How are enterprises using retrieval-augmented generation beyond chatbots? At a COLLIDE panel, leaders from UPS, TSYS and Shipt described a digital twin operations copilot, internal use cases that build trust, the risks of customer-facing RAG and how to prioritize generative AI under top-down pressure.
Key takeaways
- RAG grounds answers in current data. Choi said retrieving from external sources counters outdated training data and hallucination.
- Trust is the payoff. Viciana said richer, current context produces better outcomes users can trust.
- Digital twins plus RAG. UPS envisions an operations copilot that digests data, configures simulations and visualizes results for operators.
- Start inside the company. Internal use cases such as documentation and call centers build organizational trust before external launches.
- Start small, show value. Focused proofs of concept, not top-down hype, earn support for generative AI.
Retrieval-augmented generation, or RAG, gives large language models access to current, trusted data. At a COLLIDE panel in Atlanta, three enterprise leaders described how they are putting it to work: Youngjun Choi, Director of Advanced AI and Robotics at UPS; Kenneth Viciana, VP, Global Data & Analytics Products at TSYS; and Andreea Popescu, Senior Director of Data Science, Operations at Shipt, who moderated.
The conversation ranged from customer service to a digital twin of UPS's entire delivery network, along with the risks, governance and prioritization questions that come with each.
What is retrieval-augmented generation and why does it matter?
Choi said LLMs struggle for two reasons: their training data isn't up to date, and they pull from everywhere, which leads to hallucination. RAG addresses both by drawing on external sources such as tables, graphs and documents, converting them into embeddings stored in a vector database and retrieving the most similar content to ground an answer. RAG lets teams get "high confidence answers using external data," he said.
Viciana, whose TSYS team builds predictive products from transaction data and sells them to banks, said the real payoff is trust. Richer, current data gives models more context and produces better outcomes. Popescu added that RAG also provides something LLMs alone lack: a reference to the data sources behind an answer.
On scale, Viciana said RAG shifts much of the heavy lifting to external data sources, which makes models more efficient to deliver.
How is UPS using RAG and generative AI?
Choi described two applications. In the near term, UPS has deployed a generative AI product for customer interactions that he described as message response automation, which he said makes UPS employees more productive.
The longer-term project is a digital twin of UPS's end-to-end delivery network. Choi said UPS delivers 20 million to 25 million packages a day across 220 countries and operates 100,000 ground assets and 600 aircraft. The goal is to "make a decision digitally and then take the action physically."
When a hurricane hits, operators must digest weather and trailer data, configure simulations and interpret the results. An operations copilot could use RAG to digest the data, set up simulation inputs, run them through function calls and visualize the results, Choi said, giving operators accurate data for daily decisions.
He also highlighted internal personalization. UPS has about 500,000 employees, and each building has its own demand, package flow and layout. Generative AI with RAG could tailor operational planning to each building and cut the time operators spend digesting data.
What other RAG use cases are enterprises pursuing?
Viciana said TSYS started with internal applications such as documentation management and call centers. Proving value internally first builds "organizational trust" and buy-in before deploying externally, he said. In his current role, he is most interested in personalization: better understanding customers to make targeted offers.
Popescu said Shipt's parent company, Target, recently deployed a store companion that helps new employees quickly find process information and answers for customers.
What are the risks of RAG in customer-facing AI?
Choi offered a three-part test: is the system responsible, reliable and scalable? Responsible means strong AI governance so customers don't receive unethical or misleading information. Reliable means high confidence, with a human in the loop to fact-check. Scalable means a strategy for delivering real-time answers as databases grow across tables, documents and knowledge graphs, possibly with a hybrid approach.
Viciana said he doesn't see RAG itself as a limitation. The bigger issues are ethics, bias and the perception of models as a black box. On data, he cited the old phrase "garbage in, garbage out" and said generative AI doesn't change it.
Popescu noted that the gig work industry Shipt operates in is "not regulated that well but it's highly scrutinized," so reputation risk is the main concern. Balancing speed with responsibility means being "responsibly fast to the market," she said.
How should companies prioritize generative AI use cases?
Popescu said it was the first time in her career that a technology mandate came from the top down. Viciana's advice: "start small, show value." Generative AI won't solve every problem, he said, but focused use cases build support. His team runs a handful of proofs of concept at any time with a fail-fast mindset.
Choi said he works backward from the customer. Business owners define the problem, data scientists check feasibility and engineers build robust, cost-effective infrastructure. For his R&D team, he said, the job is to turn an ambiguous three- to five-year roadmap into a clear six-month path with deliverables that show value.
Popescu also raised sustainability, asking whether the environmental cost of AI should factor into ROI. Viciana said it isn't common today but could become an input. Choi said an operations copilot that reduces miles and trailer movements supports UPS's sustainability goals.
Explore enterprise AI use cases at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Registration details are coming soon.
Frequently asked questions
What is retrieval-augmented generation (RAG)?
RAG retrieves relevant content from external sources such as tables, graphs and documents, typically via embeddings in a vector database, so an LLM can answer with current, referenced data and fewer hallucinations.
How is UPS using RAG?
Choi described a generative AI customer message response tool already deployed and a planned operations copilot that uses RAG with UPS's network digital twin to set up and interpret simulations.
What are the risks of RAG in customer-facing applications?
Choi said systems must be responsible, reliable and scalable, with strong AI governance and a human in the loop to fact-check answers.
How should companies prioritize generative AI projects?
Viciana recommended starting small with specific use cases and proving value, while Choi works backward from customer and business problems.
Watch: Harnessing Retrieval-Augmented Generation (RAG): Enhancing AI's Knowledge & Contextual Understanding on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
