Build vs Buy AI in Collections | Mike Walsh (EXL) & Tim Collins (Pay Ready)

Adam Parks (00:01)
Hello everybody, Adam Parks here with a LinkedIn Live webinar. So glad to have you joining us today. I've got some great guests and a really interesting discussion. You know, as we were going through the TransUnion 2025 survey data, started talking with my friends here, Mr. Tim Collins, Mr. Mike Walsh about in-house versus outsourcing AI technology. And I think different types of companies are looking at it different ways.

But I thought it would be very interesting to bring these folks together today and have a discussion around really what are the criteria for making that decision? I think it's a very personal decision for an organization to make. However, I think that there's some key criteria that's important that we consider as we start to evaluate AI technology and considering just the rapid adoption rate that we have seen across companies of all types all sizes throughout the debt collection industry. think different types of organizations look at in-house development as potentially a competitive advantage. And I wanted to bring this conversation together today. So gentlemen, for anyone who has not seen you before on one of our webinars, podcasts, we're speaking at one of the many conferences that you all speak at. Tim, starting with you, could you tell everyone a little about yourself and how you got to the seat that you're in today?

Tim Collins (01:23)
Yeah, awesome. Thank you for having me here today, Adam. And it's always great to be speaking with Mike. I've been in the industry for 33 years. I've kind of seen it all done a lot. And my current role is I'm Senior Vice President and Chief Risk and Compliance Officer of Pay Ready. Pay Ready is in the multi-tenant space, so it's apartment rental.

They're a software solution. So it's also one of the first times I've not been at a collection agency or a creditor or collection law firm in my career, but still very much in the industry because we connect the clients with the agencies themselves. Very, very focused on the topic that we've got today, which is the use of AI. How do you make the decision whether it's billed, the buy, the hybrid? And I'm super excited to be talking with you guys about it. But before we get there, Mike,

Mike Walsh (02:11)
You want me to jump in here? All right, sir. So Mike Walsh, I'm at EXL. started like him. I'm a dinosaur in this industry. Started in 1996. Worked with Tim at one point. So I've been with traditional agencies, started getting into the tech about eight years ago when I jumped over to VoApps did digital collections and...

Tim Collins (02:12)
Over to you. Absolutely.

Adam Parks (02:13)
Yeah, please jump in, Mike.

Mike Walsh (02:37)
I'm really excited for this conversation because I have it pretty much almost every day in my life. So there are, I'm a little biased. I'll put that out front. Because EXL has some very cool AI built that you can use today. But I think it is worth talking about and just educating the industry on, you what to look for. If you're going to build it, let's be smart about it and know what you're getting into.

Adam Parks (03:02)
Fair. And so as we think about, you know, really what this all means, what do organizations underestimate the most when they think about building something internally? Cause LLMs are not an inexpensive thing to build. And we're going to either have to leverage some third party technology in order to execute on it. Like there's going to have to be some other LLMs that we're bringing in and we're kind of piecing together or cobbling our own solution. But like, what's that, that biggest thing that organizations are not understanding or not preparing for when they think about building technology or AI technology themselves.

Mike Walsh (03:38)
I'll go first, Tim, if you don't mind. I would say, I'll beat Tim there. So I would say it's, I say it's the maintenance, right? Like people look at it as I'm going to build this technology and I'm going to have this technology now in my, you know, it's going to be one of the tools I use to collect. I would say, Hey, if you're going to build customer facing AI, like either a virtual agent or two way SMS.

Adam Parks (03:39)
Yeah, let's just jump right into it guys. I mean, we all present all the time. I figured I'd just throw out questions and see where they go.

Tim Collins (03:40)
Please.

Adam Parks (03:50)
Okay.

Mike Walsh (04:06)
You're going to have to maintain, adjust that technology constantly. Plus I think there are roles you will need to hire or outsource, via contractor that you will need to maintain that technology, especially as the technology evolves as, laws change in States, cities, you name it. There's a lot more than just building it and let's go and now we don't have to pay for it anymore. I think there are costs to maintaining AI that a lot of people don't consider.

