United Airlines sets its flight plan for gen AI success

CIO Jason Birnbaum has ambitious plans for generative AI at United Airlines.

With the core architectural backbone of the airline’s gen AI roadmap in place, including United Data Hub and an AI and ML platform dubbed “Mars,” Birnbaum has released a handful of models into production use for employees and customers alike.

Chief among these is United ChatGPT for secure employee experimental use and an external-facing LLM that better informs customers about flight delays, known as “Every Flight Has a Story,” that has already boosted customer satisfaction by 6%, Birnbaum notes.

The flight-status service, released earlier this year, gives customers faster, more transparent and context-rich information about flight delays and updates, the CIO says. United has used the storytelling LLM to send out more than 100,000 digital notifications to date and has identified roughly 90 other use cases in the pipeline, Birnbaum adds.

As part of its storytelling ethos, the flight-status LLM will specify, for example, which precise weather event may be affecting a delayed flight and provide quick and useful information to customers about next actions. This has proved to be far better than broadly announcing that a plane is delayed due to mechanical issues or bad weather when it is sunny outside, Birnbaum says. 

“We worked hard to fine-tune this model to take operational feeds, notes from our operations teams, the crew, and all these different sources of data, and have AI take all this data and create a narrative that is more transparent, empathetic, decisive, and clear as we can be,” the CIO says. “As opposed to a canned message, we try to write a specific story about what’s going on with your flight. People hear the specifics, and they understand it and their blood pressure goes down. Our storytellers still look at the messages to make sure they’re within the realm of what we want to communicate but we’re getting much more comfortable with it.’

Fueled by cloud

United’s early success with generative AI wouldn’t have been possible without its prior shift to the cloud. Today, United has a closeknit partnership with Amazon Web Services that was detailed by several executives at Amazon’s recent Re:Invent conference.

For example, United Airlines has leveraged the AWS Generative AI Innovation Center and partners to innovate new solutions and level up the airline’s digital organization to become a leader in gen AI, said Francessca Vasquez, vice president of AWS professional services and Generative AI Innovation Center, which has engaged more than 1,000 customers and supports 1,300 gen AI use cases in 50 countries.

“United is just getting started,” Vasquez said of the airline’s “Every Flight Has a Story” initiative. “Historically United storytellers had to manually edit templates, which took time. United realized that gen AI could automate most of these messages, freeing up their staff to focus on the more challenging delay issues — and an informed customer, for sure, is a happier customer.”

United claims to be among the earliest users of the Amazon SageMaker ML platform, and it has leveraged its own United Data Hub and AWS Bedrock-based Mars ML platform to create this first batch of production gen AI LLMs. Birnbaum says Bedrock’s support for foundational gen AI models from a variety of vendors gives United developers flexibility, while the airline’s homegrown data hub gives them connected access to a vast amount of mostly unstructured data for AI development.

In addition to its flight status LLM, United has also developed LLMs for procurement and another LLM for enhancing manager-employee communications. Another new United gen AI app, currently in beta at customer sites, summarizes all relevant operational data required when there is a shift change, the CIO says.

“When there is a shift change, there is a long handoff between a person managing a complicated task and the person who comes to relieve the worker, and you’ve got about 45 minutes to an hour to handle the handoff in terms of identifying everything one needs to know for the operation,” Birnbaum notes. “We’re finding they can be faster and more effective during this change by taking in all the operational data and real-time alerting in our United Data Hub.”

United’s embrace of SageMaker and Bedrock as well as Amazon Q “is going to be a game changer for building data products,” said Mai-Lan Tomsen Bukovec, AWS vice president of technology, who pointed to United Data Hub as a transformational component in its AI journey at re:Invent.

“It gives unified access to your data, whether it’s stored in an S3 data lake or a Redshift data warehouse or it’s a federated data source,” she said, adding that “United Airlines Data Hub lets their data scientists and analysts self-serve on data requests, and it’s helping drive a data-oriented culture into their organization.”

Tomsen Bukovec remarked on the importance of having “clean data products” when it comes to implementing retrieval-augmented generation (RAG) for AI inference-based applications.

Moving beyond the pilot phase

United is among many enterprises shifting from gen AI proof of concepts to putting gen AI into production.

Based on Gartner findings, only about 4% of organizations were in production with generative AI services in March 2023. That number has increased to 21% in just 18 months. In March 2023, only 15% of enterprise organizations were piloting gen AI applications, which has since increased to 38%, Gartner maintains.

Gen AI LLMs moving into production has increased substantially, but most are for basic tasks that make operations more efficient, says Sid Nag, vice president of cloud, edge, and AI at Gartner.

“These are prime applications for leveraging AI and many organizations are doing these things,” Nag says. “But they’re very simple.”