US & Canada collections in the AI era

(upbeat music) - Welcome to the latest episode of Better Dead. I'm your host Josh Forman, CEO and founder of InDeadin. And every month, we bring together speakers from across the global credit and collections industry to share insights and new perspectives shaping the future of debt collection. - Hello everyone. Welcome to another episode of Better Dead. I'm Josh Forman, founder and CEO of InDeadin. I'm joined today by Dave, Tony and Mike, Jen Sprater to have you with us and to do some introductions of yourselves in just a moment. I'm just gonna kick off on a little bit of housekeeping for everyone and then we'll get stuck into things. Firstly, just on the agenda, we've got an exciting agenda lined up, focusing on how AI is influencing how bound voice strategies, predictive analytics and how it can improve outreach and performance and then how to stay on top of the regulatory and reputational risks surrounding that. And just from an audience perspective, I can see there's quite a few people on already and joining today. We really want to open up the floor to questions and engagement from the audience. There's two ways in which you can do that. So I'm a free one who wants to take a shot out. We'd love you to join in. You can do a request call in, button at the bottom of the screen, check your mic and camera settings and then basically provide a short description of your question and then jump in and join us on stage on that. If you don't want to jump on camera, not a problem at all, you can just select that off. And for others, otherwise, we would just prefer to drop a question in our public chat, feel free to do so. Anyone that can't make the whole session don't have any issues jumping off. Everything is recorded and we will distribute it after the session today. All right, so let's get into things, to kick things off, Dave, maybe get you to introduce yourself first plus across the Tony and then into Michael on the indebted side. Josh, for sure you have us. Yeah, my name's Dave, I run credit and we've been in the collection space for about five years now. Initially, I've had different digital collections tools, direct consumer technology and stuff like that, but where our platform has kind of really pushed the boundary is creating a clearinghouse between the different collections industry stakeholders and the either for-profit or non-profit advisors who might coach consumers through, how to resolve their debt. And so it's a quickly growing segment of the consumer population. And we see a ton of utilization on both sides of the ecosystems in terms of different tech-enabled strategies to interact with each other. Yeah, excited for the conversation and to kind of play out what we're seeing as people, you know, dip their toe in the water, push the boundary, be on the bloody tip of the sphere, you know, the kind of full spectrum there. Yeah, also, thanks, Dave, for turning. Chris, to you. Hey, thanks, Josh. Tony Pampina and the president of CBV Collections Services, Canada. We are Canada's largest full suite receivable company being we're not just a traditional collection agency, we're a BPO as well as we're a large debt buyer in Canada. We've been around since 1921. I'm also the vice president of the RMA Canada. Hi, everyone, Mitchell. The VP of Data and AI in the US side of the business and I've been focusing a lot on the development of AI. Voice recently as well as I've been the implementation of data size and analytics in the US business I've indebted. So look forward to sharing some insights today. Awesome. Thank you, Mike. That's great. Let me jump in to sort of the agenda topics that I mentioned and how we'll build on from there. So I think the framing I'd like to use for this is something that's sort of, I think, fairly relevant, literally relevant to this month that I think is particularly helpful, which is around setting the framework at least for this, which is there's a new Apple release coming out in iOS 26, literally in the next few days, where it basically introduces AI powered call screening. And we've obviously seen sort of similar different versions of technology implemented over the last few years that have had a consistent impact on right-party contact rates and in the use of voices, a channel, et cetera. And I think this is a bigger expansion of that. And I think one of the things that it has an opportunity to do is obviously filter out legitimate calls, many of which applicable to some of our businesses on here and certainly those about clients and other stakeholders basically in terms of their businesses. So I think it provides us really good sort of framework to sort of talk about these things and questions for the audience around sort of how reliant we still are on voice communications or where it works and where it doesn't. How do we adapt to some of these opportunities and know we're going to talk about AI as an outbound voice utilization tool, but you also then have it almost as a defensive component here as well on the opposite side of things. So I think Dave Tony, not sure who wants to jump in first on your piece, but would love sort of give you a couple of adopt points to work from and then build on from there is how AI voice agents are changing outbound contact strategies and the different use cases you guys have seen across collections be it in the broken payment reminder space and settlements, inbound outbound. I'll wait, are you guys taking it from there? - Okay, why don't I jump in? So in full transparency, CBV, I would say we were admittedly late to the ball when it came to AI. We only started looking at it seriously in about 2023 after countless pitches and a lot of disappointment. We began to think maybe AI just wasn't ready for our industry or we just were willing to adapt. Sometimes in 2024, they go as we saw that there was a shift and we felt that the technology started mature or we started actually meeting providers that were offering the services that we needed. After a 60 day proof of concept and 90 days of a full usage, it started to exceed our expectation. We just didn't adopt the AI, we invested in it, became a shareholder of the company called the Brela Tech Solutions. We see that outbound contact strategies has always been about reaching the right customer at the right time. But we started to see that the landscape started to shift dramatically with the rise of all this call screening that you just talked about in AI driven agents. And today consumers are protected even more with all these different carriers and apps and devices and filtering out these suspected spam calls. And that's what our business, even though we're a collection agency, but we're not marketing people. That means your traditional volume-based dialing was really no longer affected. And in my honest opinion, if you're not at the infancy today or stages of implementing AI or you're not finalizing or onboarding an AI company or a vendor, you're really doing a service to yourself, your company, your team. Today alone we drive over 151 million AI-powered interactions. That's SMS, that's email, and that's through voice chat. And that's still with 400 employees. You know, the concern we see today is with all these different spam filters. You really need to have the technology and the intelligence behind it to circumvent that. There's also other companies out there that are diving right into this space. You know, my opinion is this should be driven by the carriers and not these alternative companies, but ultimately this is what we're starting to see at least up here in Canada. You know, the difference is between these companies and