Tim Collins (04:38)
And I think Mike, you make some really good points. I would add on that what's really concerning to me, if you're going to build something, you're going to build it with today's tech, right? So you say, so I built it. I built it. I went back and I was looking at one of my old GPTs that I had built and it's, and it's built on 4.0, which is a great, great, great LLM. get me wrong, but we're on 5.2 now. And so, and it doesn't update automatic and when I did upgrade it to 5.2, the results were not what I was expecting, because it was built and I made the instructions for that GBT using 4.0. And then I tested it and all that other stuff. to your point, you're absolutely right. You're just not gonna be able to maintain it. It's not gonna be like a legacy system that you build and it's good for 20 plus years or whatever it's gonna be something that's gonna be changing so rapidly because the underlying LLM technology like Adam mentioned that's required for these is gonna change so rapidly. is, I mean, it's an AI arms race as it relates between Google, Quad, OpenAI, DeepSeek, and other players that are out there. You your Jen Sparks, Manus, all these guys. And so if you're gonna build it yourself, It's the maintenance of it. And then also the realization of whatever that initial, you know, foundation you built on that LNM is going to change and that's going to impact. And you're going to want it to change to be able to get it better. So if you built this to be competitive, you want it to be competitive and you're going to have that, those, that piece along with it. So you have to be able to balance that, that in there, like you mentioned.

Mike Walsh (06:21)
That's a really, that's a really good point to like, I always tell people, think of it like this. I've been with EXL for about two years and like four months. I'm on my second total virtual bot. it's already, the one is out. We don't even, I mean, people use it, but, and people like it and it produces, but the technology has already advanced beyond that two years ago.

Right? Like, and it's completely different. The new tech, the Ajanic AI virtual agents, if you have not heard them, it's amazingly human-like. And so think about what you're going to build. What is it, be honest with yourself, because everybody, there's a lot of terms that there's no standard of what people use to describe a virtual agent or a I would call, you know, what you see mostly in the market is

Tim Collins (06:49)
Right.

Mike Walsh (07:15)
what I would call intent-driven AI and now that agentic is here and starting the process of getting utilized by agencies, creditors. I think a lot of creditors have said, yeah, this is a step above of, even creditors who have money, have resources to invest have said, we're not a tech company. This is beyond us because it changes monthly, faster than.

Adam Parks (07:36)
That is a really interesting statement, Mike, because that's what I saw in the TransUnion survey data that, believe it or not, creditors are the least likely to try and develop something in-house and debt buyers were the most likely to try to develop something in-house. And I think the debt buyers look at it as a potential competitive advantage. And I think the creditors realize we're not a tech company, we are a creditor and we need to focus on lending money, which I think is smart. I think that seems to be the right path for those types of organizations. But as Tim talked about kind of the arms race, we start thinking about Gemini releases. So ChatGPT bumped up their release dates and like you see this back and forth for market share right now. And I think as we're in this AI arms race, it's just going to move faster and faster and faster, at least in the short term. So what what hill are you willing to die on?

Tim Collins (08:13)
Yes.

Adam Parks (08:28)
Like where are willing to stake your flag? What hill are you willing to die on? Because if you're going to build something in house, you're going to have to stake that flag and you're going to have to defend that position for an extended period of time. Cause you're not going to, it might be easier to move from, let's say a baseline of ChatGPT 4 to 5.2, but it's going to be very difficult to move from ChatGPT 4 to Gemini three to chat, you know, back to ChatGPT 5.2. What's that going to look like if you don't have the resources now?

Adam Parks (08:56)
My next question comes into the resources themselves. I know we read the headlines, know, Meta is spending hundreds of millions of dollars on a specific developer to be able to bring them in from an AI perspective. So can these smaller companies even afford to be competitive in attracting and retaining that type of AI talent in compared to a multi-billion dollar organization that is built for this?

Mike Walsh (09:23)
Ooh, that's, I mean, you need a lot. You'd have to allot a lot of money for this, Like, it's a lot of like, so we're a global company. That is a resource we constantly are trying to hire. And so we're hiring around the world for that, that talent, not just here. In the US I think it's even more competitive for that talent. So.

Adam Parks (09:28)
lot of money. It's a lot of money, right?

Tim Collins (09:29)
Right.

Mike Walsh (09:49)
You could, I mean, I think you can, you just have to commit to it. But again, I think your flag point is very key, Adam. Like, this is what you're gonna be marketing, this is what you're gonna be standing next to, this is what you're gonna be saying, this is our tech, and this is what we do on every RFP you fill out. is it gonna last? That's the number one. Concern I have when I hear it. And I think that's what the creditors set. Like the creditors saw the change. They were all building. When I got to EXL August of two years ago, that was my number one competitor was we're building internally. And that has flipped. Because what they built is the old, like Tim said, is what you're building is today's technology. Today's technology is outdated in three months. ⁓

Adam Parks (10:21)
there.

Tim Collins (10:34)
today.

Adam Parks (10:35)
Yesterday? Yeah, was updated yesterday.

Mike Walsh (10:37)
Yeah.

