Breaking the Back-Office Bottleneck: Insights from ITC Vegas & Insurance Innovators

Brian Weiss: Good morning or afternoon, everyone. Thank you for joining our webinar today. The title is “Breaking the Back Office Bottleneck.” Myself and Chris are going to really bring some insights that we brought back from both the ITC conference in Las Vegas and the Insurance Innovators Conference. Both of those happened within three weeks of each other, so we thought we’d not only bring some insights from those, but also some of our forward thinking from Hyperscience these days.

I am joined by my esteemed colleague, Chris Bloomfield, and you can tell we share a color palette in our wardrobe here—at least on the day we took these screenshots, Chris, right? I’m Brian Weiss, I have the black background. Chris has the white background. It says nothing about our disposition, just not a “light side of the force or dark side of the force” thing. It’s just the monochromatic way we’re going to roll today, Chris.

Brian Weiss: So, just as a quick agenda, we’re going to have a conversation about the common themes and challenges that both of us heard and continue to hear not only at the conferences, but out in the market in our customer base. Then we’ll give you a sense of our approach from a technology and strategy standpoint for these technologies. And then Chris is going to do a live demo specifically for some of the ways we’re using Gen AI inside the Hyperscience platform and live claims.

Before I do that, if anybody on the call is not familiar with Hyperscience, a quick introduction. We were founded roughly 10 years ago by data science and machine learning PhDs who, having just sold their first company to SoundCloud, turned their attention to bringing ML deep learning technologies—what we now call AI—to the core of the enterprise. What they realized was that with the appropriate compute power, if you could function and do what people do with, say, handwriting, you could solve really deep problems that generate a lot of friction inside an organization.

So they went after those to begin with, with a model-centric approach. Over those 10 years, we have grown into a very successful company, really built on the backs of a product-centric vision that AI and ML can solve these problems, but needs to do so in a responsible way. We have built a really innovative approach to “human in the loop”—so machines and people working together to drive an accuracy outcome with a harness around it—as well as a platform for making the creation of those models simple and easy. They realized those two things would be critically necessary to really bring AI into the core of the enterprise.

Brian Weiss: We are backed by tier-one investors. We’re well-capitalized—Stripes, Tiger Global, and Bessemer being our main investors. We operate in partner ecosystems with top-flight partners. And then the other part of the company is we grew up in very regulated markets, government entities, and financial services. This is for everyone who’s concerned about PII, private data, and secure data. We operate in air-gapped environments, for example. So, well before anybody had concerns about where data is going and how models are being trained, we delivered on-prem solutions as well as now private cloud SaaS and even in a FedRAMP High environment.

We are also well ensconced inside the insurance industry. We have a smattering of customers listed here who are benefiting from Hyperscience. This is anywhere where there’s high surface area to the public, where you’re getting messy forms, messy data—you all know the problems and what it creates. So whether it’s underwriting, claims, enrollment, these are all bread-and-butter use cases for us. And the customers on the left are using Hyperscience to generate competitive advantage by building trained models on their data on top of Hyperscience’s platform to get accuracy and automation rates in the 99th percentile. That’s sort of unheard of with the prior technologies that have landscaped this market.

Brian Weiss: So, Chris, as we jump into it, let’s talk about the conferences. What we heard and what people are struggling with. Some are a surprise and some not so surprising.

Chris Bloomfield: Yeah, so Brian, we spoke to a lot of people across those two conferences and we took all the feedback, all the different viewpoints, and condensed them down into what we’ve classed as the “Five C’s.”

Chris Bloomfield: I guess number one is no surprise really, and it’s really the Cost to serve for global insurers. What we see is that cost pressure from the top down: do more with less. Or in an ideal world, keep your cost to service the same, but do more with your customers. And typically what they do is focus in on the automation technologies, and they’re typically focused there with rules-based, RPA-based kind of systems. They want to now dip their toes into the world of AI and the promise that that will deliver. But for a lot of insurers, they seemingly can’t afford to be able to go into the AI space.

Brian Weiss: Yeah. You balancing that kind of long term…

Chris Bloomfield: Exactly. You think about the lifetime value of a customer and balancing that with short-term cost constraints, that’s a real headache for insurers.

Brian Weiss: Yeah. I heard the same thing. People try to understand where to place bets and understanding they don’t have unlimited funds, and how to go about it and sort of thinking about cost and overall cost reduction. How about Complexity?