AI helping to spam, to reduce the spam. This is the synergy of our industry right now and it's here now. I hope that kicks off something. - Yeah, thanks, Danny. One of the things I think about guys is the first ones I think about this is from the consumer's perspective. You know, I get these calls all day long. My number's on conference attendance lists and things like that and then makes it into the ether. So at this point, you know, I don't think of my phone if somebody's not in my contacts. If they really want to get hold of me, they're going to get a hold of me and text me. Or email me first. So as I think through like, what's the framework that's durable? As we go through so much rapid change and just try to just update your frameworks very, very rapidly. The first break point in the waterfall for me is, do I want to talk to this person or do I not? If I don't want to talk to somebody there, so many tools of my fingertips, such that I don't have to. And I can just block that communication. Then the collection space that leads to very negative outcomes because then lawsuits are the only thing left to lenders. As they go through and that's obviously not the outcome that anybody want if it all possible. But as we go through, if you take the opposite side of the fork and the road and you think about all those many, many communications that people do want to talk. And it's just about making the trains run on time. Like you said, Josh, payment reminders, updates on, hey, you said you couldn't pay last month. Can you pay missed this month, that type of thing? Those types of things where it's really just a matter of lost in the organization. You've already established positive intent to pay and things like that. There's not a lot of creativity you need. It's just doing the blocking tackle. That's really, really low hanging fruit. And obviously, I have 26 is going to be a huge impact for the industry. I think most folks don't see it coming and don't realize just what to drop off. They're going to experience and they affect just their app and calling. In that world, it shouldn't be that surprising. If you follow the intent of where the carriers have been going and where the regulators have been going, it's one new tactical method for limiting app and communication, but it's not inconsistent with the themes. So as we think through, what is the future of the collection space, there's definitely the app and element of going through and saying, hey, we need to get a hold of consumers, make them aware of the accounts, you know, if those accounts slip through the cracks, but it's as much about creating inbound channels. And we think of a debt settlement as one of the groups that kind of pushes the envelope in that tech adoption and what's possible when inventing those to perform factors as we look at credit counseling agencies, financial laps, consumer attorneys, tick-tock influencers. Like there's there's actually a lot of different groups driving consumers inbound to engage, but those consumers come in with their own bias and their own expectation based on the channel they come in through, so you need to then adapt to that. But a lot of those tactical calls, I think of it as, how do you create as many tools that then you can play with and put together? - Yeah, just to add on a point you made, Dave, and then Tony yours as well, I think the example you gave of the amount of people just to call on a daily basis, right, of the ones you don't want to answer and you're just doing everything you can to block it. I think this is where the omnicunnel piece has a big role to play because the consistency of message across multiple channels is allowing us to be legitimate. So if you had a miss call from me, Dave, but then I follow up with an email and I get the text, like you said, they're an act consistent threat of like, "I can't know Josh, I can't know." Now I've got more contacts that's been delivered through multiple channels that helps me enable that ability to drive the inbound conversation, which I think is one of the key things you mentioned there and sort of using that. I've also, we've recently expanded into, particularly in the Middle East, also Mexico, but particularly in the Middle East where channels like WhatsApp are ubiquitous, right? Like they are everywhere. In this country, you do not text people, text seems what people did for like 30 years ago. No one sends text messages anymore. It's like, it's a stamp if you get a text message because no one sends them. And then phone calls are not on and you reset your banking passwords and everything through WhatsApp. But interestingly, if you look at the WhatsApp API infrastructure to enable communications, it's basically driven by a consumer opt-in approach through an API. And so when you look at things like payment reminders as a great example, you can use that channel as would you like us to have to call you if you miss a payment or would you prefer to send it through WhatsApp? If they tick that, it actually sends the verification through the WhatsApp API that now allows you to send that message because it's been consumer-led. And I think it's a really interesting way to think about some of those structures of like, how do you, throughout different parts of the process, yet that engagement will allow to come through. And I think that framework, like you said, I think your comment around which ideally should sit with the carriers, I think that totally makes sense up until the point that the phone calls or the text message is no longer run through the carriers themselves. And that's where, you know, when we start to think of a world where everyone has ubiquitous internet delivered by a styling, no matter where you are, you don't need the phone carrier anymore. And it's just a data sort of plan. And then you run that over a WhatsApp or an I-message or a Facebook messenger, et cetera. Now you sort of got to democratize that out a little bit differently. It's a different way that it involves. >> So if I can add to that, Josh, there's a couple of companies in Canada right now that, you know, the spamming of calls is significant, right? So there's a ton of calls that we never get to go through. And there's, you know, and you're trying different, you're changing your DIDs and trying to circumvent. Because, remember, that's all driven by consumer traffic right on what's considered spam or not. But there's a couple of companies now that are being creative where, you know, when the calls come out, they can put their branding in and as the calls coming out, as persons receive the call, they can say, hey, opportunity to discuss settlement, anything. So it prepares the individual that it's not going to be a confrontational call, because let's not get ourselves. They all know why we're calling, right? So the end of the day, if you're never, unless we're downing a number that is a spoof number, for the most part, you know, they're traditionally familiar with the collection cycle. But, you know, you're dealing with individuals that are in collections for reasons that, you know, the health situation, loss of work. These are unexpected circumstances in someone's life. And they don't know how to deal with confrontation or whether it's embarrassment. So at least this way, you know, you're able to communicate with these individuals before you get the call. And it may allow a certain percentage for those people to pick up that call. - Totally great. The point around that messaging piece, like you said, like what iOS 26 is going to use as what iOS 25 already did for voicemail screening, right? Which is amazing, because now someone's leaves a voicemail message