Tim Collins (10:39)
It's insane. I do think there's another point here though, is that, know, Mike, if you can get the talent, that's great. You gotta be able hold onto them, which is going to be even harder because everybody wants that space. it's, know, being a global player helps because you can go into other markets to able to find that talent. I do think also that there's this misconception today because, you know, some people are using ChatGPT. So they're like, hey, look at, am, and it's easy for me to use and I'm a really good prompter.

I'm just going to go build it myself, right? Because they can't get that talent. And so, and there's some tools, there's some, you know, there's tools that are out there, you know, people see them. It's like the either vibe coding or you could use a, you know, a GPT with ClaudeX and it could code for you. And then nobody knows what to do with the code, but they've coded. You've got all these other tools that are out there, but you could use something like a Lovable or a. I'm playing around with Gumloop right now and it's, you know, it has the capability to build. Would I turn that on? No, no, because I don't have all the backend stuff. I don't have the security stuff, but I am playing with it to see what it could possibly do to help me when I'm, you know, reaching out to like a company like yours, Mike, to be able to say, okay, this is what I wanted to do, you know, and have these capabilities. So at least I have an understanding of what these, what these tools can do. But, um,

Mike Walsh (11:45)
you

Tim Collins (12:06)
that, you know, I can just go build it myself. I'm just going to go. And I love it people tell me, I'm going to build my own voice AI. It's my favorite one right now. It always makes me chuckle. I'm just like, yeah, go for it. mean, I mean, me know how that works.

Mike Walsh (12:14)
Bye.

Adam Parks (12:17)
Half of the companies that told me at ACA in July that they were building an AI voice bot are now doing something very different. but I mean, even by the, we went from ACA by the time I met with some of the same people at the RMAI Summit or at DCS, just a couple of months later, they've pivoted their whole organization. But that's again, why I'm not betting on some of these smaller companies and you know, look.

Tim Collins (12:21)
Yeah.

Adam Parks (12:41)
If you were a local restaurant and you were selling acai on the street or something, and you want to build your own app to be able to like do some minor things, inventory management, whatever, like that's great. Go use a Lovable, like go use Claude build yourself a tool, but that's not a financial services level product that you're going to build with that tool set. It's just not realistic.

Tim Collins (12:49)
Go for it. Yeah. And as you get more more exposed to consumers, so you're right, a little vendor on the side build it up in Artifacts and Claude and the way you go and you've got, you know, an app, you know, within minutes. But if you're sending out content to consumers and they're responding to it, you know, in a highly regulated financial services, I mean, the risk there is massive because AI is that is, is, you know, at scale.

If it goes off the rails, which is one of the big arguments, it can go off the rails big time. It's like when we saw that, I think it was the Air Canada case, right? Where Air Canada had a bot and it was giving all these deals and stuff. And the courts came back and said, your bot did it. You got to own it. There's like $20 million. So it's those risks when you're your stuff.

Mike Walsh (13:42)
you

Adam Parks (13:48)
Yeah, you own the output.

Mike Walsh (13:51)
Yeah, and I think Tim, even taking that a step farther is like, OK, so you're going to answer the RFP to a creditor if you're an agency, let's say. the questions are, who built it?

What did you use? Like, do you own the LLM? Did you use ChatGPT? That's not what they want to read, right? Like they don't want that information. Like, so if I would say if you built it, I think you have to take some sort of open source and make it your own, right? Like, and, and, and hunker that down so it can not access the internet. So there's a lot of steps where even if you built a great using anything out there, right? Like, Is it secure? Like you said earlier, Tim, I don't, that's above my pay grade, but I would be leery. Is it going to pass the smell test for an RFP? If you're, if you're pulling all these different pieces together to make an AI product and you own like one, you know, one of the 10 pieces, is it your product anyway?

Tim Collins (14:53)
Yeah.

Adam Parks (14:55)
What does the product privacy assessment look like? What is the final privacy impact assessment going to look like for the organization if you are stacking other things on top of each other? What's that gonna look like?

Tim Collins (14:59)
Yeah.

Mike Walsh (15:08)
Yeah. What is the model evaluation, the prompt engineering training document? Who do you have to go get to fill that out for you? Like make sure, because you can build whatever you want, but it's got to pass a creditors or regulators. You know, think of those two terms, sniff tests to make sure that everything that you say is there is there. There are no biases built in. There are, you know, There's so many things, like think of a birth date. Are you taking that information? All of you have an ageism lawsuit, because half a percent of your settlements are higher on elderly people. So I just think there's a lot to consider. I would also say if you're going to build it, you need great tech people, but you need collection people who understand this domain, this industry, to be there with them every step of the way.