Chris Bloomfield: Yeah. So three types of complexity I think about, Brian, and it was kind of largely echoed through the conversations that I had on the panel regarding complex claims. So really you’ve got the complex content as it flows inside your organization through your front door, through all the different channels that you have. Coupled that with the complex business processes that are built upon years of tactical solutions that have suddenly become the strategic norm. And then lastly, actually, do you know who’s actually interacting with you on a daily basis? There are bad actors out there. Is the content that they’re providing as part of this particular claim actually “hooky” in nature? How do I know that this is good or it’s fraudulent in nature?

Brian Weiss: Yeah. Staying in the—and we all, all these start with a C and that’s not by accident—in terms of Compliance, Chris, a lot of what I’m hearing is folks want to be able to leverage the modern technology, but look, we live in a regulated business. This is data that has to be managed responsibly, whether you’re training models or using models. So I think there is a lot of paralysis potentially, and I think almost rightfully so. We don’t really understand where the regulations are going to go. Everybody in the insurance industry knows that as a data steward of your customer’s information, you’re constrained by what you can do and how you can do it.

And then Customer Experience, right? I mean, Chris, for me on my side, I can tell the difference between a customer who’s trying to save cost and manage their bottom line versus one who has flipped and is now getting aggressive. Because it’s one thing to do things cheaper, but if you can do it faster and better—focusing on the customer experience, getting a claim across in days instead of months, being able to quickly evaluate onboarding—there’s a group of folks who, as you get more technology and investment, you get that advantage of customer experience. You start to play on that front.

Chris Bloomfield: Yeah, I totally agree. And we live in this kind of modern world, this kind of instantaneous high that we all get from interacting with different services. I mean, I can order food off of my watch, and I want to do the same with a claim. I want that as easy and seamless as that kind of experience. And that kind of leads us really, Brian, into that Competition piece.

Brian Weiss: Yeah. I mean, you get such an advantage of competing on customer experience, but you see all the frustration with the old processes, legacy systems. You just can’t go faster and get better, right? And then competitive, right? I mean, look, there’s a lot of pressure to use the latest and greatest. People are aware that—we call it FOMO, right, Chris? Everybody’s got it.

Chris Bloomfield: Absolutely agree. The people that came to the stand described it: they see themselves as this juggernaut that’s on this highway. And actually what they’re seeing in their rear-view mirrors are these little motorbikes, these little zippy InsurTech startups that are able to embrace AI. And they’re fast descending upon this…

Brian Weiss: Yeah. There’s a lot of eye candy in these AI startups. And what, you know, we’ve been in this for 10 years, and a lot of those folks are grabbing a third-party model that they don’t own, and they’re putting a wrapper around it, and they don’t really have control over what’s underneath it. So I walk the floor as a technologist and I can spot the sort of “Johnny-come-latelys” with whatever the latest model is underneath the hood. Lots of issues with that. But I do see that enthusiasm to want to try and use what’s out there and get ahead of it.

Now at the same time, look, there are roadblocks that we know of and our customers express on a continuous basis.

Brian Weiss: Number one is that people are handcuffed by legacy approaches. Everybody’s doing some sort of document process. You have to manage this inbound messy data in some way. And the entire first generation of technologies was built on rule sets. And rule sets cap out at what, 40, 50%? And so there’s this awkward relationship between what the machine does and what the people do. And do I focus on making the machine better, or do I focus on driving my BPO down to a different cost basis? And that ends up, you get hamstrung on that paradigm. And what our customers are doing to be successful is getting up and over that.

It’s no surprise that if you’ve been around for a long time and you’re in it, the long tail spend becomes such a drag on your ability to accelerate growth. The deeper you get into the, or the longer you take throwing—I would say I’d call it “good money after bad”—and not just making the switch to a model-centric architecture, the bigger this number is going to get. 70, 80%. That’s a real challenge for customers, and we can help get over that.

The other one is just the operational bottlenecks of the manual processes. They’re hard, the data’s complicated. These manual processes, and Chris, it’s not getting simpler. This vision that everybody’s suddenly going to agree to use the same form and it’s never going to change—it’s going the other way. Gen AI is going to create more variability because you’ll have dynamic forms, you’ll have all of that. It’s going to get considerably messier before it gets cleaner. So being able to keep up with that is an important investment to make.