you can watch and you're like, actually, I do want to grab that and you can grab it while they're leaving the voicemail. I think to your point, Tony, you'll be able to relay these messages into the screening component. And that actually presents a whole new opportunity where when you look at the phone, you can actually see, as you said, the sort of the message you want to get across without even having to get to some part of that process. And I think that's going to potentially play a big role in there, which so I'm going to kick across to you to think this is an interesting sort of segue. One of the questions I had for you was around sort of AI enabling the richer, more natural conversations across that, because one of the things that I've always felt on the digital side is, whilst you get all the mass distribution and you obviously get a lot of data that you can leverage behind it, you also lose sometimes that human touch of the good negotiation piece, the building, the compassion and the rapport. That's where I see a big opportunity, particularly with the LLM component in AI. We'd love for you to talk about that. And feel free to go in to briefly some of the stuff that obviously you've been doing on the AI slided in debt as well. Yeah, absolutely. So I also want to just touch on the previous point around iOS 26 as well. And I think there's a score of thought where for the consumers who have intention to pay it, you know, have the financial capability to do so, this approach almost, you know, with regards to having that message that Tony used spoke about, and also having an AI voice that's natural sound. It almost kind of warms up the person to engagement. And I think where in the early stages of this process of implementation across the industry, but it would be 360 C in the years time or six months time. How that has changed are consumers who have begun to engage with, you know, businesses such as ourselves with the digital channels because they felt that was more aligned with how they wanted to resolve their account compared to, you know, phone call with the human agent as well, right? So it's like the benefit of having, I'd of speak to like getting the information without having to speak to human as well. So there's that element. For the voice AI party in particular, I think over the last two, three years, we've really noticed that the voice models have become natural sounding because of estimates in the voice technology. So like the actual model that does the speaking, said the translation, if you like, of the text or audio sound. And that transition has meant that a time you could always be in a conversation and it will sound like you speak to a human as well. And in terms of like, you know, how that will pan out, I think there will be people who will be caught off guard by that, there will be individuals who will prefer that. And it gives us opportunity to be able to utilize that tool as part of the on the channel solution to be able to engage customers in a different way. And I think it's very exciting. We've done a fair bit of that research work in-house ad in-dedit as well. We partner with Editor Bios to develop the first e-house build version of this product, which is about to go into testing in Australia actually. So we're excited about that. And I think over time, what we'll probably see is that there'll be a bunch of models that I'll be specifically set up and tuned and customized for collections and even different types of work that is being done by the organization as well. And I think that's something that will become on prominence over the coming years. I will be curious to just understand across the panel as well. How we see that particular tool being utilized as a mix of human voice as well. Because I think with AI agents and their voice, they're very good at certain things. But I think at least from the initial testing right now, as the conversation kind of goes deeper and things become a little bit more complex, there's a tipping point where the human will probably do a better job, at least from our point of view, in making sure that the account is resolved. Yeah, be interested in here if anyone else has any thoughts on that, mix between the human and the AI voice and how it can transition between the two as well. Yeah, I don't mind taking a shot at this and I'll say this. So we have, as I said, three divisions of our company. One of our divisions is a purchase division, right? So we buy a council's trust, that's. And we had as much as 150 to 200 people offshore in our India operations. And the challenge there, sometimes linguistics is a challenge. And we moved it to, in 2024, to a champion challenge with a digital agent, transfer agent, just right party verification. So simplified stuff, no complex discussions. And over a period of 60, 90 to 120 days, we went from 100% offshore to 0% offshore to 100% digital for right party verification. We saw our right party contacts move up 3x to 4x to the point that we weren't able to service accounts 'cause we were getting flooded with inbound calls. And that's when it transitioned to, we need to start moving this into handling complex accounts. But I can see there's a balance. But on the flip side, like you said, on the collection agency side, my collection agents are dear to my heart. At the end of the day, they deal with all the complex situations. And I think David mentioned as well, the purpose of AI, my opinion. I know the people in the AI world are going to hate me. But the idea is small balance, low hanging through, not highly complex, you're dealing with sub-recation accounts or litigation accounts. And it's not like it can't do it, it can't do it. I've stretched these conversations to the max. But the difference is, you know, perfect blend is to have AI with real people. We're at the end of the day, these calls, if the AI is built properly, it's not a vacuum. But there's an exit strategy where a consumer wants to escalate to a manager. It can happen real time. This is what, in my opinion, builds a perfect hybrid. You know, you're not gas powered, you're not electric, you're pure hybrid. And I think it really builds a unique model where, 'cause it's all about customer engagement, right? If you don't get people on the phone, you're not going to get a result. You're not going to make a sale, you're not going to make a collection, you're not going to do any of those things. So, I think giving the customers the opportunity to have the best of both worlds, or you're not waiting in a call queue. And you're able to address things right then and there is the right answer. - Yeah, I think it's a good point, Tony. And David, I became for your view on this, 'cause sort of tying a little bit deeper into the actual sort of execution part of this. And so, Tony, building on from your point, I think if I just zoom out of the collections component of AI for a second and just think holistically, I'm in the camp of verticalized specific models are probably where we're going to go for different applications. Like there's not going to be chat GPT is not going to be your accountant and your doctor and all of those things, but there may be a foundational component to a very specific model that is built for GPs, for example, from a medical perspective or a CPA from accounting perspective, et cetera. And I think that's been starting to be validated when you're seeing different models for different things. We saw a anthropic yesterday, just raised a small amount of money of $13 billion at an observed valuation based on like the usage of coding, in particular, right, in that product. And when you look at the numbers of growth in