Mike Walsh (15:59)
because if they're not, if collection people are not helping to build this, you're building a customer service, AI, that may have issues when it gets specifically into this space.

Tim Collins (16:13)
Yeah, agreed hands down and I think.

Adam Parks (16:13)
So one of the things that we've talked about in the past, go ahead, Tim.

Tim Collins (16:18)
Now, was just going to say, you know, what we're going to see from a regulatory perspective that Mike was talking about and that level of transparency, you know, the, you know, we saw the executive order come out from the White House about states not creating, you know, roadblocks to AI. Let's just say that that's not effective or as effective as people think it's going to be. The states that have already come out with AI regulations, we're going to see this patchwork. and you're gonna have to make sure that it's compliant in Massachusetts, California, Colorado, Utah, Texas, know, just as kind of you do. We've been lucky with privacy. Privacy's kind of followed what we saw in California, which is a GDPR light, as I call it. What we saw come out of Europe, you know, you have the right to know where it's being collected, you have the right to delete those kinds of things. So, but I'm afraid with AI because it's moving so quickly, right? People don't clearly understand it. You have statutes like California that have AI defined in an overly broad fashion that it could include basic scripting, if you will, things that we've been doing since the beginning of programming, not necessarily generative. So you're going to have to have that level of expertise for anything that you're to be presenting. But I think that's coming. And I think that's going to add a significant burden. Yeah, you could build something today. Let's say you can get past all those other hurdles. But now you've got this regulatory compliance risk assessments, you missed or the ISO 42001 standard, all of those other kinds of things. Now you're going to have to have that stuff and you're going to have to probably have a bigger organization. But Adam, I apologize that I cut you off.

Adam Parks (17:48)
No, that's me because I'm on the other side of the world. Like broadcasting from Brazil has its challenges. I appreciate you guys. No, it is lovely weather. ⁓ So Tim, when you were talking about the, it is a little bit warmer here than probably where most of you are right now. I think it's like 80 degrees Fahrenheit here right now. And I have to say Fahrenheit because everybody gets confused.

Mike Walsh (17:52)
Thank

Tim Collins (17:53)
It's lovely weather though. Lovely weather.

Mike Walsh (17:58)
Probably a little warmer.

Adam Parks (18:09)
So, you one of the things that you were talking about was, you know, Trump comes out with this executive order. We're going to, you know, we're going to do preemption at a federal level, but I don't see any of that becoming a reality unless Congress steps in and does something when they haven't done anything about privacy. So I don't know what makes us think that they're finally going to get their act together and try and put together, whether it be a regulatory body or something that allows us. Maybe we'll get extra lucky and like, let's just hope for, you know, an early Christmas present here to where maybe we can wrap privacy and AI into one organization and have it not have an omnipotent director that sits over it, you know, that becomes a political punching bag, no matter who's in an office. But I don't think that's realistic. So I want to turn the conversation here for a minute to the data that feeds the tools. And so Mike, you and I have talked about this a couple of times, Tim, you and I have talked about this on AI hub is, you know, the AI and the tool sets that we're using are only as good as the data that we're going

Adam Parks (19:04)
And Mike, know one of the things that you've told me that you guys provide before is assistance and actually I don't need all of your data and I can help you organize and structure the data that I do need to see. Now, if I'm going to build something internally, now I have to build it and I have to structure all the data. Like how realistic is that? And is that amplifying the amount of resources necessary for me to build a tool in-house?

Mike Walsh (19:27)
I think it is and it isn't though, right here. Like I'll try to not be too biased, but I think, you know, I've seen so many different collection systems. We have a lot of right? We have, and we gain data. I think we don't utilize it effectively. Like we have all these interactions with customers and I'm not sure we're using that to its fullest, but It depends. I think the challenge is you're pulling data from different spots, your dialer, your email server, your CRM. How is it feeding in the CRM? Can you access what you need to create, let's say, a two-way AI which can communicate with the customer in real time? That may be a challenge. do see. So then is it, are you going to create a data lake, have all this information feed to that system? You got to make sure that system's secure. You got to firewall the heck out of it because that's going to have all the data. And then you need to be able to pull it as you need it, especially with a virtual agent. Like a virtual agent is like this. A text, you get, you know, you'll have a person can respond six hours later to it. after you send it, but a virtual agent, think it's a challenge. It's a challenge, but you have the data. so many prospects tell me, I don't know if I have the data. If you're collecting accounts and you have all the regulatory information to collect an account, you have enough data to do it. It's just, is it organized structured or.

Adam Parks (20:55)
But if you don't know you have the data, it worth actually trying to build something? If you don't even know that you have the data, is it irresponsible to be trying to build your own tool set?

