Insurance Thought Leadership
I'm interested in what you’ve written about using agentic AI in distribution. Could you explain the framework for how you're thinking about the use of AI agents in the agent and broker space?
Abby Hosseini
AI has been part of the insurance industry for 25 to 30 years. We’ve been using different disciplines, such as natural language processing [NLP], deep learning, machine vision, and robotics, in the front office, back office, and middle office. A few years ago, we saw the rise of robotic process automation (RPA), and chatbots have been around for 20 to 25 years in the form of FAQs, NLP, and search engines. The latest trend involves generative AI, which produces content or knowledge, while agentic AI takes it a step further by doing something with that knowledge.
Agentic AI orchestrates processes or uses machine learning to achieve specific outcomes. There's a distinct difference between generative AI, which focuses on knowledge creation, and agentic AI, which can make processes happen.
Companies are trying to understand the new technology, deciding which large language model to use, and considering security, governance, and risks. There's been a lot of focus on choosing among Google, OpenAI, Anthropic, and Microsoft, but less on actually implementing the technology. A significant challenge is that, without good data, there is no AI. Many process owners in insurance and banking struggle with poor data quality or lack confidence in the data used for decision-making.
Insurance Thought Leadership
Why is the data not clean enough, and what can be done about it?
Abby Hosseini
Good question. That's a good segue into the distribution side of the house, because data is created by agents, mostly independent agents, as well as by customer service and claims adjusters. If you're lucky, you have good data entry tools and core systems with some governance or control around what to input into the system. However, poor data entry is causing data quality issues. A lack of governance around the definition of the data or the ability to understand what a field means is another issue. For example, if a field is called "premium earned," what's the exact definition of it? Because there's a lack of common definition and people have taken liberties with how they input data, you end up having challenges with creating a single view of the customer.
For instance, if I say "Jane Doe" versus "Doe, Jane," or spell the word "senior" versus typing "Sr.," the computer doesn't understand these differences and can't tell that the entries are the same. When machines operate on the data, they need some normalization or cleaning.
In insurance, we expect agents to cross-sell, and for them to do that they need a good understanding of the household and knowledge of what products the person or household is using. A lot of that is missing.
If you have four or five different underwriting systems, as most large carriers do, there is no connection between them. For example, when I was at Mercury, if you were a homeowner customer, you were instantiated in the home system. If you were an auto customer, you were instantiated in the auto system. But there was no linkage or index that says this person here is the same as that person there.
There are a lot of these kinds of problems, but there's also the fact that there's a new age of abundance when it comes to the availability of data, particularly from the internet. Being able to wrangle that data, purchase it, bring it in, and integrate it with your own data has been a challenge.
Insurance Thought Leadership
What’s the solution to the data cleansing issues?
Abby Hosseini
As far as data entry is concerned, you need a robust user interface with good business rules for every field. For example, when the pandemic happened, the company I was at realized we didn't have email addresses for about 45% of our customers. We asked ourselves why and discovered that the email address was an optional field in the agent portal. Someone, 10 or 15 years ago, decided emails were just nice to have. Then the pandemic hits, and you need to go digital, send notices and information, and you realize you can't quite do that.
We had to go back to our agent portal and make it clear that this field is not optional. But even then, people still mistype things. You need to have email address validation in place to catch these mistakes. Software can handle these issues, but there also needs to be human commitment to ensure the data entered isn't just lazy typing or missing critical information.
In insurance, particularly in personal lines, there's still a debate about who owns the customer—the agent or the carrier. This can lead to issues where the agent might not want to share too much information.
Insurance Thought Leadership
You've talked about three specific areas where agentic AI could be applied. Let's start with the sales assistant. What does the current landscape look like, and how might it evolve with AI?
Abby Hosseini
The concept of AI copilots is already coming to fruition in many respects. These assistants aid salespeople with next best actions, offering contextual products or help, while providing a lifeline for policy, procedure, or product knowledge that the salesperson might be missing.
In complex product sales, like insurance, it's impossible for a human to remember all the specific rules across different locations. For example, knowing the garaging rules in Wisconsin versus Arizona. AI solutions can provide this information contextually, in line with the sales process. Anything related to guidelines, policies, procedures, cross-selling, or next best actions can be facilitated through these Gen AI solutions.