that business, it's insane. It's gone from a billion of revenue to five billion in eight months. But people are using it and it's working. I know Mike, you and I know this internally on the engineering side we're seeing it happen in real time. The reason I bring up this is if that is true, there's going to have to be a collections or something specific in that area, which is where you then have this tradeoff of internal built versus external. And you obviously can get some part of the rails there. But Davis came for your view because we've seen this happen on the digital side as well around what stuff can be easily ubiquitous, taken off the shelf and provide some part. And where is there the proprietary nature of the stack? And to me, a lot of it sits around data. Your business is a great example of this when you connect those two data pieces that becomes so key. But yeah, we love your view on where you see this going from the ability for any of us just to buy this from Salesforce versus the need for some of this to be built internally. - Yeah, so I think Tony, go back to what you said in terms of the human capital side of this, right? So one question in the back of every operator's minds that a lot of people were hesitant to say how loud it is, there's going to be many, many, many rolls that are just gone. There's no way around it. But it's those rolls gone. That doesn't mean that people are gone, means those rolls are gone. And in a world where a bunch of rolls are now gone because humans aren't required for those things, that could be a pessimistic way to look at things if there's no additional work to be done. But the reality is there is so much work to be done by intuning these different models. And Josh, to what you're saying, you can't just take a open source model or one of the models from one of the closed wall providers and immediately deploy that. You've got to do so much training, so much additional testing. Now, the core models themselves are really, really good at natural language. So I think they're actually well-attuned in that regard to being used in the collection space, but they need very specific guard rails, very specific tooling. And the collection's workforce, the operators in the space are really, really well equipped to go through and be a part of that. And frankly, that's more interesting work than being on the phones, talking to consumers all day. Some people who love that, you know, God bless them. That's not the other systems for us in it. And I think there's a lot of other systems people out there in the collections industry. We've met, if there's an old adage, it's better to be a warrior in a garden versus a gardener in a war. So as we go through and think about, hey, how can the collections industry get ready for this? Quicker people can just get over the hump of, hey, there's a lot of jobs or a lot of roles that are going to be drastically changing. Let's have a organizational transition plan out of that to say, okay, in this future world, what is all the tooling we need to build? And where do people's personalities match well with that? Because to build these jobs, you don't need to be technically, frankly. You definitely need the mics here, you know, who are doing all the infrastructure set up and all those different weaknesses. But the training is such a big component of its, you know, the easiest analogy is if you use Sora from up in AI, right? You ask it, hey, I want a picture of a cat sitting on a car and it gives you four pictures. It doesn't give you four pictures out of generosity. It gives you four pictures because when you say, here's the best cat sitting on a car picture, it's going to then feed that back in. The collections use case is identical. You know, take four calls, have a personal listen to it, which are those four calls for the best. Great, that's all the training the AI model needs. But it needs that expert opinion and input from the collections operators to go through there. So Josh, I think that the foundational model, A, it's really expensive there. But frankly, there's other people spending their money on that type of things. In my mind, there's so much value to be created, building the tooling that works really well with those models and the training that goes along in the fine tuning, such that, such that, in my mind, that's where the verticalization really, really comes in, because frankly, the speed that the foundational models maybe is so, so quick. So you kind of just want to make sure that you're not slowing them down and just letting them do their thing off to the side. Yeah. Daniel, I want to bring up a point. I think you made such a great point there on, on sort of the role piece. And I'll reference the recent podcast I was listening to, I think I have a really good example, which is the sales force has automated away about 6,000 full-time customer service roles since implementing, they're basically the agent force working internally before they've made it external. But they've replaced that, putting 6,000 roles into, or the equivalent of that salary bands, into their basically SDR team on the sales side. And the comment from Benny off the CEO of Salesforce was since the time they founded the company until today, there's been 100 million inbound leads to Salesforce, which they have not responded to, because they haven't had the time. Like, you think about being in a good position, getting 100 million inbound leads that you don't have the time to get back to, but what he explained was like, "Look, if you're a salesperson, you're gonna pick between getting the Bank of America lead or, you know, Josh's plumbing companies lead, which one do you go for, right? You go for the biggest accounts for the biggest opportunity." And so, but AI is allowing that to be completely changed now. And so pairing the free up basically resources with the ability to shift that across into another part of the org. And then I think to your point, Dave, it's like just so much other work to be done. Like, Antonio, I'm sure you see this in your firm as well. Like, you don't get back to every mispayment. You don't get back to every opportunity to upsell into a better settlement campaign. And so it's like, and this is the question I posted to the audience here in the chat, is like, if you could automate away 80%, the pessimistic viewpoint, Dave, as you said is, I'm just gonna free that up and I'm gonna make 70% of the done margins, not 40%. The practicality is just gonna repush that P&L somewhere else because there's so much more revenue upside opportunity in what we're seeing. And I think we're seeing this, for example, with engineering, right? Like, no one's getting rid of their engineers. They're just doing more coding. So it's not like there's change in that way. But, Tony, I love your view on where you see that. And maybe there's a question like to the audience, like, what would you do? - I think the world's got this impression that this is like Armageddon here with the way AI is going to be. AI is about making things more efficient and redistributing the wealth, you know, regardless of what country you're in or content and finding good people today is virtually difficult. So you got everybody in each company multitasking, doing two, three, two, three roles, burning themselves out, just trying to keep up with the Jones. And, you know, for us, I'll give you an example. We talk about AI in the voice. We talk about on the channel with SMS and email. But, you know, AI can be, and we use this on a daily basis with our QA, you know, a typical QA person will do somewhere around 400, you know, QAs a year at a certain value, right? And that's