Mike Walsh (21:05)
And maybe those are different voices in the organization saying there's no way, why are we going to do this? I hear that a lot. We think we can build it, but we'll never get it done. How long is it going to take? I hear that all the time. I'm sure, Tim, you've had these conversations where we should build it, and then the guy next to him goes, we'll never get it done. Who's going to make this their full-time job to be in charge of it?

Adam Parks (21:12)
It could be.

Tim Collins (21:13)
Right. Yeah.

Mike Walsh (21:30)
Who's going to report to them? We have to hire people. Like, where's the expertise coming from? It's a lot. Now, if you commit to it and this is the path you've chosen and you have the money, you know, I know.

Tim Collins (21:32)
Exactly.

Mike Walsh (21:46)
A guy who's using tech, he's built some of it, he's pieced it together and they're very effective. We're good friends, we talk about it all the time, he calls me for hints, things like that. So it can be done, but it's just a long road. timing is something to consider with that.

Tim Collins (22:08)
Yeah, and I think that's kind of what we've talked about before, right? And I think back, Adam, to your question, if you don't think you have the data and you're gonna try to go build something you don't think you have the data, then I put you in the batshit category, crazy category, because you don't have the fundamental blocks to build what you need, right? So that's a legal term. You can look that up. But in all seriousness, that should be your...

Adam Parks (22:24)
That's it.

Mike Walsh (22:27)
Thank I'm dealing with that after.

Tim Collins (22:33)
That should be your first clue. We need to go get help. So if you're sitting around the table and a bunch of people are going, I don't even think we have that data. You're like, okay, thank you for playing. We're going to move on. But Mike's point is the one we've kind of talked about already. So if you have the resources, great, that puts you in a great spot. If you can get the expertise using those resources, that's great. You've checked off that box, you know, as we talk about that checklist. Do you have the commitment? to do this, right, which is we're not just gonna make this part of Mike's job. We've gotta get somebody in here who's going to be doing this full time forever for as long as we do it because it's gonna continue to change. And so I think it's easy for us to check some of those other boxes. Like I don't have the resources, perfect, move on. I don't have the, I can't get the expertise. Okay, move on. the commitment one's a little bit harder because that can change if the CEO changes or ownership changes, that can change very, very quickly. And most people don't realize what that level of commitment really means. It means like dedicated, focused, willing to continue to spend, you know, when, you know, somebody might come in and say, look at, need to get ready to sell. So we're going to stop all this other development and stop everything else. It's that lack of commitment.

I think in our space, I think in any industry really, to be honest with you, it's pretty easy to start. It's the execution and implementation to get it across the finish line, which is where we all struggle. know, whether it's January 1st, we've got our New Year's resolutions, know, great starters. The gym is packed that first week. Week two, it's a little thinner. By week three, it's like back to normal. You know, the people that are committed, yeah, I know, yeah, who are committed.

Mike Walsh (24:15)
Yep, thank God.

Tim Collins (24:19)
who are gonna continue that piece. And that's just part of human nature. So you have to really dive deep into that element. And assuming you can get past those other ones, have the resources, expertise, that commitment is gonna be crucial. And it's like looking at the team and saying, hey, you committed to do this. I'd almost have them sign something and have it notarized and hung up on a wall somewhere so everybody could look at it say, no, you said you were committed. We're gonna have to do this. So, no, right, yeah.

Mike Walsh (24:44)
Now, I think, Adam, real quick,

Adam Parks (24:44)
It's not a terrible idea.

Mike Walsh (24:48)
based on those numbers you were reading, like how debt buyers are higher up on this, and I think a lot of what they're building is models, right? Like now, if you wanted to build models to evaluate.

Mike Walsh (25:00)
portfolios and I mean, probably the expertise, no one's more capable with that knowledge than, cause they've all been using models, right? They've just had humans running it. So now you add, or you bring in a group to build you some machine learning and algorithms and AI to make these models better. That's where we started. were, EXL got into the AI business for collections through modeling for creditors and debt buyers. So I think that's something that the industry can do. I think that's more achievable. Like what do we. We always say on our demos is, you know, let's start with the outcomes. What, what are you looking for? What do you want to accomplish? Because if you want to do everything in AI, it's very hard. You're going to need, right. But if you want to do something like that and you're a debt buyer and you want to evaluate even, even a collection agency if you want to create your own internal scoring using AI, I think that's achievable on your own, you know, with, with a minimum

Tim Collins (25:52)
Right. Right.

Mike Walsh (26:08)
commitment. You still gotta have someone who knows, you know, that damn it.