And then the other bottleneck, I see this all the time as a CTO, is people are kind of paralyzed by the wash of technology options. Build versus buy. Great example. Do I buy somebody who’s offering me a bag of APIs and a very expensive consultant and data scientist to build something in-house? Do I try and make sense of it with the parts? Is there a platform that can do this job? If you look at the world of RPA and IDP and enterprise systems, should I be looking at rethinking the way these interact or just go for these sort of point solutions, best of breed kind of thing that I’ve been working with all along? Most CIOs out there—and this is by 2024—55% are going to increase investment. We’re now what, a month out? So either in that 55% or not, I met a lot of people who were in it and didn’t know how to make these decisions.

Now, AI is of course on the tip of everybody’s tongue in all of this. And by the way, for those in the UK, this is Rodney Dangerfield. He’s kind of like the equivalent of our Mr. Bean. So, you know, you can get it. Otherwise, it’s a little startling as to what “who is that guy” wants to do. Chris, what are you hearing about the AI hype? And coming from our position of being pretty deep on this for 10 plus years, what are you hearing?

Chris Bloomfield: Yeah. So what’s really interesting is like narrowing the focus and focusing in on areas that are going to deliver you ROI. But humans being humans, a new technology paradigm comes along, and what do we try and do with it? We try to make it solve probably the hardest and most cognitive problems that exist inside of our organization. And it tends to also look at front office kind of applications. But what we’ve found over the last 10 years, Brian, is that real focus and the gold for AI is actually in the back office and back office organization.

Brian Weiss: Amen.

Chris Bloomfield: So you mentioned earlier these swivel chair processes. Have a look where you’re spending a lot of time manually entering data in systems, or you’re outsourcing to BPOs simply because you don’t have the bandwidth or you don’t have the money there to be able to operate at those levels of exceptions. Or those back to that swivel chair where you dip in and out of multiple systems. These are the ones that are really ripe for AI disruption.

Brian Weiss: Yeah. And if you break it down by sort of the value chain… compelling customer experiences. The interesting part of all of these value chains is that the document is really at the center of it.

Brian Weiss: And so the back office becomes the unlock for most of these companies. If you can get ahead of the technology curve, first of all, customer experiences are driven by faster and better processing. Faster and better processing requires you to change and think about your operational inefficiencies. Where am I using people that I thought I could never use a machine, but I can? And then finally, operational efficiency requires you to modernize your systems. So modernizing your systems allows you to improve efficiency, allows you to deliver that customer experience. People are seeing this kind of a chain here.

The other piece I bring out there is that there’s a huge appetite to do Gen AI. To say, “You know what, if I had an agent that knew everything about my business, my contracts, my claims, my customer base, and I could just ask it questions along the way, wouldn’t that be great?” Yes. Now, what’s the key element? You have to feed it your enterprise data. You’re not going to take your enterprise data and feed it to a Chat GPT public frontier model. So the first step of Gen AI is actually getting a clean data estate for training, labeled, et cetera. And by the way, that’s what Hyperscience does. So there’s this realization that the technology that you’re using for just what you would think of as automation is also the unlock to these Gen AI use cases.

But Chris, let’s shift gears a little bit now and give folks a baseline as to where Hyperscience lives, our technology approach, and how we’re sort of moving this market.

Chris Bloomfield: Perfect. So let’s start really around that whole complexity piece. And Brian, you spoke to it just there: turning this kind of content into rich contextual data that can be utilized for Gen AI applications as well as your day-to-day business.

Chris Bloomfield: So here on the left, this is a non-exhaustive list of content that flows through your insurance processes. And the way I like to think about this, Brian, is simply: I’ve got controllable format—that’s something that I’ve created inside my four walls, you know, application forms, claim forms—which probably really represent only about 5% of the content. But the vast majority of it that flows through my business processes aren’t going to be in a controllable format. It’s something that’s been created outside of my four walls. And you mentioned earlier, with AI, this is only going to get worse.

Brian Weiss: And even when you control it, there’s variability, right? This is another version of the same thing. Next year, it changes. Like, “Oh, it’s going to be the same.” No, it’s not going to be the same. So there’s always this chasing of rules-based approaches that have been kind of the bane of the market.

Chris Bloomfield: Absolutely. How many times have we seen sort of brokers take an insurer’s form and put their logo on top and it kind of breaks automation?

Brian Weiss: Breaks all the old systems immediately. Yeah.

Chris Bloomfield: Yeah. And here on the right-hand side, your insurance processes, but they’re only going to be able to run as smoothly as the data that flows through them. So it’s really important to focus in on the data quality that’s fed in. So what’s really interesting about more traditional approaches, and even those with newer technology today, is that they focus on fixing that left-hand side through a document-centric architecture, making the document the center of the problem.