whether that's because the company wants to be proactive or you're got contract mandate by clients. Well, that 400 can be done in three minutes, right? So that's the difference. It's all about efficiency and then redeploying the staff to other areas where you're multitasking and maybe gaps and whether it's business development or retention, whatever the case may be. So, you know, and again, you know, it's all about educating people, you know, where there's an opportunity to do a hybrid, it's great. And it goes back, I think, with Dave said, you need to have somebody who checks the checker. Like, you know, the thing that's concerned with AI is about is hallucinations, right? Whether the system goes rogue. I did. I've never seen anything go rogue, right? But hallucinations, if the system and coding, like you said, Josh, you're investing in that, it's getting better every hour, every minute. So that's just my two cents that, you know, you still have to have big brother, check the checker. But the idea is, you know, if you use it wisely and you've got good coding behind it, you know, you're, you're, you're far ahead of the game. >> Yeah, and another thing, like for this was something like, you know, I think that the, those of us who nerd out on this stuff all day long, the thing everyone's focused on is the rate of change. And it's just changing so darn fast. So when I think three, six, 12 months out, the things that are going to be possible, it, it, it feels like you don't need to like, you know, look too far in the future for those things to be real. The danger is that then, you know, leads into us talking about those things as if they're here today. So, you know, when you say there's so much work to do, like weren't like one percent integration of these things. There is huge, huge amounts of blocking and tackling on all the edge cases and all the, the different user journey handling. And so into your point, like you guys started in 2023, it sounds like, and, you know, that's not that long ago. So, so for other organizations, I don't think anyone should be intimidated by, hey, like how do I get started? Like the sense intimidating. I think the technology is actually super approachable and super easy to understand. Once you start getting in there and working with a little bit and the cost to get started is so, so low. All you need is a monthly subscription and to start embedding it in your daily workflows. So, so I don't want to, you know, give the impression that, hey, this is like so advanced stuff. It is changing so, so quickly, but, but it's, it's very easy to get involved and kind of wait at the water. So, how do we say this? Just dip yourself into the pool. You don't have to go into the deep end. Whether you start off with SMS or email and just voice verification, right party verification. There's so many ways for you to get, you know, climatized to it before you start putting a larger investment in it. The moment you start seeing the value to it. And again, it's not about replacing staff. My intention is never to do that. I, you know, I'd rather have staff and AI work together. But when you don't have the luxury of having extra staff, you have that, you have finger and I use this as an example. When I walked into a client, I made him listen to an AI call. So, would you think of that? That was awesome. I said, how long do you think it would take for me to get an agent from data higher to get to that quality? There are at least four or five months. I said, well, I can have 500 of those within 30 seconds, right? And that's the benefit, right? Now, it doesn't happen overnight. There's a lot of teaching coding that goes into it. But the reality is it's available to you at your fingertips. Get climatized to it and then you'll see where you can go. You know, listen, we talk about collections. You know, radiologists are an area of concern because like I think Josh, you mentioned, you know, I got a friend that just lives on chat GBD, doesn't go to a doctor. He doesn't go to a mechanic. He just lives on chat GBD. And the reality is more than 50% of the time, chat GBT gets it right to a certain degree. So, you know, that's the reality where things are going. And I agree with Dave, you know, things are improving every single day. And it's insane. It really is surreal. Yeah, I could definitely be that person as well, Dave. Like I get my blood work done quite frequently. You put your blood work into chat GTT. It's absolutely amazing. And it's like you don't get that level of quality from even a standard GP. It's pretty remarkable. I think it's, and Tony, I was going to make a similar point to you, which is, I think, and Dave touched on it. The rate of change is so fast that this idea of implementing some like org-wide, you know, bet, you know, we're going to bet the whole house on this opportunity and go all in. By the time you get anywhere deep into the implementation, the landscape's changed again. And so I think your opinion, Tony, on like incrementally dip in your toe and finding where there's value and experimenting is definitely the right path because it does, does just move so fast. Wanted to use that as a jump-off point to, and moving around a little bit on the agenda here, but I think you brought it up well on the guard rails, Tony, on your last point around. You can hallucinate. We can have these situations we need to validate for. There's multiple dimensions. You've got the technology piece of like hallucinations and the code and the data integrity and all that sort of stuff. You've then got regulatory, which is different as we all talk about different countries and the markets we all operate in. Whoever wants to go first, Tony, Dave, how do we see that in, I guess, evolving? I think again, the regulatory landscape is quite different in the U.S. and I think Tony, one of the benefits of the Canadian business we see in ours as well is a lot more easier to experiment and try things. Certainly we see that in places like Rotam and the Middle East as well. And Australia, to be honest, for as well as Mike mentioned, our first testing ground. But yeah, open, open floor and end for the audience as well as how you see this evolving. Let me take a quick stab. So my opinion, I think regulators are still playing catch up. They're trying to, they're trying to figure out jurisdiction and oversight because technology is moving at lightning speed. You know, you've got the registraros here in Canada and then you got the CRTC. They don't play nice in the sandbox together, right? CBV, we didn't want to wait till this rulebook. It's written. We, we build our own AI governance right to share. We talk about, you know, the things, the buzzwords, fairness, transparency, accountability. So this way we keep our system, you know, we try and keep it ethical, safe, compliant. I think those are the things, right, that from our perspective, because we do want to establish, like what I think David mentioned, those guardrails to provide unintended outcomes. We audit, we audit frequently. So again, I think that's what goes back to big brother checking the checker, right? And you know, what we set and reset, you know, remediation processes for unexpected behavior, right? So we have the verbages that we throw up where in our QAs, if we see something offside, we're able to address it right then and there. But we still see the regulators playing catch up because they're at least in Canada, the two don't play nice together. One thinks you'll supersede each other and this goes back to the same idea when collections was going into when SMS and emails came into play. Before it was written rules, people were diving