 |
Dr. John Sviokla is co-founder of GAI Insights. He previously was a strategic adviser at Manifold and former senior partner and chief marketing officer of PWC. He has almost 30 years of experience researching, writing and speaking about digital transformation — making it a reality in companies large and small. He has more than 100 publications in many journals, including Sloan Management Review, WSJ and the Financial Times. |
Paul Carroll
To jump right in: You and your colleagues are living on the cutting edge of generative AI, so what are some of the latest examples you’ve seen of powerful uses?
John Sviokla
There’s amazing work being done at Blue Cross/Blue Shield of Michigan. About seven years ago, they began moving beyond functional excellence to build a platform that set them up to offer all sorts of new services to keep customers healthy.
Ten years ago, they probably had 1% or 2% government business. Now it’s more like a third, and something like 30% of the premium dollars for Medicare Advantage are going to preventative care. That can even mean things like helping pay a light bill or getting someone a pedicure, because it turns out that when they give away pedicures they discover necrosis [dead flesh] far earlier, before it can lead to all sorts of complications. They save hundreds of thousands, if not millions, of dollars.
When AI came along, they were primed to take advantage. They now have three generative AI products. One gives customers an idea of what benefits are available to them and allows for member self-service. This is super important, given how many benefits there are. One improves members’ security profile, which is needed because of increased access. The third provides an ability to review and manage contracts, which are putting pressure on the insurer because of all the new supply of services and providers, both medical and nonmedical, across the platform.
They’re not only offering those products themselves but are making them available to the other Blues.
Jerry is going wild with customer service. We’ve written up a case study. They have about 5 million customers and help with insurance selection, refinancing and other issues related to cars. Since April of last year, they’ve been using generative AI to handle customer service via chatbots and text. They went from just over 50% of people getting responded to within a day to 100%. Most now get a response within 30 seconds. They've gone from 100% of issues going to a human being to 89% going to the robot—and the 11% that remain with humans get there faster for obvious reasons.
They're way more scalable. They can grow this business without having to grow the customer service function. ROI is about $4 million a year.
Drug companies say generative AI has decreased the amount of time from discovery through clinical trials by 60% to 80%. J.P. Morgan claims over 300 productive capabilities. Studies from BCG, Harvard, MIT and so forth find increases in task productivity of anywhere from 10% to 15% at the low end to as much as 50% to 60% on software.
Paul Carroll
You and two colleagues recently published a piece in Harvard Business Review about how companies aren’t designed for generative AI. What do they need to do so they can accommodate these capabilities?
John Sviokla
The big thing is that you have to redesign work to account for the fact that employees have a new conversation partner: one made of silicon.
To design for that new dialogue, you have to understand the task at hand, work to discover the new interaction patterns between employees and AI and then codify those patterns and spread the frame. But this isn’t the same as a linear decomposition of tasks, which is how most systems development is done. This is much more about shared discovery, because as the individual is doing the work their job is changing.
They’re conversing with a different sort of conversation partner, one with a hive intelligence and encyclopedic knowledge. And the conversation is codified into the knowledge base.
We've got a turbocharged conversation.
Paul Carroll
How do you recommend people get started?
John Sviokla
I think you get started with low-risk parts of the organization. That’s certain kinds of customer service and certain kinds of inquiry, such as people looking for a job or employees with questions on benefits or policies.
Then you think of this like the Quality process. You want a certain number of white belts, green belts and black belts. White belts, very much like in Quality, have some basic understanding, such as what an LLM [large language model] is and how prompting works. They need to know how to train a model. A green belt knows how to teach a white belt and has managed at least one project to completion. Black belts get more into the technology. The white belts are basically smart users, green belts are kind of in the middle and black belts are closer to people who actually start to build stuff.
Paul Carroll
You’re always working on the next idea. Where do you think you’ll go next?
John Sviokla
We need to get past sequential processes and understand how to redesign team-based work, so I’m thinking about collective cognition and how to take advantage of our new partner, the machine. We’ve had 120 years of automation of the physical world, 67 years of automation on structured data and 18 months of automation of unstructured data. There’s lots of opportunity.
Paul Carroll
I get excited about all the unstructured data that’s becoming available. I see companies using AI to make claims or underwriting more efficient, but getting access to all the unstructured data out there will take us to new levels of understanding.
John Sviokla
A lot of times, underwriters are working on descriptions that are pretty thin, that are historical. Well, this is about enriching those descriptions—understanding the semantics, understanding the functional interaction, being able to probe and assess all the dimensions of a risk. You're really expanding the bandwidth of underwriting radically.
Paul Carroll
Thanks, John. It’s always great talking with you.
