Too Good To Go, a marketplace connecting consumers with businesses to reduce food waste, wanted to understand why partner stores stopped using its platform. By Amazon Bedrock and working alongside AWS Partner Mistral AI, the company developed a solution to analyze feedback from thousands of monthly store interactions. This gave Too Good To Go insights into key issues and led to actionable product improvements and a more tailored approach to partner retention.

Opportunity | Investigating the Cause of Retention Issues

Too Good To Go operates a marketplace where bakeries, grocery stores, and other food businesses can sell surplus food at discounted prices. With over 100 million registered users and nearly 200,000 active stores, the company has saved over 400 million meals from going to waste. When some stores stopped supplying or left the platform entirely, Too Good To Go struggled to understand why.

The company’s existing processes for performing a root cause analysis were inefficient and didn’t provide the depth of insights needed to address retention issues effectively. Too Good To Go identified an opportunity to examine its sales team’s interactions with stores that had stopped supplying for a period, seeing these conversations as valuable data points that could inform improvements to its product and marketing strategies.

Solution | Achieving 85–90 Percent Accuracy in Categorization

Too Good To Go built a solution using Mistral AI’s models to analyze and categorize the reasons that stores stopped supplying on its platform. The solution is built on Amazon Bedrock, a fully managed service that offers a choice of high-performing foundation models. To develop the solution, Too Good To Go extracted information from sales team interactions, establishing broad categories for common issues and manually categorizing a subset of entries to create a training set. Through careful prompt engineering and iteration, the team achieved 85–90 percent accuracy in categorization.

The team chose Mistral AI in Amazon Bedrock for its multilingual support, strong performance in categorization tasks, and cost efficiency. “The main advantage of using Mistral AI models within Amazon Bedrock is that we had a lot of confidence in the system we were working with,” says Daniel Redgate, product analytics lead at Too Good To Go. “We didn’t have to worry about data being sent to an unknown service provider to an unknown location. And we knew that if we wanted to ‘productionalize’ the model long term, then it would be much easier working within the AWS environment.”

Outcome | Rethinking the Traditional Onboarding Approach

The implementation has generated significant insights and improvements for Too Good To Go. “We’ve identified a completely new category of stores, which will require a different onboarding process and potentially different touch points with our sales team,” says Redgate. The analysis revealed that this new category represented approximately 20 percent of analyzed tickets, prompting a fundamental rethinking of the company’s onboarding approach.

Too Good To Go’s solution also uncovered that many stores were struggling with the platform’s scheduling feature, particularly around vacation periods, leading to unintended inactivity. This insight has prompted the company to improve communication around product features and consider changes to its user interface to make the platform more intuitive. The company is now exploring new applications of AI, including image analysis to provide more transparency to consumers and optimization recommendations to partner stores—furthering its mission to reduce food waste through technology-driven solutions.