With that said, the way document-centric architectures work: either a document is processed, or it ends up in a document-centric exception queue.

Brian Weiss: And a garbage bin. Yeah.

Chris Bloomfield: Yeah, exactly. A team of people sat there. I always think about this like a teacher receiving homework, but I’ve got to read the whole thing. “Oh, that’s where the problem is.” And so I’m spending an inordinate amount of time reading an entire document to find out where the actual issues lie. Little wonder that we end up in a world where the BPO market is… staggering stat, isn’t it, Brian? Was it $550 billion?

Brian Weiss: Yeah. Something like that. It’s astonishing. The BPO market’s $550 billion of what machines have failed to get right, and that we have failed to actually make an efficient process around. That split is pretty dramatic. I often talk to customers—let’s not talk about technologists, talk about your people costs and why are you spending that much money where you are?

Chris Bloomfield: Yeah. And the reason why this doesn’t work, to an extent, is because it’s largely static in nature. It kind of acts like this dumb terminal. So I mentioned earlier: document comes in, doesn’t meet my template, does it meet my business rules? Oh, fantastic. That passes. If not, it fails. A simple pass/fail mechanism. But the problem is, Brian, it’s the documents that pass. And we’ve seen this many, many times when we scratch under the surface: the content within those documents that have passed are actually inaccurate.

Brian Weiss: Yeah. And you end up in this bunch of rework. It’s crazy. It’s shocking to me when somebody says, “Well, my STP [Straight Through Processing] is 90%.” But you’re going through some legacy system that “I put a hundred documents in and 90 came out,” and therefore, yay, I’m 90% in my technology. But then you read those documents and their accuracy is 11%. Like they’re horribly wrong.

And so the idea that we need to stop talking about STP and stop talking about the overall “what does it take to get a hundred percent of the documents right?” The legacy solutions are not accountable to accuracy. They’re just like, “Yep, I processed it.” Okay, that means me and you, or just our customers, have to go build some sort of accuracy QC after the fact, and then they gotta send it to people. So this relationship between the black box that’s not accountable for accuracy—they don’t even tell you what the accuracy is in a lot of cases, but it’s proud of its STP. Like, I don’t get it. Are you right? Are you wrong? Like, you still have to spend all of this time and energy at the document level. So it’s not even like it says, “Oh, I don’t know about this field.” It’s like, “Nope, I got it right, but I’m wrong.” So you’re kind of up to you to make it work. And that’s the problem with these legacy technologies, Chris. You end up with too much time, too much investment in the wrong place.

Now let’s talk about Hyperscience a little bit.

Chris Bloomfield: Yeah, absolutely. So we recognized this particular issue a long time ago. As you said, we’ve been in this business for 10 years, and we are very, very comfortable with the fact that we cannot possibly automate documents. It is just such a wide variety of documents that exist out there. But what we can do is largely automate the actual process of turning documents into highly correct machine-readable data. And the best and most scalable way to do that is modeling human-type behavior using machine learning, using artificial intelligence, and enabling us to orchestrate to target accuracies.

Chris Bloomfield: So we have this turnkey platform, so you can consume that as a service or, for certainly those organizations that operate in rarefied environments, we can deploy on-prem. And what we do is we drop in discrete models that are built upon your data and our digital workers interact with these models. But the key here, Brian, is actually you, as the insurer, can turn around and go, “Actually, the output of each of these models needs to be at least 98% accurate, or 99% accurate.” And we orchestrate to that.

And what’s really great is that we break down all of this kind of work into what I would consider micro-tasks. i.e., “What’s this page belong to?” “Where is this interesting data point?” “What does it say at this particular point in a document?” And we do the really smart thing and go, “Uh, I don’t think I can hit 99% accuracy on this particular task. Hey, human, can you come in and help me?”

And effectively what we’re doing, Brian, is borrowing a small sliver that was either outsourced to a BPO or you got a large sway of staff, and bringing them actually properly into the loop—not after the fact—into the loop with our digital workers working in tandem to ensure that we are able to extract that content at any level of accuracy that you want.

And then we can go further. Now we can start orchestrating the actual correct data with your business processes, with your business rules, and be able to surface up 360-degree insights of your insurance workflows and your actors as they flow through. And we see here on the right-hand side, these aren’t just goals, these are realities. And Brian, you’re gonna talk about a few different samples there. Now we are hitting those 99-plus percent accuracy targets that our customers are looking for us to achieve.