Adam Parks (26:12)
I think it's the use

Tim Collins (26:13)
Yeah. Yeah.

Adam Parks (26:13)
case that's most likely for us to be able to develop internally, right? Like the debt buyers are probably the most capable for it, but let's go back and look historically about how AI entered our marketplace to begin with. The first companies that came out were scoring companies basically, right? Atunely and other organizations that aren't in the industry any longer, right? But that's where it kind of all started was who is going to be able to build some tools that help us to prioritize accounts, score accounts, and model potential liquidation before purchase. Mike, I definitely hear the idea of the debt buyers building out their models and being able to better understand the investments in accounts. The scoring and segmentation, those kinds of things. Do you think that's the best use case in which debt collection companies can build something internally. We spent a lot of time about why they shouldn't, but once you get that data into Snowflake and we start thinking about where is it most realistic guys for the industry to actually build something on their own.

Mike Walsh (27:14)
I think there, and I think there are a lot of administrative processes that are probably, you know, I was listening to the. NVIDIA CEO on Joe Rogan. was really good podcast. But he said if your job is the process, then your job may be in question in the next five years. so if you're filing or just checking, I think there's an administrative behind the scenes. It's not necessarily scoring, but just handling so much, you know, busy work. those type of projects where it may not truly be AI, it could be machine learning. I look at it as robotic processing with a little intelligence that says, a human should look at this. A human should look at that. because this letter is not just a cease and desist, but there's a threat in here of a lawsuit. Stuff like that. I think there's opportunity there. Those processes that are highly manual, That's where I would look if I wanted to build something to make, to give me an ROI quickly that is achievable to test the effort to build something like a complicated scoring model or something like that. That is going to be, you know, one bad equation on a scoring model. It's, it's, it's also tough, but I would say manual processes are behind the scenes of the, are the place to start if you're to build internal.

Tim Collins (28:36)
Yeah, and I agree with Mike. I think there's a whole host of things that you could do that, you know, they're not consumer facing, they're internal, you you still have to get past the security piece. So don't be using the free version of ChatGPT to clean up a data set for you. So you still have to get past that some of some of those things. I do think, Adam, to come back to your point, especially as relates to the the debt buyers, if you will. Yeah, they're very comfortable in building the models.

Mike Walsh (29:00)
Yes.

Tim Collins (29:15)
You, just have to be careful that model building is not, you know, the same as using some of the advanced, you know, AI tools that are out there, the generative AI. A lot of models are built around machine learning, which is a form of AI. And I think that's where this really, this really helps is, you know, you can use machine learning. It's an algorithm. You can set it up and runs. You don't really have to worry about it changing much. Rarely do the algorithms change, but they do change from time to time. So it's fairly stable the generative AI stuff is what is changing dramatically. And I do think there's this, sometimes this fallacy to be able to say, you know, I built something using RPA, I should just be able to build something using AI and it should be easier, right? And it can be, don't get me wrong, but it can also open up all these other doors that we've necessarily talked about. I do think, you know, to Mike's point, There's probably a bunch of stuff in the back office that is a great place to start. And I wouldn't even start with AI. I would start with what is the process and why do we do it that way? Really, you you can love them or hate them. doesn't matter. But Elon Musk has been so successful because he always starts with the process, right? And he removes everything that doesn't, that does not add value and is not required. He just removes it. Right. And he sometimes removes stuff to a point where it breaks. And then he goes, okay, that's where it breaks. Now we have put this back on so it doesn't break. So he has the bare minimum. If you started just looking at your processes first, and then you could say, okay, now what can I automate? Maybe you're using a Power BI tool or an RPA or something along those lines, or even some basic scripting could get you by. And then you could say, I'm now I'm graduate to maybe, know, RPA or machine learning or whatever. And as you start to go and your level of expertise start to shrink, we should all be able to review our and figure out what we should be doing and not doing and asking those questions. Anybody can do that. Anybody who's watching this can do that. Using scripting, using RPA, using AI, that's different and it goes up, the level of complexity goes up and that's really where you need to bring that level of expertise that Mike's been talking about in. So I wouldn't, if you wanted to, you could take your current process, upload it to ChatGPT and say, what am I required to do here? What could I eliminate? They would actually streamline the process and it would give you a pretty good start. Now would you still want to be the human in the loop? Maybe want to verify that with compliance or outside counsel before you just go and implement. Yeah, we don't need to send a model validation notice. ChatGPT said it's fine. Just send them. Yeah, now that the CFPB is out of business, don't even need an NBN. Just go right forward, right? You're always going to want to be questioning stuff, but I think it's starting with those fundamentals. to Mike's point, there is great places where you could use AI and some of the tools that are out there.

