Join our daily and weekly newsletter for the latest updates and exclusive content on industry-leading AI coverage. learn more
American Express is a huge multinational company with approximately 80,000 employees. As you can imagine, whether they’re dealing with Fritz laptops, whether they’re struggling with WiFi access.
But as everyone knows firsthand, interacting with it, especially chatbots, can be an irritating experience. Automated tools can provide vague, non-specific responses or walls of links employees must click on until they reach something that actually solves the problem. That is, if you give up because of frustration and don’t click “Get Me a Human”.
To overturn this worn-out scenario, Amex injected the generated AI into an internal IT support chatbot. Chatbots interact more intuitively, adapt to feedback, and walk the problems to the user step by step.
As a result, AMEX has significantly reduced the number of employee IT tickets that need to escalate to live engineers. AI will become increasingly able to solve problems.
“It gives people the answers, as opposed to a list of links,” Amex EVP and CTO Hilary Packer told VentureBeat. “We’re back to work quickly, so we’re more productive.”
Verification and accuracy “The Holy Grail”
IT chatbots are just one of Amex’s many AI successes. The company does not lack opportunities. In fact, the dedicated council first identified 500 potential use cases across the business, reaching 70 at various stages of implementation.
“From the start, we wanted to make it easier for our teams to build and comply with Gen AI solutions,” Packer explained.
This is delivered via the Co-I enablement layer. This provides a “generic recipe” or starter code that engineers can follow to ensure consistency across apps. The orchestration layer connects users to models and allows them to exchange models in and out based on their use cases. The “AI Firewall” envelops all of this.
She did not go into detail, but explained that Amex uses open and closed source models and test accuracy through a wide range of model risk management and verification processes, including searched high-end generation (RAG) and other rapid engineering techniques. Accuracy is important in a regulated industry and the underlying data must be up to date, so her team spends a lot of time maintaining the company’s knowledge base and verifying and reformatting thousands of documents to source the best possible data.
“Verification and accuracy are the holy grail of generator AI now,” Packer said.
Reduce AI escalation by 40%
The most frequently used technology support feature in Amex – Internal IT chatbots were a natural early use case.
The first is equipped with a traditional Natural Language Processing (NLP) model. Specifically, we integrated closed-source Gen AI to provide more interactive and personalized support by integrating open-source machine learning bidirectional encoder representations from the Transformer (BERT) framework.
Instead of simply providing a list of knowledge base articles, Packer explained that chatbots will attract follow-up questions to users, clarify the issue, and provide a step-by-step solution. You can generate personalized, relevant responses summed up in a clear, concise format. And if workers have not yet got the answers they need, AI can escalate unresolved issues to live engineers.
For example, if an employee is experiencing connectivity issues, the chatbot can provide some troubleshooting tips to get it back to WiFi. As Packer explained, “You can interact with your colleagues and say, ‘Did it solve your problem?” And if they say no, it can continue and give other solutions. ”
Since its launch in October 2023, Amex has increased its ability to solve it by 40% without having to transfer it to a live engineer. “We’re getting coworkers halfway through,” Packer said.
85% of travel counselors report efficiency with AI
Amex has 5,000 travel counselors, helping to customize the company’s most elite Centurion (black) cards and Platinum Card members’ itinerary. These top-tier clients are some of the company’s wealthiest and expect a certain level of customer service and support. Therefore, counselors need to be as knowledgeable as possible about a particular location.
“Travel counselors are spread across a variety of fields,” Packer noted. For example, one customer may be asking about a must-see site in Barcelona, while the next customer is asking about a five-star restaurant in Buenos Aires. “It’s trying to keep everything in every person’s mind, right?”
To optimize the process, Amex has deployed Travel Counselor Assist, an AI agent that helps curate personalized travel recommendations. So, for example, this tool can extract data from the entire web (such as a specific venue’s opening, visit times, nearby restaurants) combined with its own Amex data and customer data (such as restaurants where cardholders are most likely to be of interest based on past spending habits). Packer said this helps to create an overall, accurate and timely view.
AI companions currently support Amex’s 5,000 travel counselors in 19 markets, with over 85% reporting that the tool saves time and improves the quality of its recommendations. “So it was a really, really productive tool,” Packer said.
While AI appears to be able to take over the process completely, Packer emphasized the importance of keeping humans in a loop. Information obtained by AI is combined with travel counselors and institutional knowledge to provide customized recommendations that reflect the interests of the client.
Because even in this technology-driven era, customers want recommendations from their fellow people who can provide context and relevance, not just general itineraries that are put together based on basic searches. “You want to know that you’re talking to people who are going to think about the best vacation for you,” Packer pointed out.
Assisting and coding companions from Ai-Enhanced
Among dozens of other use cases, Amex achieved 96% accuracy with AI “co-workers help center” (similar to IT chatbots). An enhanced search optimization that returns results based on the intent of the searched word rather than the literal word resulted in 26% improvement in response. AI coding assistant has increased developer productivity by 10%.
Amex’s 9,000 engineers use Github Copilot primarily for testing and completing code. Packer explained that there is also a talk-to-code feature that allows developers to ask questions about the code. Ultimately, the company wants to extend it to end-to-end software development lifecycle (SDLC) and API documentation.
In particular, Packer said that over 85% of coders have expressed satisfaction with the tool.
“Not only is it working, but when your colleagues interact with it, do they like it?” Packer said. “We had pilots who said we could achieve the results we wanted, but we haven’t got the satisfaction of our great colleagues. Do you want to continue doing that? Is that really the right outcome for us?”
Source link
