This is the first entry of the Praxos Development Diaries (“Dev Diaries”, for short). Every month we’ll be discussing the design choices, decisions, successes, and challenges behind Praxos. In short, the inner workings of an AI company developing software for insurance brokers.
My name is Lucas, and I'm the co-founder of Praxos. Two months ago, my co-founder Soheil and I began our project of building our very first feature, the Praxos AI Proposal Builder. Our goal was to help agents craft proposals in 3 clicks with no setup, regardless of the lines quoted and an insured’s risk profile. This first step has been, however, anything but easy.
To call our first model an Artificial Intelligence would have been a misnomer – “Artificial Dummy” being a more apt term. The results spoke for themselves, too, as the AI churned out proposals that were little more than jumbled jargon and incorrectly classified lines.
Quite a few tens of cups of coffee later, we asked ourselves: “how do agents learn about insurance?”
[I'll pass it to my co-founder Soheil for this section, who will talk more about what this means, and how we made it happen.]
Hi, this is Soheil. I'm the co-founder of Praxos, and I mostly deal with technology development. I'll be talking a bit how we cracked this problem in the past month.
How to teach an AI Insurance? Fundamentally, the question is almost comical. How do you teach an AI, an entity with no concept of risk, about a product made to mitigate risk? After all, insurance was born out of a very real need from sailors going on long journeys, who wanted the assurance that their families would be cared for if they were to drown at sea. It grew to protect people against the abject misery of having their livelihood destroyed by natural disasters and having to build from scratch. How could an AI ever understand something so very human?
Thinking that the best way to teach AI insurance would be to have it try to answer the questions a client could ask about a quote, we began to write out a massive list of questions. It took us embarrassingly long – please don't ask how many tries we gave it – to realize the futility of this approach. A single page of a simple personal policy will already have the answers to more than a dozen questions, and when working with longer documents, a single question may have dozens of answers, depending on minute details of the question itself.
Clearly, this wouldn't work. What is more, even if it could answer every question, it simply didn't understand how they related to each other. it was just making thousands of oddly shaped puzzle pieces that wouldn't (and frankly couldn't) fit together.
But what was the next step? We considered using the document layout to make better puzzle pieces. After all, if the document already fit together, then if we made all the AI reshape the puzzle pieces it generated based on what the document looked like, they would be able to fit together, right?
The reality was, well, different. There was simply no uniformity in the ways different companies shaped their proposals, and thus, the AI could not learn anything by looking at the structure. The puzzle pieces were only deformed further, leaving us with what amounted to some sort of multicolored word salad.
But it was, in fact, this word salad, which paved our way forward. We had a crucial realization: if you mince and dice each insurance document into a salad of words, minced and cut differently according to the layout, salads made of the same ingredients would still taste the same. This is of course a long winded and overly metaphorical way to say that auto insurance salad will be different from property insurance salad, and liability insurance salad will have its own unique, positively exotic flavor. Have the AI taste a few thousand auto insurance salads, forcing it to tell them apart, and it will slowly become the world's foremost auto insurance sommelier.
And so, that became our path forward. Our AI gorged itself on 500 thousand insurance documents across all lines, sublines and companies, learning to appreciate the slight oaky notes of Collision coverage and the sublime hints of Fiduciary Liability Third Party Claims coverage. Knowing that two documents were both auto insurance, yet that they were clearly different forced it to notice the little differences that made them unique, and soon, it could break down every document of insurance into all the different pieces that made up its unique flavor.
Armed with this extensive knowledge of how insurance documents were similar and different at their word salad level, we asked our AI to build puzzle pieces again. This time, however, the results were very different. It knew that the hint of collision coverage could not come without adding a few cars into the blend at just the right time, and so, it began creating the RIGHT puzzle pieces. suddenly, it understood insurance as humans do: a unique product, created by adding hundreds or thousands of non unique ingredients in just the right way.
Today
Last week, we rolled out updates that were a direct result of these breakthroughs:
- An online dual display editor to help you review and edit proposals and quotes side-by-side. Direct links (work in progress) to source pages are also included. This should eliminate time spent going back and forth between files.
- Broader coverage support, including personal lines and employee benefits. You can skip spending time on creating proposals for these lines too.
- Template support to style proposals according any agency's standards. Because spending time styling these documents can quickly become a time sink on its own.
- A new AI engine with near-perfect accuracy. Small brag: we trained this beauty on over 200,000 documents.
To us, the greatest possible achievement would be turning the proposal process into a fire-and-forget task for agents.
The Future
As I think about next month's client meetings, and our own product development roadmap, I cannot but jump with enthusiasm.
Picture the following scenario: you’re having an introductory Zoom call with a prospective client. This is the CFO of a mid-sized manufacturing company specializing in eco-friendly building materials. Having recently experienced growth from increased consumer demand and positive press for alternative materials, they are now exploring comprehensive risk management as part of their plans to expand operations and establish a new factory west of the Appalaches.
As the conversation unfolds, Praxos’ AI, which has been discreetly listening in on the conversation, gathers information, analyzing the client's potential vulnerabilities. Once the call concludes, it compiles a detailed risk profile, drawing from mitigation strategies and information collected in your book of business. It takes 5 minutes to craft a tailored programme and send it to the CFO, just as this information is also saved in your agency’s systems without the need to spend time on maintenance.
This is just one possibility. Truth is, automating the entire servicing workflow is now within reach. And with it, helping agents reclaim their time to focus on what they do best: advising clients and prospects on their risk.
This is the imminent future at Praxos. And we’re really excited to bring it about.
If you read so far
On Friday, May 10rd, Soheil and I will be hosting an Open House on MSFT Teams, 3:00 to 4:00pm ET. If you’re interested in talking insurance, AI and technology, sharing feedback, or just meeting with us… we’d love to have you!
An invitation link is available here. Drop by to say hi!