US Foods successfully built and scaled a generative AI sales tool by following a strategic, business-need-driven approach that leveraged modern technology, proprietary data, and cross-functional collaboration.
The initiative, led by David Falck, VP of Machine Learning Engineering at US Foods, transitioned from a proof of concept (PoC) to a production-ready tool that significantly enhanced sales productivity and reduced operational costs.
The process began with a clear focus on addressing specific sales challenges rather than starting with the technology itself. Falck’s team engaged with internal stakeholders, particularly the sales team, to deeply understand their needs and workflows.
This ensured the tool was designed to solve real-world problems, such as time-consuming administrative tasks and inefficient sales processes. For example, the tool ultimately reduced tasks that took sales reps 3-4 hours to just 20-30 minutes, giving them hours back to focus on customer interactions.
Technologically, the team utilized Amazon Bedrock, a fully managed service providing access to high-performing foundation models, to build the tool. In just 1.5 months, a single team member developed the initial PoC using Bedrock, demonstrating its feasibility.
The choice of Bedrock allowed for flexibility, enabling the integration of evolving generative AI components, such as Anthropic’s Claude models, while maintaining a modular architecture. This adaptability ensured the tool could scale and incorporate future advancements.
A key factor in its success was US Foods’ modernized data warehouse, which provided high-quality, proprietary data on ingredients, food lists, and other assets. Falck validated this data’s ability to deliver unique insights that generic AI solutions couldn’t replicate. By grounding the tool in this rich dataset, the team created a differentiated solution tailored to US Foods’ specific business context, enhancing its effectiveness for sales optimization.
The PoC was rigorously tested to confirm it addressed sales challenges effectively, showing clear value over off-the-shelf alternatives. After validation, the tool was rolled out to over 4,000 team members in a pilot program.
This rollout was met with enthusiasm from the sales team, who provided feedback for feature enhancements and new applications, fostering a collaborative optimization process. The modular design and ongoing input allowed the tool to evolve, meeting real-world needs as they emerged.
To further amplify its impact, Falck’s team partnered with US Foods’ digital team to integrate additional analytics and machine learning models. These enhancements improved product assortment decisions and enterprise search capabilities, extending the tool’s value across the organization. Cost optimization was also a priority—techniques like caching and selective use of Claude’s models ensured efficiency without compromising performance.
The scaling process was supported by Amazon Web Services (AWS), which provided not only the technical foundation (e.g., Bedrock and Pinecone for vector search) but also strategic guidance through programs like the AWS Generative AI Innovation Center. This partnership helped US Foods move from experimentation to production in 2024, aligning with a broader trend among AWS customers.
Ultimately, the success stemmed from starting with a business need, leveraging cutting-edge yet flexible technology, utilizing proprietary data, and fostering collaboration across teams. The result was a scalable, impactful tool that saved approximately 500 days of sales team time annually, transforming how US Foods supports its sales operations.
