From PoC to Production: US Foods Scales Generative AI With Solid Strategy

Generative AI is on virtually every company’s agenda — but many consumer goods companies are still struggling to move beyond the first proof of concept (PoC). To get a generative AI initiative off the ground, companies need a solid strategy based on clear goals and cross-functional collaboration.

David Falck, vice president of machine learning at US Foods, and his team created a new sales-focused tool using Amazon Bedrock, a fully managed service that offers a choice of high-performing foundation models from leading AI companies. In this blog, we’ll explore the steps David’s team took to successfully move this application from proof of concept (PoC) to production.

Step 1: Identify a Meaningful Use Case

When approaching any digital innovation, you want to start from a business need and work backward, rather than focusing on the technology itself first. That’s exactly what David and his team did: they talked to various teams across US Foods, including sales, not just to know what these internal customers wanted, but to deeply understand them and the context of their needs. Thanks to this research, David’s team was able to identify a truly durable, long-term need — to be able to do more, in less time. 

Fortunately, David’s team has cultivated strong relationships with the sales team and was able to dive deep into day-to-day workflows. The team identified time-consuming, repetitive tasks that could be automated, empowering salespeople to focus on high-value work, such as being in front of their customers. 

Step 2: Assess Use Case 'Quality'

Oftentimes, we see companies rushing to “do generative AI,” but without a clear strategy. This can lead to wasted resources and ineffective solutions.

The best approach is to prioritize a handful of the best ideas instead of spreading resources thin across many. One widely accepted framework is IDEO’s approach to design thinking. This approach looks at whether an idea is:

1. Desirable, from a user point of view
2. Economically viable
3. Technologically feasible

David determined whether his idea was desirable by including salespeople in the design and testing process early and often. To decide if the solution was economically viable, he worked with his strategy and finance teams to evaluate the cost savings and revenue opportunities that his proposed solution could drive.

David’s team also tested whether the proposed solution would be technologically feasible, given the state of generative AI today. They evaluated ways to build a flexible architecture that could incorporate new components, such as foundation models, as generative AI continues to evolve. 

Most importantly, David dove into US Foods’s newly modernized data warehouse and validated the quality and availability of proprietary information on ingredients, food lists, and other assets. Based on this analysis, his team determined whether this data was capable of providing unique insights that off-the-shelf AI solutions couldn't match.

Step 3: Build the PoC, Test, and Iterate

In just 1.5 months, one person on David’s team developed a PoC for the sales tool using Amazon Bedrock. During this stage, David tested the PoC and validated that the tool was able to address sales challenges — demonstrating clear, differentiated value that couldn't be replicated by generic AI solutions.

Step 4: Secure User and Executive Support to Scale to Production

Upon the completion of the PoC, David presented the application to his executive sponsors. He demonstrated how the new sales tool had the potential to save US Foods millions of dollars and create revenue streams in the future. The stakeholders then approved additional funds to build upon the PoC, with pilots across the country.

Step 5: Roll Out, Monitor, and Update

Currently, US Foods is rolling out the sales tool to more than 4,000 team members. This pilot program has generated tremendous enthusiasm among sales teams, which are actively contributing valuable feedback on feature enhancements and new applications of the tool. With this collaborative approach, David and his team can optimize and evolve the tool to solve more sales challenges and meet real-world needs.

At the same time, David’s team is working with their digital team partners to deliver additional analytics and machine learning models. These technologies will help optimize product assortment decisions and vastly improve enterprise search capabilities based on the data that US Foods is currently capturing and generating. This collaboration will further amplify the tool’s impact across the organization.

Applying These Steps to Your Own Generative AI PoCs

2024 is the year of moving to production with generative AI for many Amazon Web Services (AWS) customers like US Foods. As you begin to explore what’s possible, you will first need to determine how you can measure and track the business value of your application as you scale. Second, you’ll want to make sure that you have the right infrastructure in place. Continuously monitor and optimize your applications to account for costs, latency, accuracy, and advances in technology. Finally — and this is an important one — make sure you are using generative AI safely, ethically, and responsibly by implementing a solid compliance and governance structure.

AWS offers a variety of programs to help you get the most value out of your generative AI investments, including training for technical employees and executives; a Generative AI Innovation Center that pairs AWS science and strategy experts with your teams to help them plan, execute, and scale generative AI use cases; and an AWS Partners program that offers you access to partners that have demonstrated expertise delivering generative AI solutions on AWS. 

Ready to begin your AI journey? Learn more at aws.amazon.com/retail-consumer-goods/generative-ai.