in back in 2010. And then, you know, we had we had Castle created, right? Which is, you know, that overseas of the language of SMSes and emails to make sure it's preventable. Apparently it's, you know, it's one of the most honest written documents in the world. They say, but again, I don't think we're, I don't think we are where the regulators are where they need to be. I think they're concerned about dipping themselves into the pool themselves on overstepping because the reality is this technology is making it much easier for an outreach. And outreach isn't necessarily collections. A reach can be for several things. And it adds value, especially when you can't find staff to fill those voids. And I can only speak to the US market because that's the one that we operate in. In my mind, there's, there's a couple of layers that that oversight is, I wouldn't say broken in, but, but significantly hampered. The legislative cycle time in the US is so slow that people are still operating off of legislation from the 1970s. And the regulators have been trying to then, you know, interpret, okay, what's that mean for the technology of today? And a lot of those regulators are really well-intentioned, but now a lot of those regulators in the US are gone. So, so as we think through it, it is, in our minds, the technologists that are really kind of the last line of defense looking out for consumers, best interest. And there's always the, the veil threat of if the technologists and if their operators don't do a good job, then the regulators come back in and slap somebody's wrist for not doing the right thing. I don't really worry about that because I've met so many well-intentioned operators in the space who care so much about positive consumer experiences, getting the right information to the right person with the right guidance and the right context for that individual and making sure that everybody's treated fairly. So I really do think it falls in the operators. And again, that comes back to the speed of all this stuff changing. You cannot rely on case law from 10 or 20 years ago, or legislation from 10 or 20 years ago because, you know, it was people's wildest dreams this type of stuff would be true. So, so you have to go through and say, "Hey, not what's the letter of the law, but what's the intent of the law?" And let's carry that forward and ensure that the protections were put again in place with the different models and the tooling and the controls in the context is going to be something that someday all of us will be proud to tell it rain kids about, "Hey, I was responsible for pushing the industry forward in this regard. After those cranks you're going to care." They were like, "Yeah." That was some old fashioned industry where it's like, "Okay, you know, now everything runs in real time and stuff like that, but, you know, we can at least be proud of it." Yeah. Dave, I was going to add to your point there as well, just around the regulatory framework, probably not being ready, and that's really around, I think, because models haven't really been used for these, how to front my use cases in the same format as we have been in the past. And the example that I refer to is it took a while for self-driving cars to be approved for use on public roads, right? But then the technology has been available for quite a long time, at least, you know, in like, I'm close environment testing. And a lot of industries, including ours, you know, are playing catch up on that front. I think fundamentally, like, the model is always going to be at some point incorrect because that's what it does. It's like a probabilistic model and there'll be 100% chance that'll be incorrect at least once. And it's, I think, it's about demonstrating that error rate and allowing us as operators in the industry to be able to demonstrate that to the different bodies and having that be the new yardstick humans and, you know, with QA processes, humans make errors as well. And I don't think that, like, connection between the models being that kind of human, like, agent is also going to make errors has kind of transpired to the regulations yet. And it'll be good to see that I'm come through and being treated as like its own separate kind of entity instead of being honest, hallucinated, they're for it's bad. But it might have hallucinated once, but that one time it just kind of creates that impression that these things are good. So it's, yeah, I think we're definitely on the journey, but, you know, model work to be done there from our side as well. The other side, like, you know, we as humans like to believe that we're superpowers. In reality, we are all at the end of the day. We walk out of the house, we take in a bunch of information through our eyes or ears or nose, you know, or our hands, and we react to it. These all elements are just doing the exact same thing. They're taking in information and they're reacting to it. Now, the amount of information they're taking in right now through a little context window is tiny in comparison to the information that we take in with all of our senses. But imagine if we walked out of the house one morning and, like, tripped and then, so they said, go back in the house and I'll have to come back out for another 10 years. And just, you know, a lot of us, it's like that where we need to think about it as, okay, we're going to expose these elements to a lot of different information. When you think about, okay, well, there's a lot of situations that we as humans step into where we don't know the right answer. But we still take a head on and we have the intellectual capability to react to those things. We need to think about these models as we're helping them react to these things as well. Yeah. And I think that the only piece I'd add on is on your early point, is that, like, with the technologist sort of being the sort of final frontier, I also think that one of the issues on the regulatory side of things is just to immediately think about the downside case as opposed to think about the upside case, which is, you know, like, you flip it around the other way. Like, you can always use this example in various webinars, but like, um, many of us hadn't had a flat change and needed to call the American Airlines through, you know, Canada, like, want to see if you're in Australia and you get like, there's a flat disruption. So now there's a three hour wait and it's like, this is the most inconvenient thing. You can't imagine there's not a person who wouldn't be like, there's an AI agent, it means I'm getting my call answered in one second and you're going to resolve my problem that he was like, no, I'm just going to opt for the three hour wait option. I guess it's insane. Right? So you're providing that convenience for the consumer. This can be true on the other side as well, right? And we see that, so that with digital, like, the view was, there was going to, you know, overly harass the consumers and all these other things. And if you think about it, if you try and send a million emails a day, Google will solve that for the consumer, right? It's called spam filtering and the carriers are doing at their own side. The incentive structure is there. And so like you have to deliver the right message or the right volume and deliver it in the right way. But the convenience is the consumer doesn't have to harass with a phone call in the middle of the day and they can deal with it in their own time. And so I think there is plenty of examples that give positive inspirations of where this can go. Mike, I had a question for you as it leans into this and like, I think there's a distinction point in my mind between