Brian Weiss: Yeah. It’s really high levels of automation. It is a fundamentally different approach about how to use AI and ML. So I, even at these insurance events, I walk the floor, people say, “Well, how does your model compare to Textract?” That’s the wrong question. What you need to ask is: “What do you do when Textract is wrong?” Do you call Amazon and say, “Hey, I got five documents that you were wrong on, I’m gonna send them to you and I want you to make your thing better”? This is not possible.

So Hyperscience does exactly that. You’re actually curating narrow proprietary models that get better over time. So don’t ask, “How good am I compared to Textract?” Ask, “How can I make that third-party model better?” And the answer is, you can’t. So you need to shift the conversation around: invest in a platform that allows you to control the destiny of the accuracy and the providence of these models. And by the way, data security, right? You can run it on-prem, you control the data, nobody sees it except for you, you have full providence over that information.

Look, we’ve been in business for 10 years. We have a lot of customers in the insurance industry who have been with us long enough that you start to see very significant results at the end of it.

Brian Weiss: Chris, I’m going to give a couple examples here along the lines of what I talked about before. Customer experience: Legal & General is a customer of ours since 2019. They’re using Hyperscience for new application and customer service processing. They have optimized their onboarding so severely with Hyperscience that they’re seeing outsized results. I mean, they’re on an absolute tear, and they’re winning awards for customer experience. So the benefit of getting all this operational work and getting model management is you start to really provide a different product out to the market. And Legal & General is doing that.

The Department of Veterans Affairs—it is maybe one of the, I mean, it’s among the largest insurers in the United States. They provide claims services to over 9 million veterans. We have been in the Veterans Affairs for I would say three years now. Since using Hyperscience, they achieved over $470 million in savings. And that’s just taking old rules-based processes and putting a model-centric architecture in play. The project actually got a lot of notoriety. President Biden called it out recently as exemplary use of technology for efficiencies. But what for us is probably the most important thing is that veterans used to wait almost three and a half months to get a claim processed, and it’s now down to about a week as part of using Hyperscience.

In another one, this addresses legacy systems: Corebridge. They’ve been a customer for a better part of, I think, almost five plus years now. And they just chunked away at their legacy processes and put in models, put in Hyperscience. So they now curate overall 70% reduction in data entry time. So you can extrapolate cost savings, but a great example here that they’ll call out: they’re getting 99% accuracy on handwriting. It’s 95% compared to 10 in the prior solution. And that’s not a typo about ROI, Chris, that is actually 1012% return on the investment to Hyperscience. So customers who make the switch to a model-based design that they control, and they flip their mindset from “there’s gonna be a garbage bin and I’m gonna send it to a BPO” to “you know what, I’m gonna put humans in the loop and not only speed up my processing but curate models at the same time,” they see outsized benefits in the market.

I also want to point out before we shift into a demo here is that Hyperscience is a small-ish company, and we have made some significant technology bets in the last 10 years. And really, if you now look at what analysts are saying about not only Hyperscience but the market, we made the right bets. Specifically, model architecture is the future. You cannot continue with a rules-based approach or legacy. A lot of legacy vendors are just bolting on some AI to the rules-based approach. So you don’t get the operational benefits. They go buy a third-party model that they can’t control and try and use it for exceptions.

But the idea also is that you will be able to use a portfolio of models. There will not be one model to rule them all. Their recommendation is invest in a platform that allows you to control a portfolio of models that are the right tool for the job, for the right price. Why would you use 60 billion parameters to get the square root of 34? You need a calculator, right? A sovereign model to do the work is really the approach. So this ensemble approach is where we’re going. Templates don’t work.

And look, they’re very complimentary of Hyperscience. We bring a bunch of IP and models to the table. You get those out of the box with us—handwriting, all kinds of great stuff like that. But we also allow you to integrate an LLM if you want it. Like, don’t use an LLM just for fun. Use it when it’s useful. And they’re great probability calculators for words.

Brian Weiss: Across the board, analysts are loving what Hyperscience is doing. I as a technologist particularly proud of the Forrester Wave. We are a fraction behind a company called Google, who is the farthest along in strategy. So we’re considered very, very visionary in terms of our strategy and approach. So you’ll see that across the board, our customers are getting the benefit, and the analysts are starting to see it.

So, Chris, we’re at the bottom of the hour. Let’s give the folks an example of in action in the insurance industry. Let’s give them a little flavor for it, shall we?

