitative insights and analysis. Our approach to Marketing tends to be very quantitative, as are our Finance, Supply Chain, and Operations groups.

One of the most exciting partnerships we’ve invested in over the past year is our work with our Culinary team, who are a highly creative and chef-driven organization with really impressive culinary backgrounds. Specifically, our team has built a tool powered by machine-learning that allows the Culinary team to input recipe attributes into a drop-down menus during their weekly recipe writing sessions to predict recipe demand, reach to different customer segments, and overall order rate based on the portfolio of recipes. Because we have 8 recipes for our 2-person plan, and the tool spit out a score after inputting simple variables, we called it “Magic 8.”

Initially, the inputs to the tool were fairly simple (pick a protein by cut, and then specify the dish and cuisine type) but it generated 10–15 outputs, ranging from the predicted weekly order rate (including an upper and lower bound), estimated weekly profit and margin, distribution of recipe quantities, and more. And, as we worked with the Culinary team to understand the full scope of their needs, our list of requirements grew and grew, as did the tool’s complexity.

Over time, after the tool did not enjoy the same level of adoption that we hoped, we had a candid follow-up with the team. Did the tool not generate the right outputs? Was the functionality incorrect? Did they not trust our work? No — it was simply too complicated, and by serving everyone, we’d served no one. And, not only did our laundry list of outputs provide way too much detail, we were impeding the creative process and creating distrust of the tool during the recipe writing sessions because it was too constricting by simultaneously optimizing for recipe appeal and profitability at the same time (To make things worse, the “front end” is currently powered by Google Sheets, which we’ve pushed to its very limits).

So, we simplified. We moved away from focusing on the trade-offs between profitability and customer appeal (which was the initial mission of the tool), and honed in on optimizing for customer preferences with the ingredients we’d have on hand. And, we combined the different outcomes into a single output that generated a 1–5 score (color coded into red, yellow, and green) with a “3” or above representing a desirable recipe. We then agreed with Culinary on the acceptable number of “red” or “yellow” recipes in a given week, based on supply chain and operational constraints unrelated to the creative process.

We’ve learned a lot from these efforts to bridge the gap between quantitative (machine learning) and qualitative (creativity and intuition), but the main one is to keep things simple, and focus on the user even if that user is internal. Internal tools shouldn’t be hard to use, and neither should data products. We don’t need to show of the ML bells and whistles for colleagues to trust us — our commitment to serving their needs is what created trust, and simple, clear communication helps to bridge gaps in skillsets and backgrounds.