Just understand as you add more more PI into it or you use more and more complex tools, those risks go up and they go up not linearly, they go up exponentially. So the exposure for using generative AI to respond to consumers when it hasn't been trained, hasn't been built correctly can really go off the rails very, very quickly.

Adam Parks (32:25)
I think you're 100 % right. And the scalability of the AI operations, I think is one of the scariest things that we have to consider. Because if it's good, like you said before, if it's going to go wrong, it's going to go wrong in grand style. There's going to be no, you know, small mistake because everything is going to be exponentially amplified. But when we look at how organizations are looking at it, you know, previously, and I was just actually going to pull some of the data so that we can talk about it in a little bit more of an interesting way because what really interests me from a debt collection perspective is how debt collectors are looking at it from a use case perspective. And I think there's use cases like we've talked about that apply across any kind of an organization, Your local restaurant to, you know, somebody selling pens, like it doesn't really matter. There's some use cases that I think are general and generic enough for everybody.

And then on the flip side, we've got six use cases that are laser focused on the debt collection industry itself. For me, when I look at it, I see scoring and treatment, data sourcing and appending, chat and written communication, voice communication, quality and compliance and negotiation. Those are the six that I've identified that are really specific. And I think if you show me any vendor in the space, I can fit them into one of those categories, at least their products pretty easily.

But what I think is really interesting here is how, when we look historically at the industry. The first people that came in were scoring and treatment. Then we started seeing this compliance use case start to come to the forefront where it was, we'll identify what they, like Mike, you were talking about, we'll identify what the, who needs, which calls need to be seen by a live person. And we start looking at it from that perspective.

Mike Walsh (34:00)
you

Adam Parks (34:11)
Now we seem to have jumped over a lot of these other opportunities here and we've gone straight to voice AI. And I don't know if that's because it's the sexy new, you know, thing that is now possible, but like we've jumped over so much of the low hanging fruit to go after the thing with the biggest risk. Now granted bigger risks generally means bigger reward, but do you think that we've kind of skipped a few steps here and now we're focusing so far down the path or you know, is that really where the value is?

Mike Walsh (34:41)
Well, I'm biased, so I'll just say that, right? Because I have a pretty good virtual agent. So I'm just going to put that out.

Tim Collins (34:46)
Fair. Fair.

Adam Parks (34:46)
I mean, have, but you have some bias, you also, but you built a lot of things to get there. You didn't just jump. You didn't go from here. Let me listen to some compliance calls to now I'm making outbound calls.

Mike Walsh (34:53)
Correct. I would, I think there's right, like just the marketplace pressures, right? Of hiring agents, maintaining agents, training agents, rehiring, the extra volume that is out there, you know, the high charge-offs, the pressure from creditors who want to make their budgets too and have so much debt that.

Mike Walsh (35:23)
I think there's a push to get, if I can save money on the phones and if this can perform like a human, near a human, close enough, I can get to more of my volume and I'll figure out a way to get, right? Like there's so many and you know, we just got a case study. Now it's not in the U.S. It's from Australia and the numbers were really good. And I was shocked, but I why people want it so bad is I think people are making decisions not to touch certain volume because it's either too small a balance, risk like it doesn't fit into the scoring models.

My problem with scoring models is I could be a great customer, have a great job, be paying my bills, and then I am unemployed and maybe I'm unemployed for four months now. it's now savings is gone, unemployment, I can just feed my kids. And now I'm in real trouble. What good is the score that was six months, pulled six months ago off my credit period? you know, so I think there's all these pressures that, I think that's why there's the push. Do I think it is voice is the best? I tell everybody to start with intelligent email slash two way SMS, right? I think that is the fastest, easiest. It's very secure. It is intent driven in AI. So you know what it's gonna respond. It would pass Tim's check. That's how I look at it. then you would, I think you'll see ROI. then voice, know, voice is, the technology is getting better every day. And, you know, I think let's say we did this this time next year, I'm gonna have a lot more. collection agencies, creditors on it. like this came out, we went live in June with the newest technology. And we did it at a big scale, but I don't have 50.

References, know, like I just you can't implement that many that fast, but I think So I think it's there I think people want it and they see those numbers that the math is amazing, like if you look at a price of our US agent right party contact, it's ten bucks At that you know pretty much near shores probably seven ish Offshore, let's say five maybe three at the best This is you know, less than half of you know, this is pennies. So The math is so good. I think it's just so exciting But do your research make sure That you know, they know collections are not you know, there's a lot of stuff started in marketing or payments now Hey collections is huge space. Let me bring it over Be careful. Be careful.