all the ML activity that's been happening in the industry over the last, you know, certainly nine years that we've been in the business, but, you know, with which channel do we use and what type of content we put in front of customers? And there's always been this question from clients and regulators on like, what is the black box having cited, you know, how the model's been informed? I think it's been easier to explain that on the ML side. It gets a little bit more opaque when you start to think about sort of the pure hello ML inside of things. Like the question I have for you is like, what does it really mean to have the AI be explainable in collections and like, whether it's to a client or a regulator, like, you know, when we have to ask you like, how is it working in a way that can be explained? Like, how do you think about that and what's your views on it? Yeah. I think I'll, I suppose take a step back and kind of go back to the ML piece that you mentioned as well. So if you kind of get to the very earlier days of them, or even before ML where it was used to be called statistical modeling like my first job after we finished university, yeah, we built statistical models. And you were able to actually work out like based on like a prediction of something, what was the variable or the feature that was causing that particular output to be as such? Because the models were much simpler, you know, people in the days where it was running, we were running linear models, they were maybe be like 30 or 10 or 15 or 50 columns, right? And it's a lot easier to actually analyze from a technical point of view and just to be able to investigate what's going on in the hood. But now we have neural nets of transformer architectures that have hundreds of billions of individual weights that actually, even if you just scale down to, like, let's say you had a neural net with a thousand nodes in different layers, you see what would be able to know what's going on because there's no kind of direct representation of the actual model weights to a real life understanding of a feature of like the data that's going in. Like one of the weights would actually just be something related to us and kind of wavelength in the audio file that's coming in. So the actual explainability piece, I think, comes really around to being able to make, I guess, the process and also individuals who are kind of trying to understand the process are comfortable about how this process is going to not loosen. So it's your point earlier, Tony, these models tend to loosen it, but also if you build them in a certain way, they will be less likely to. And from the kind of regulatory point of view, and also making this process become a little bit more portable and understandable, I think it's about making sure that the right guard rails up and also proving that the guard rails are not bridged because you can't actually go into the model and actually see what's causing output output to be run. I think the way to do it and to kind of at least also internally and externally in terms of how we're thinking about guard rails is to run a series of tests and making sure that these observations are within certain boundaries. But also the real reason people, I think, want the model to explain how the model behaves is because of concerns around how it will be output and results that are incorrect. So if we kind of tackle it from the angle, I think there's an opportunity to educate the wider industries and also just some users in general of how we're able to view these. So that, indeed, we have made observations. Like with the voice tool that we are developing internally, we've made observations that using just a GBT model off the shelf and just using prompt engineering, just putting in the system prompts and in the context windows, it's not as consistent because the model is so big and you're putting this tiny 100k context window into the model. It is going to result in, especially when the conversation gets longer, less predictable behavior. And I think that's the concern around what's actually going to, if it's going to cause issues. And in our industry, other industries, I think there's optionity to your point, devolue around using both human feedback reinforcement learning as well as the supervised fine tuning to be able to kind of make that a lot more predictable, smaller models that are fine should also tend to be a little bit more predictable. So there's a bit of a journey to kind of make sure that what we're setting up right now is going to be widely adopted. But I think when they're at least eight or nine, it's really exciting to be kind of making sure that everything is going to be kind of working alongside what the kind of orientation is how it is supposed to be. Like you made a good point there about the measurement. And I think one of those pieces is actually consistent with how you build these different voice parts. So when you're building a voice part, you're not saying, hey, AI model girl, talk to the consumer and figure out whatever they need. You're saying, hey, here's like 20 different voice spots. Each with a single purpose in that conversation, one to authenticate, one to do the hardship discussion, one to do all these different places. And these voice spots like can like rotate in and out of conversations such that it feels like a contiguous conversation for the consumer, but it's actually a lot of different people they're talking to and like a group call. And when in the testing regard, then it's actually easier to say, hey, like, I don't need to test the end to end thing, I can actually isolate, okay, in this tiny little piece, there was an issue, but the other 19 pieces were good to go. And you can isolate and dig in there. So I think when you break things down into very bite sized pieces, it makes it easier for the operators to work on those pieces and fine to and it makes easier to test and prove that the validity of the function there. So yeah, I think that measurement pieces is critical. Yeah, definitely. So conscious of time, everyone, and there's actually been a few questions and comments that have gone through. So I might just do a quick recap on some of these and then I'll pass them across Dave Tony and you might yourself just to address some of these, but just quickly, there's a point raised by a net that I want to address. I think it's a really good one, which was around the significant opportunity to reduce manual work and after call work. I think this actually provides a really good example of both Tony where you mention in terms of different you're telling the water and then Dave in terms of not needing technical like really technical people to solve this we've done this internally. We used the technical teams in the business to be able to sort of work out what infrastructure and tooling we needed to have in and whether that's the n8n of the worlds or the other sort of tooling and infrastructure you can use, but this is one of the exact use cases when non technical members of our teams, our operations teams, et cetera, are able to stitch together off the shelf software that we're already using to replicate exactly what agents are doing after the calls, which is filling in, you know, dispute forms or client servicing forms and pacing them into other areas and sending them off. All of that sort of stuff can be stitched together today in a really, really easy way. And when you look at the amount of like that's non productive time not being on calls and not being on chats with customers, I think a next point is super valid and like I've seen that firsthand in our business and that's been led by, you know, our VP