The level of intrusiveness for these copilots is customizable. They can function as a sidekick that the salesperson consults when needed, or as a more supervisory presence overseeing the human's actions.
We're not even considering the direct-to-consumer model yet. In online quote-and-buy scenarios, customers often get confused about certain terms or need product education. They may have questions about what they're buying. In this context of self-service for complex sales, the AI copilot can play a crucial role in educating the customer and helping them navigate the buying process.
Insurance Thought Leadership
I have to say, I always find it funny when I hear the term “next best action.” The first time I heard that, I thought, well, why wouldn't you tell me the best action? Why just the next-best? I quickly figured out that the term refers to the best action to take next, but I still chuckle in my head.
Abby Hosseini
I think of it as an "if-then-else" scenario. When a customer calls and presents situation X, Y, or Z, there's a fork in the road based on that dialogue -- you either go this way or that way.
While you're solving the initial problem, you might identify an opportunity. That becomes what we call an "aspect section." For example, if you're talking to a homeowner, maybe they don't have an umbrella policy or a car policy. Or if you hear a dog barking in the background, you might ask if they have pet insurance. These opportunities become what we call the next best action, primarily in the context of cost.
Insurance Thought Leadership
Turning to the second of your points about how agentic AI can be used in sales: Beyond acting as a copilot, what would a sales agent in AI look like?
Abby Hosseini
Well, there can be instances where the AI isn’t just providing knowledge or product information but is managing a process. For example, in underwriting, if you have an expensive novelty car or a Monet painting in your house, you can't just write the policy; you have to have some proof. In the old days, pre-pandemic, we would tell the Porsche or Lamborghini owner to drive over to the agent and show them the car. The agent would go out, take pictures, and send them to the carrier. Then the pandemic hit, and nobody could visit anyone else.
So, how do you support that sales process? It needs to be a self-service process where the customer can take pictures and upload them on their own. The customer could call the carrier, and the carrier sends a text message back that says, "Click here." The agent then takes over, guides the customer through taking the picture, uploads it, and brings it back into the underwriting world. This process, which could take days or be very cumbersome before, is now much faster with a good digital experience powered by some sort of orchestration or AI. That's an example of assisting the sales process.
Insurance Thought Leadership
Pardon me while I go take a picture of my Lamborghini and send it to you.
Abby Hosseini
The other thing I wanted to mention, which I'm pretty excited about, is the potential of interactive AI in customer education. When you buy a new product, like a Cadillac EV or a Samsung TV, you own it and bring it home, but you might not know exactly how to operate it or have questions about it.
The ability to use interactive AI to educate the customer about the product is exciting. It's not just one-way communication but interactive. Today, we might Google questions like, "Why doesn't my TV turn on?" or "How do I turn off the child lock in the car?" If you're lucky, you might find the answer after some searching. But imagine if this process was more intuitive, where you could talk to an avatar and ask, "Tell me more about the child safety lock," and then immediately follow up with, "How long does it take to charge my car?"
That thick manual we used to get in the glove compartment could become an avatar you can talk to for contextual help. This isn't just about the sales process but also about post-sales product loyalty and experience, which can be really helpful.
Insurance Thought Leadership
I can imagine that sort of thing certainly applying to insurance. Instead of having to go through your big thick policy, you could ask, "Am I covered for fire?" or "How am I covered for fire?" and get specific answers.
Abby Hosseini
Yes, there are already some vendors out there that do this. They have systems where you upload your deck page, and they will interpret it and come back to you, telling you what your coverage is and what your exposure is.
Insurance Thought Leadership
To get to your third point on agentic AI: How do you see AI fitting into customer service?
