How is AI moving insurance workflows from ingestion to action?

Transcript

Keith Raymond: We’re seeing it happen with human in the loop all the way through, making sure everything’s done accurately and correctly.

Cameron Scott: Submission processing is where we started as an organization. We thought that was the biggest kind of bite out of the apple of practical application of AI to assist with extraction of submissions as they come in.

Cameron Scott: And it’s quickly moving from, you know, the generative extraction elements to prepare for reviewing it as opposed to doing that data entry to the agentic flows, where we are seeing customers have tremendous interest in taking a step further, start communicating with the insured or the broker with the AI, with tailored, nice custom messages to let them know that they missed providing a key detail.

Cameron Scott: Cross-reference my underwriting manual or my underwriting required information and email them requesting their license number. Email them requesting additional information so that when an underwriter actually picks it up, that submission has not only been ingested, but handled.

Cameron Scott: And we’re seeing a shift from ingestion to handling where we actually have ready submissions as opposed to having to constantly look at a submission, fire back an email to the broker for a request for information, look at another, and really can help you with that immediate action, immediately actionable items, as opposed to just needing to respond for one piece of information.

Keith Raymond: It’s kind of in line with what we’re seeing. And, you know, agentic, I think, is moving that needle a little further and really making things from a speed of processing and a consistency of processing, you know, taking it to the next level.

Keith Raymond: If I look at, you know, the claims process and the first notice of loss, you know, we’re seeing agentic capabilities being used for the initial response, right? And some of the things that you just mentioned, Cam, making sure that all the elements have been received appropriately, everything is accessible, all that kind of thing.

Keith Raymond: And then responding back with a personalized email or something like that, all done with an autonomous agent. And then working its way down the line, I’ve already seen some workflows where autonomous agents are being used all the way from first notice of loss to payment in the claims processing, right?

Keith Raymond: Very rare, but we’re seeing it happen with human in the loop all the way through, making sure everything’s done accurately and correctly.

Keith Raymond: And in the claims processing, excuse me, in the underwriting, as you mentioned, I think there’s a lot of opportunity there from the standpoint of just, you know, efficiencies gained from initial submission, you know, the ingestion, summarization, initial risk analysis, you know, all with traceability.

Keith Raymond: So any underwriter can go back and look at, you know, the documents from which all the information originated from very quickly. So the essence of efficiency. Efficiency gain is big and really the outcome of that is that improved experience with the customer, broker, whomever they’re interacting with, whether it be an adjuster or an underwriter.

Lawrence Ngo: One is to be able to just purely extract text at a higher accuracy to begin with. Two is it’s more robust to, like, changes. So whether it’s a change in handwriting, whether it’s change in contrast in the different images, it’s able to handle that a little bit better.

Lawrence Ngo: And then I think the big third piece is that because there’s some intelligence behind the extraction as well, there’s a lot of weird things that happen on application forms, as you may imagine. So, you know, people not following directions, people putting things in different fields, and being able to have a layer that can kind of reason through that a little bit really is a huge value add from previous generations of extraction.

Lawrence Ngo: At the end of the day, it’s really important to have high accuracy of extraction because it, back to Keith’s point of efficiency, like your gain in efficiency is really a function of how accurate that system is.

Lawrence Ngo: How much intervention you have to put on, how much intervention you need on the outputs of the AI, and then ultimately to how much you trust the system at scale and at steady state too.

Cameron Scott: We need to dial that right back to what’s this very specific problem that we can solve that will prove trust, that we can wrap in proper governance to get people excited about AI, to get people excited about removing those mundane tasks from their desk.

Cameron Scott: Whether it’s automatic cancellations based on unstructured email content, or even prompting those for review by a user, preparing those cancellations, declinations, a very specific business case at first, as opposed to thinking that AI is just going to broadly replace an underwriter. That’s not the case.

Cameron Scott: We are much more seeing it as what’s the specific function that they can get off of their plate in order to spend more time doing value-add responsibilities, more time underwriting the risk to improve their loss ratios.