Chris Bloomfield: Yeah, let’s do it. So hopefully we’ll be able to bring this to life with a medical claims demo. So you’re going to see all of this different kind of content flow through Hyperscience, how we orchestrate to accuracy, and then how we also orchestrate to business rules and validations, and then also how we integrate into Gen AI solutions as well. So with that said, Brian, I’m just gonna share my screen.

Chris Bloomfield: So we mentioned earlier about controllable and uncontrollable formats, Brian. So this is a great example of the controllable suddenly turning into the uncontrollable. There’s a lovely structured claim form, but because it needs to be printed out and I need to fill it in by hand, I’m taking it from mobile, suddenly become less compliant. And as a human, I naturally make mistakes. We’ve got a lovely little pen in the way, and we’ve got some kind of shadow on there, which is typical things that we see on a day-in-day-out basis flow through the Hyperscience platform.

Chris Bloomfield: And then alongside that, there’s gonna be supporting evidence to back up this particular claim. So a classic example of what we would class as a semi-structured use case is around an invoice. But not all invoices are equal. Everyone uses different software mechanisms. Everyone presents the data in a slightly different way, even if that data’s consistent. So this is actually quite a complex invoice that we see here. It’s made up of multiple tables, and we’ve also got nested tables in here. So here, parent and child rows describing the treatment. And here’s the classic unstructured use case. This is supporting evidence from a recent cardiologist appointment. It’s handwritten, as you can see, a few mistakes.

So just before I started sharing my screen, I submitted these into Hyperscience, and we’ve got a specific workflow that has been created in order to process these different types of forms. So we can see here we’re in the realm of processing. And as a refresh, we’ve moved into what we class as “manual identification.” So let’s just drill in a little bit and see what Hyperscience and our digital workers and our models have done to date.

So the first thing you’ll notice here is that we’ve actually been able to correctly classify some of this content as it’s flowing through the Hyperscience workflow. So a digital worker is taking our classification models saying, “Hey, you need me to be 99% certain that I know what these pieces of paper belong to.” Well, yep. I’ve collated two pages to a hospital claim form. I’ve collated two pages to a medical insurance claim form. But these other handwritten notes, and there’s an email body as well that was submitted alongside this… but we’re gonna deal with that later.

But hey, Brian, I’ve got to a point now where you are asking me actually on identification to be 99.5% certain. Yep, it’s fine in certain fields, I’m struggling a little bit. I need to bring a human in. So I’m literally now gonna shift gears and I’m gonna be somebody that’s working inside of an insurance organization. I don’t need to be an SME here at all. And all I’m doing is literally helping the machine do some identification work. And it happens to be on this nested table.

Chris Bloomfield: So here we can see that the machine has correctly identified all of the child rows, but each of the parent rows associated with it. So I’m quite happy with the work that they’ve done now, but this is where the problem is. So a machine’s lowering its confidence, and “I’m not quite sure, Brian, I can hit your accuracy requirements. Hey, human, can you review?” And indeed, I can actually see that they’ve missed out this parent row here. So I can very quickly go in and just help the machine and say, right, okay, we’ve got some bed charges here. Let’s just add in the nursing charges over here, and we’ll extend that. And we can go in for the OT charges. And then lastly, for the professional fees, let’s just make sure that we are actually capturing headers correctly.

And that’s it, that’s my job done for helping the machine on that particular point. You know, driving towards this accuracy orchestration. Now what we’re going to be doing, the last piece here before we start looking at the validation part of the process, is finalizing data trust. And now we need to convert what’s written on the page into machine-readable text. And because this is demo, I just wanted to highlight to the audience today how that works.

So we’ll wait for a few more seconds as we are extracting the text from the page, and we can see that this is now being updated. This is our digital worker turning around saying, “Actually here, you’ve asked me to be 99% accurate with all the transcriptions. I just need a little bit of help this time actually on a different form.” And this will be the one that contains some of the handwriting.

So here, this is a UK postcode. It said, “what I think it’s read,” but as a UK resident, I know this is incorrect. So the machine, it’s actually done the right thing, it’s actually put its hand up and said, “I think I’m going to get this wrong. I can’t hit your target accuracy. Hey human, can you come and help me?” So literally come in and just help the machine. Over this date for whatever reason in this bounding box, it’s looking there and it’s going, “It looks like a date, but that’s really low my confidence. Hey, human, can you come in and adjust that for me?” So with 24th of May, 2023, and that’s it. That is all my role done now in terms of building that data trust.

