Advanced Computing in the Age of AI | Friday, March 29, 2024

Geico Walks with Watson on AI Journey 

IBM CEO Ginni Rometty (source: IBM)

“A journey of a thousand miles begins with a single step” intones the ancient Chinese proverb, and it can be applied to enterprise AI journeys. But what’s the right first step? And what steps should be taken after that?

This was the big topic of conversation at IBM’s recent Think conference, an enormous industry confab that took over several square blocks of downtown San Francisco in and around the Moscone Center.  In her keynote address, IBM CEO Ginni Rometty declared the AI era is coming to the end of its first chapter, characterized by “random acts of digital” and PoC AI point solutions – many of them utilizing customer-facing natural language processing (NLP) applications. The next chapter will move toward AI scaling, in which AI is infused in mission critical workloads across the enterprise.

“I think we’re all standing at the beginning of chapter two of this digital reinvention, and chapter two will in many ways be enterprise-driven,” she said. “It will have three dimensions: it’s all about scaling AI and embedding it everywhere in your business. It’s hybrid, driven by mission critical applications. And underpinning it is trust, and that’s all about responsible stewardship of data.”

Scaling AI from prototype to enterprise was discussed at sessions throughout Think by IT strategists from the insurance, banking, automotive and healthcare industries, among others, and they described a similar evolution: build AI one successful step at a time. Or, in Rometty’s phrasing: “think big, start small, move fast.” The obverse also is valid: beware of a first-time fail, which will reinforce AI aversion within the organization.

Rometty and Geico's Greg Kalinsky at IBM Think

Layered on the think-big-start-small approach is building in a technology infrastructure – comprised in part of big compute and storage, an “AI IA” (artificial intelligence information architecture) that also handles data cleansing and warehousing – that can support AI at scale, Rometty said, along with a conscious, strategic commitment by the C-suite to embrace AI.

This process has unfolded over the past several years at Geico – the $25 billion/40,000-employee giant that insures 30 million vehicles (along with houses, apartments, and so forth). Adoption of new technology, customer service and pricing are three tenets of Geico’s corporate strategy, according to CIO Greg Kalinsky, who appeared with Rometty during her keynote. Geico was one of the first big insurance companies to undergo major digital transformation some years ago. More recently, the company has applied IBM Watson NLP AI technology for customer support and sales interaction.

Kalinsky explained that it’s particularly difficult to achieve market differentiation and competitive advantage in the insurance industry because it’s “a little bit of a commodity.”

“It’s tough to find a foothold,” Kalinsky said. “If you think about it, we’re selling a product that you don’t want to have, that you’re forced to have, and once you have it you hope you never have to use it. Think about getting up every morning and getting motivated to sell that product. So how do you distinguish yourself? We do it, we believe, with outstanding customer service and with low prices. How do you do that? We do both with technology, that’s always been our mission.”

Having started as a mail order company more than 80 years ago, Geico over the ensuing decades evolved toward the telephone as the primary means of interacting with customers before moving to the internet and then mobile – by 2013, Forrester Research ranked Geico’s mobile applications no. 1 in insurance; of late, the company has transformed itself digitally in pursuit of greater efficiencies and to derive more customer and market insight.

A few years ago, Kalinsky began hearing about the potential of AI and, working with IBM, decided on a Watson PoC project that was “somewhat different,” he said, in that unlike other Watson applications, Geico’s would be customer facing rather than expert facing, such as applications developed in the medical field. The difference being that experts tend to know the correct next question to ask while customers may not, customers are less predictable – and therefore present a more challenging AI/NLP environment.

Kalinsky admits he had reservations about the first Watson application.

“Truthfully, I was nervous that people weren’t ready for AI,” he said. “I was worried that in our zeal to find a better customer experience, more dynamic and data driven, that we’d alienate people and scare them off.... So I’d say the first stage of AI at Geico…we didn’t have much more of standard than: ‘Do no harm.’ We didn’t want to alienate people, we didn’t want to lose business, we didn’t want to scare customers away.”

Another cause of concern for Kalinsky: he knew if the first attempt failed it could undermine, or at least delay, AI adoption at Geico.

“So we watched how this was going very closely,” he said. “We’d review transcripts from the early customer interactions, verbatim transcripts, to see how people are reacting in a conversation with Watson – because they didn’t know it was Watson.”

Kalinsky found reassurance in particular from a Watson-customer conversation that occurred one night at about 2 a.m., in which Watson closed a sale, then thanked the customer and asked if there was anything else the customer needed.

The customer replied, “‘No, thank you, you’ve been wonderful, I appreciate it. You know, we late night workers have to stick together.’”

That customers were not put off by (or know they were) interacting with Watson was a positive step. But Kalinsky and his team also found that using Watson with auto insurance customers wasn’t getting the sales traction they wanted. “We did no harm, but we couldn’t get the improvement we were holding ourselves accountable for. So we said, ‘maybe this application is just a little too structured to really feel the benefit.’”

So they re-directed Watson at selling renter’s insurance: “We were able to reconfigure it because it was a slightly less complicated application…in a way that we eliminated some of the traditional flow in the application, but then we mitigated the disruption to a customer who didn’t know what to expect next by layering in Watson. Bottom line: we achieved a 40 percent increase in our closure rate on those policies, it was a tremendous success.”

Having achieved that, Kalinsky & Co. went back to training Watson for the auto insurance application in pursuit of the “same sort of lift.” The team came to realize they weren’t fully leveraging Watson because they were applying it to the old way of selling, “that, if I’m being honest, isn’t terribly different than the old, structured, monolithic application form somebody would’ve filled out on paper after tearing it out of a magazine 50 years ago. So it was time to change.”

The goal, he said, was taking advantage of Watson’s dynamism, it’s ability to handle the unpredictability of auto insurance customer interactions, the randomness of the customer’s next question or comment.

“What we needed to do was take what we knew about the customer, in the form of data, break the application itself into microservices, and then recompile the application at different points in time, allowing Watson to lead the customer in a way that was meaningful to them, individually,” Kalinsky said, “and that’s the process we’re doing right now in earnest.”

The Geico-Watson lesson, Rometty said: to get the value of AI “you can’t just layer it on your old way of working. You’ll get some benefit, you’ll do no harm, but…then you broke it into microservices – because you wanted all that flexibility built back in.”

EnterpriseAI