Summarize this content to 2000 words in 6 paragraphs
Amex GBT is about to deploy its first big generative AI project. And there are several more in the pipeline, the culmination of an AI initiative that the business travel agency established in February.
Skift spoke with Marilyn Markham, who’s in charge of developing a company-wide AI strategy, about the company’s progress since then.
Markham, vice president of engineering and AI strategy, is part of a team from various departments, established to identify ways that AI could be useful across the company. The team from 30SecondstoFly, an AI startup that Amex GBT acquired in 2020, leads engineering for the projects.
“I think that it’s not obvious what gen. AI is going to be helpful for, and we need all the people putting their minds to it,” Markham said.
“I feel like we’ve done most of everything we could do with current technology, and so this opens new frontiers for us.”
Amex GBT (American Express Global Business Travel) has become a giant since spinning out from American Express a decade ago. Last year, the company booked $6.3 billion in business travel on behalf of employees at about 20,000 client companies. Amex GBT has about 18,000 employees.
The First Projects
The first project is a tool to help travel agents quickly digest travel policies, which can be 20 pages or longer. It’s especially helpful for new hires, Markham said.
“They may have 100 clients they’re servicing. Do you trust people to have 100 different travel policies in their head and know them perfectly? With gen. AI, you just give it the travel policy, and you can interact with it. You can just talk to it normally.
“We found that it had tremendous accuracy. It really, really is good at finding that information.”
The next project is for email servicing. The AI reads through customer emails and extracts all pieces of the query — a flight, hotel, and car rental, for example — and pairs that with the respective company’s policy. It presents the top three options, and the agent can make changes if needed before generating the final offer for the customer.
Today, agents have to do a lot of manual searching and re-searching, then a lot of copying and pasting in an email. “They don’t need to do that anymore. They get to review and curate, and have more head space to maybe personalize it.”
“All in all, the whole thing is faster than the original, and it’s a better, more consistent experience for the traveler.”
Markham spoke at the Skift Data + AI Summit in June about early progress on the AI initiative at Amex GBT.
Taking Ideas from Across the Company
Markham started the year by writing an AI policy for the company and establishing an AI oversight committee, which includes people from the legal team, privacy, compliance, HR, talent, and commercial. “And that actually just tells you how far reaching gen. AI projects are,” she said.
The team built a secure internal platform, powered by third-party generative AI, for staff across departments to experiment. Staff members aren’t allowed to use public generative AI websites for work because of data privacy concerns.
“Every few weeks, a team who had asked for access pops up with an awesome project that … delivered lots of efficiencies, or it allowed them to do something they didn’t think they would ever get funding to able to do.”
“It’s just been such an unlocker of creativity in the company.”
“Once you get a real product idea, you go into normal product development cycles. You have to go to the product team and pitch your idea, pitch your ROI, and then get that approved.”
“We’ve asked everyone to be open minded that AI is not your silver bullet.”
Human resources, for example, is looking at AI for end-of-year employee surveys. “When you have thousands of employees around the world and they’re using different languages, it’s a lot of work. But you can ask [a large language model] to summarize the top three sentiments. I wasn’t expecting HR to be such a big adopter, but they are.”
Determining Return on Investment
It can be fun to experiment with AI, but Amex GBT is still a business. So projects that get into the pipeline have to have a strong business case, Markham said. Generative AI companies often charge enterprise clients per word, so it can be expensive.
“That’s mainly the reason why some of the [proofs of concept] did not succeed. It was not that it wasn’t able to do the job. It was because it was too expensive at our scale.”
The company also looked at tools like Microsoft Copilot, which charges enterprise clients $30 per month per user, according to its website.
“We were not able to justify the cost. Yeah, it’s helpful. It can help you write your email a little faster. It can maybe correct your Word documents. We didn’t find that it generated good PowerPoints.”
“Multiplied by 18,000 people at the company, it’s becoming real money. You get into the few millions.”
Quotes have been edited for length and clarity.