Chris Bloomfield: So let’s have a look behind the scenes of what else this digital worker is doing, and we are able to bring that to life, Brian, with what we call a flow run. This is that workflow that’s been curated, built on discrete blocks which describe different functions. So we’ve done all this data, we are now building a claim shell, and we can start seeing that we’re building a picture. And we can see also here that each of these blocks that we are operating in are turning green, which means that digital worker is completing all of these different discrete functionalities on the content.

So we’re gonna use this claim reference later. Let’s have a look through and we can see things like entity recognition. So in highly regulated environments, we’re gonna need to understand where PII-based information is contained in all the documents, whether they are structured, semi-structured, or completely unstructured. We can see that we are highlighting using out-of-the-box models from Hyperscience to be able to determine person’s names, organization names, et cetera, and create on-the-fly redacted versions for your longer-term archiving. See things like your FOIA requests, GDPR, and that kind of thing.

And so, as this is a demo, Brian, we’ve created a scenario where we’ve fallen foul of the business rule. And this business rule is relatively straightforward: we do not cover for preexisting conditions. And actually what’s happened here is that this is a preexisting condition validation is required. So now we’re gonna curate all of this different data from disparate data sources. We are going to auto-classify email bodies and Gen AI models. And we have this kind of “better together” strategy, Brian, where we have the sovereign-based models that are able to really focus in on high accuracy curated on your data. Then let’s use Large Language Models for large language problems and help us think, summarize the content, easy to read in digestible formats for claims assessors. And ultimately what we are now doing is presenting that up in a validation interface.

Chris Bloomfield: So I’m gonna move out of here and I’m going to perform… Now this would be more of a subject matter expert kind of role, a claims assessor, rather than just helping the machine find some things or write some data on their behalf. Here I’m presented with a complete pre-claim. So we’ve got the claim reference number that appears up here. I’ve got all of the different elements that comprise that particular claim. And on this right-hand side, where all of that information and actually what Hyperscience and the digital worker has been able to pull together, is based upon all of this information—including the customer sentiment here—is that this should be rejected. And it should be rejected on the basis of “we do not cover pre-existing conditions.”

Indeed, I can click there and I can zoom directly into the part of the document where it says actually this is the date of the first symptoms. And you know, I’ve got other information here. I mentioned that use of “Better Together” and Gen AI. We’ve got a complete summary of information that’s generated from Large Language Models. So what I’m gonna do here, I’m gonna make a decision, and actually I’m going to agree with the Hyperscience digital worker, and I’m gonna reject this particular claim, and I can add some case notes, and I’m gonna complete that task.

Now, what we’re gonna be doing now, Brian, is wrapping this up so we can have a quick look back into the flow run. We’ve got all of this great data, it’s highly accurate. We’re sending out a just-in-time message, so bear with me.

(Chris’s audio cuts out)

Brian Weiss: Yeah, Chris, while you do that, I’ll point out two things here. Number one is we’re actually bringing the decisioning to the user. So because the dates don’t match of the symptoms to the actual policy—which we went and got from a third-party system and determined what the policy is and read that—that we’re able to recommend rejecting the claim. At the same time, back and forth for accuracy point of view… for example, we’re taking responsibility for the accuracy of the output here. This is very different from a black box approach. We just dump it to you…

Chris, I’m not hearing you, but I assume everything’s okay.

All right. So looks like Chris has lost a little bit of audio here, folks. So I’m gonna try and narrate for what he’s showing you here. What he’s explaining is that the blocks and flows process allows us to orchestrate an outcome. And that outcome really, and what you’re seeing here, is actually he’s putting a note out to the end user that their claim was denied. But our ability to not only understand the data that’s in there, but also be able to orchestrate an outcome or a recommendation based on data that might not live in the documents, that live in other systems, is a core part of the Hyperscience platform.

Brian Weiss: So Chris, I think got some audio issues there. I can’t hear Chris, but I’m assuming that we’re all good here. Just to close out, one of the hallmarks of Hyperscience in our business is really around security in the enterprise standards associated with it. We operate at rarefied levels. Obviously SOC 2 Type 2 accreditation. As of last week, we are FedRAMP High for our government customers. We provide worldwide 24/7 support, multi-data center availability.