of ops et cetera and not within sort of the CTOs and the engineering part of the functions as well. So I think that's a really, really good point. A couple of the questions, maybe Tony, I'll flick the first one across to you, like the question was around sort of adoption acceptance and engagement across different age groups for these digital agents and towards like digital in general, but within your agency like, is there a particular hyper cohort of consumer that's engaging more on the AI side of things or is it pretty generic across all buckets of your customer group? You know, we're seeing better responses obviously with SMS and I always use this, right? You know, within with an SMS, your phone vibrate, your phone rate, you know, it's an impulse, right? Emails, I'll check my emails maybe once a day, twice a day from a personal perspective phone calls, you know, depending on whether I see it spam or what and whatnot, it determines whether I pick up the call. Response times are better with SMS, we try and engage and segment the portfolio based on younger generation gaps. So the younger generation will use SMS on the older age groups will use email and then we'll spray everything in between between voice and because I think you've got to really target and segment the portfolios recognizing which my daughter and my son or they're 21 and 25 full and range, they don't pick up the phone unless they know it's in their contact list. They don't pick up the phone. But SMS impulse, right there, pick it up the phone, email, not so frequently unless they pop up notifications, right? So, you know, we try and base it, we try and base it on the generation gaps, you know, you start using SMS with individuals that are in all the security, very difficult. Now in the flip side voice chat on the individuals that are in the generational groups of baby boomers beyond, we're starting to see an acceptance, you know, it goes back to the old bank scenario, right? When the banks came in and they moved from brick and mortar to telephone banking and telephone banking sucked. Let's be honest. But when they move it to online banking, you know, everyone started to adopt. Now you're not going to be able to, you know, control the folks that still like walking into a bank and dealing with a teller. But for the most part, people recognize if I have the ability to deal with things head on rather than drive 15 to 20 kilometers, a gas of $1.50 a liter or 40 a liter, I rather deal with, I can deal with it from an accommodating perspective. So, you know, I see people willing to adapt. But I think if you're going to do the right thing for your business is strategize it based on where you believe the right type of connects you're going to come from by which vehicle. And I would add to like that, Tony, and that, and I think you're part of the question sort of raised is obviously as you point out, Tony, there's age groups that just have an aversion to phones or much more digitally native, they originated digitally on the credit products et cetera. But I think one of the biggest things we've proven in the digital collection space is that it does go through all age groups. Like there are plenty of people who just would prefer to be able to serve something in their own time and their own convenience, albeit there's definitely portions of that. And I know the models that Mike's built and our side to select channel template et cetera driven by those categories is done at a consumer level. It's definitely not segmented on age et cetera. And so I think that digital adoption uptake, you know, we'll transcend all barriers. Again, still conscious of time, there was a not so much a question, but I thought an interesting point raised here from Simon around something he's seen in the UK with sort of basically additional data in this case affordability data being used to sort of leverage how we can be more sensible. Like we see that our UK business as well, like for those not like familiar, there's basically a requirement to do income and expenditure assessments, which is before you set up a payment plan, you need to capture that because the UK is one of the leaders in open banking. You can pull that data so easily and consumers are already adapted to doing so. So rather than just setting up a payment plan based on sort of a negotiation, you can feed the data, understand exactly the right payment plan for consumer feed that through. And this is again where you could imagine trying to fill out those income and expenditure forms manually. There's three or four hundred questions you have to do and agents used to have to do that in the UK. You can now pull that through open banking run that through AI and all of a sudden you get an outcome, which is like the day of the right payment plan might be $20 a week and for might it might be $35 a week. And I think that sort of stuff is where we can, whether it's in the US or Canada, we can look to see somehow does that sort of stuff feed in feed saving further. Last final minutes, I want to thank everybody obviously for joining. If there's any more questions, I want to come through the chat and get them through quickly. Dave, Tony, Mike, any final thoughts from either of you? Tony, do you want your first? You know, AI is surreal. It's moving in a direction that we never thought if you go back 20 years, I say this all the time. When the terminator becomes a documentary and AI becomes like that, then we are living in a completely different world. So I'm excited about it, you know, little nervous because you're walking into parts unknown, but of you adapt it properly and you use it in a progressive manner and a safe manner. It's going to, it's going to do wonders. Yeah, there's so much opportunity, but it is really intimidating. You know, I mean, anybody like dips their toe in the water immediately gets infected with like, you know, big dream syndrome and it's hard to go back once you see, once you see all the opportunities. So, you know, I wouldn't, I wouldn't sit on the sidelines if you are, definitely dive in. And there's a lot of other people wandering through the dark, kind of feeling at the edges of the map the same time to learn with. So it's, it's, it's well. I was just going to say in terms of, you know, practical steps, I think it's going to understand what the edge of what the AI is capable of. Yes. So, if there is a news cases in your, in your firm, your business that you're wondering, oh, I wonder if I will be good for that. I think it's good to just do a quick test that is going to result in outcome that's perfect. In a year's time, it will because the models are continuing getting better. So when you set up your infrastructure and data collection and different kind of workflows to utilize what we have today, and the labs will, you know, come up with better models about a year's time. They may be able to actually be, you know, 99% accurate instead of 90% accuracy. So just understanding what's possible and keeping that as part of your kind of, uh, thought of view is, um, is good to kind of, uh, I think it's good practice. Yeah. Definitely. Awesome. Boris Johnson, thank you. I really appreciate it. I think the audience did as well. And thank you for everyone who tuned in and for now asked questions. We'll get the recordings out for those who had to drop off and those who couldn't make it today. So yes, Sean, there will be a copy of the recording and we'll get that across to you. Thanks for doing this. Just on time. Appreciate it. Yeah. Thanks for everyone. Appreciate it. Take a time. Thank you. See you. Thanks for joining us for another episode of Better Dead. Don't forget to subscribe.