Abby Hosseini
In customer service, there's traditionally been a two-pronged approach. There are about 1,800 chatbot vendors in the market, and chatbots have been around since the early 2000s. So, there's nothing new in terms of providing customer-facing FAQs. The focus has been less on customer service representatives having an FAQ system and more on self-help options for customers, which have been hit or miss. Some younger people often prefer not to talk to anyone and want to engage with bots to get what they need. However, the nature of insurance is such that you can't really sell anything with a chatbot. Eventually, you have to talk to a licensed agent. Sometimes, you don't have the digital processes to provide everything because you have to go through the agent. For example, with some carriers, if I buy a car and want to add it to my policy, I have to call my agent, who then has to call the carrier. It can take a day or two to get that done. If I want to add a name to my policy or drop a driver, all of that takes a long time.
With AI or some chatbot capabilities, I can now have a guided experience that asks about the new driver and then processes the backend changes to add that information. More and more digital policy support can be done with bots to a certain degree. You don't want to disintermediate the agent, and there are complex scenarios where you can't quite get it done without one. But if you look at your call center and analyze how many calls you're getting and why, there's probably a chance that 20% of those calls could be automated and avoided with self-help.
However, you have to promote the capability, as well. For example, in my previous company, we implemented what we called EFNOL, or electronic first notice of loss, for clients. Once we rolled it out, we had roughly about 10% to 11% adoption. I asked the head of claims why adoption wasn't picking up, and he didn't know. So, I went on Google and searched, "How do I file a claim with company XYZ," and the system came back with a phone number. It didn't mention that you could do EFNOL.
There you go. We didn't nudge the customer. We were still using old methods to advise the customer on how to engage. You can't expect that, just because you rolled out some digital experience, it will get adopted. You have to force it, nudge people, promote it, and make it a better experience than waiting on the phone. Those are parts of the strategy for promoting digital solutions. It's not enough to just build these agents and put them out there.
Insurance Thought Leadership
I love the theory of chatbots but often find myself yelling at them. Have you ever dealt with a chatbot you like?
Abby Hosseini
Quite a few. They help with the mundane things. Has my payment been received, or what's my bill? When is my policy going to be canceled?
But you have to give the customer a choice. I can start in chatbot, elevate to a click to call, or elevate to a screen-sharing experience or a video sharing experience. Sometimes you start in one channel and then you go to the other channel because things get more complex. But it gets very frustrating when the chatbot doesn't answer your question and doesn't offer you any other way to engage.
Insurance Thought Leadership
If you give me five options, none of them are what I want, and there's nothing that says "talk to an agent," I get very unhappy very quickly.
Do you have any final words of wisdom?
Abby Hosseini
When I was a CIO and CTO, we had 10,000 agencies and 40,000 agents. We provided some level of delegated management of the users and the organizational hierarchy of these large agencies. However, onboarding was a struggle. Compliance with NAIC and other licensing adherence was a struggle. Commission distribution, commission calculation, and analytics around agents' performance were always a struggle.
When you look at where good analytics and AI can come in, I think it's across the whole spectrum: onboarding, agent education, compliance, commission and incentive management, and then reporting and analytics from a productivity perspective. In every aspect, some form of machine learning, analytics, or AI can be applied to really modernize that space.
There's risk involved with not having properly licensed agents. There's always an appetite or a need for just-in-time appointments. There are all kinds of human errors with commissions, and there's third-party reporting to regulators that has to happen. When you look at the whole value chain of distribution management, it's antiquated at best because most carriers have legacy custom commission systems and clunky onboarding.
That got us thinking about how, with some of these newer technologies that have come to the market, we can automate and streamline that experience.
Insurance Thought Leadership
OK, this is super. I really appreciate your taking the time.
About Abby Hosseini
 | As principal and chief digital officer, Abby Hosseini leads the strategic development and evolution of Exavalu industry solutions and practice areas. His strategic consultation and assistance to client CIOs/CTOs during digital and operational transformation ensures organizational growth for clients. He has over 33 years of experience as a senior technology executive, including 22 years as CIO/CTO, driving large and complex IT transformations for some major organizations like Mercury Insurance Group. At Mercury, Hosseini also managed the company’s core insurance platforms, including custom software and Guidewire InsuranceSuite and InsuranceNow (SaaS) platforms implementation over 14 years. Hosseini holds a bachelor’s degree in mathematics/CS from UCLA and completed his MBA from Pepperdine University. |