The key part here is that you can deploy it your way. If you want to be on-prem, by the way, any of the three clouds—if you wanna use Amazon or Microsoft or we have a partnership now with Google—we’re on the marketplace for all of those, deployable in all of those places. We can also deliver it as a full SaaS product with various levels of security, depending on whether you’re in the government or not.

So technical advantages real quick here. Hyperscience was born as an ML company. We were not a rules-based approach that we sort of come late to the project. We started this idea of AI for the back office. And so that means that we’re very different. We have orchestrated “human in the loop” as a way to drive accuracy, to have an accuracy harness around the models that are producing data. We bring the user in to solve that problem in real time. We have orchestrated a way for model management. It’s going to be an ensemble approach, but you, the end user, can control the quality of those. You see the transparency. And when you bring in a third-party LLM to the Hyperscience platform, you get that same level of harness around. You get guardrails around their use of the model. And that’s really, really great for folks who are having to look at AI governance boards.

We have delivered that on a turnkey platform. So we are delivering end-to-end not just understanding the information you’re looking at, but that automation of the decision. So Chris showed you an example where we pulled the necessary information outta three different kinds of documents—tables, handwriting, et cetera—but then we helped the claims adjuster make a decision because these dates don’t match and it’s not right and the claim actually doesn’t cover that. And we brought an LLM to summarize that work. So orchestrating the next steps, the decision step after you’ve already understood that, is what Hyperscience is also very, very good at. And our customers get a lot of value from it. We automate not only the decisioning, but also the process of curating these models so they get you to that 99% that is the hallmark of a lot of our customers.

Brian Weiss: So I think what Chris has lost audio, which is okay. But what I want to do before we close out here is… I’m based in the US, Chris is in the UK, and I know a lot of you folks are from the European theater here. So, look, Chris has probably forgotten more about this space than anybody will learn. He’s a very, very deep technical expert. He’s seen a lot of these not only challenges, but also the modern approach to how to break through them. So if you’re interested, happy for Chris to meet with you folks. Scan this QR code, give us a call, we’ll give you more insights about where we think the market is going and of course Hyperscience’s specific opinion and technologies to get there.

Um, Chris, are you back? Say something for me.

Chris Bloomfield: I’m… Yeah, I do apologize.

Brian Weiss: That’s alright. This is why we work together. I kind of know how that demo’s gonna end, right? So if you go silent movie on me, I can just narrate for you. So hopefully based on the moving pictures in our combination of voices, you got a sense of it.

I know we’re at 45 minutes on the hour. There is one question I do wanna answer. Somebody asked here that “we’re currently evaluating Gen AI models for claims process and not seeing the results we hope for. How is Hyperscience different?”

Well, first of all, Large Language Models, they’re probability calculus for words. They’re meant to generate content. And they’re not really beholden to accuracy and they don’t tell you when they’re accurate or when they’re not accurate, when they’re not confident. So what we see is absolutely those models will get some things very correct, especially when there’s words involved. But when they’re looking at shapes on a page and things… I mean, we had a travel operator who had very stylized confirmations of your trip. And an LLM got it spectacularly wrong. They inverted the locations you were going from Barcelona to Munich on the wrong date because the date it was booked… So the confusion matrix when position on a page is involved is very high. And LLMs will also then try and generate content that’s not there, right? They will hallucinate. So if you’re looking for these 15 things and there’s only 13 of them, it’ll try and generate the other two even though it’s not there.

So there are lots of downsides to trying to push a VLM or an LLM to do what it’s not designed to do. And Hyperscience really solves that by giving you the ability to train a model on the specific data you’re looking for and tell it what’s correct. And it’s deep learning. It learns it and gets accuracy. And then we use an LLM to ask a different question like, “Hey, are they changing or are they canceling?” Those kinds of questions are great. So it’s really about finding the right tool for the job to get that automation and accuracy done.

And I know, Chris, I think we’re at time right now. We’ve got still looks like 39 folks on the line. So we do have a series of questions in here and we know who they’re coming from. So Chris or I will personally reach out and give you answers to those questions. And thank you all very much for joining us today, Chris. Uh, it’s nice to see you. Maybe next time I’ll switch and do the white background. You do the dark background?

Chris Bloomfield: I think so, yeah. And I hardwired my headset in next time.

Brian Weiss: Hey, you know, I mean this technology stuff has great promise, right? Whatever, until the audio goes out. But I hope this was useful time. Look, the insurance industry is changing radically. It should be. And I would love to see it change faster. If you’re interested in our help in doing that, please give us a call.

Chris Bloomfield: Great. Thanks.