Tim Collins (37:56)
Right. Yeah. Yeah, absolutely. I do think, you know, to just add on to Mike's point, there is so much publicity around AI taking away jobs that that's where people go and look and say, okay, if I were to take away jobs, which would they be? You know, and it'd be like, well, if I could cut my payroll in half by adding AI to Mike's point, that makes my bottom line look so much better. But most of us are in competitive markets. You would have to still be competitive.

Right? So you may lose some of that competitiveness because there was a study that was done not too long ago that said it's easier to break a promise to a bot that it is a human. And so I think some of that exists. You know, I talked to Tim on Friday and I promised I was going to pay and I'm paying. So you have to factor in all of those other things, but I think people look for those big, big wins right away. And if you could just make some process, you know, three to 5 % more efficient and that frees up and then just do that again and again and again. Then year over year, you get better and better and you leave the voice AI stuff till it gets figured out and really processed right. think, I think that would be, you know, a great strategy for anybody. I will say this voice AI does not fix the RPC problem. So if, cause consumers don't look and go, Hey, it's a bot that's calling me. I'm going to take this bot call. be right back. No, they look and say it's a call. I don't recognize who it is. I'm not picking it up or

Adam Parks (39:13)
Bingo.

Tim Collins (39:23)
I'm on an Apple phone. I'm going to have my bot answer that, right? So I don't know that number. It looks like spam. Perfect. I'm just going to have the bot answer that for me. So it doesn't fix that problem. What's that?

Mike Walsh (39:31)
It's funny you said that, because it's funny you said that because in that case study, it was almost identical, right? Right? Right part of contact. It was like 3.6 or 3.7. The bot was actually a little...

Mike Walsh (39:47)
tick lower, but it made seven times as many calls, right? Like so, so you're right. Nobody, a call is an interruption. That's why I always tell people start with two way SMS. It gets delivered. They can respond whenever they want. They can do it at, you know, at 9 PM. Yeah. Looking at their bank. Yeah.

Tim Collins (39:51)
Right. Right. I make it up on volume. Yeah. Yeah, good call out. Or layer it in. Yeah, layer in your SMS to say. Yeah, layer into your SMS to say, do you want to call right now?

Adam Parks (40:13)
Isn't the RPC going to be driven by the data? Right? Like more so than the tool or the outbound communication tool itself, it's going to be driven by the data that you feed it. So if you're going to send seven times the volume of outbound calls on bad data, aren't you going to get pretty much the same result between a bot outbounding or whatever? I think there's other things that you could do to change it.

Tim Collins (40:37)
Yeah.

Mike Walsh (40:38)
Yeah. I mean, I think those, that's the tool that, you know, that I think is starting to come out is, Hey, we have data to predict right party contact, but me, Mike, yeah, I just.

Tim Collins (40:50)
Yeah, when? Day? Time?

Mike Walsh (40:54)
I think because of cell phone, I remember an agency I worked at, and we were also a BPO, we did a massive study on every account internally and we had some large accounts, it was huge volume. And we realized that what we thought was the hot times like lunch and dinner was not. And what we found was almost every time during the day of operating hours was equally valuable for right party content. So I don't believe. There is any time, I think now that people have a phone in their pocket, it's, do you come across a scam likely? Do they feel like answering it? Are they just bored in an airport? That's when I answer my possible call and mess with the spammers. You know, what, what it, I just don't, I don't know. I think it's random based on, okay, I gotta call him and answer. Or I'm a person who always answers, or I'm a person who never answered. Like, I don't know what those tools are gonna use. What a data set would it, is it, does your past predict the future? Maybe, but. who like, we used to call our friends on Friday after work and say, hey, how you doing? Where you going? Nobody does that anymore. They text them. So like, you used to order a pizza on your phone. Nobody does that anymore. You know, like, I just don't know how you, what data is out there to tell you you're gonna get more right. And I've had a couple, I've talked to a couple of companies who say they haven't and I'm excited. I'm gonna actually look at a couple in January, but they would have to show me something, you know.

Tim Collins (42:06)
Bright.

Mike Walsh (42:27)
just think it's random.

Tim Collins (42:28)
Yeah. But I think to your point though, Adam, I think if you said, hey, look at, am I going to spend money on voice AI or should I spend money on better data? I would say spend money on better data so that you could, you know, have the best phone number if you're to call them or an email address so you can send them an email or an SMS where you know that they own the phone because that is safe for them. Maybe potentially launching voice AI at this time. So I think it is to your point. you have the same number of dollars, where do you want to be able to spend those and figuring that piece out? You still have to have AI. think if anybody takes a message away from this webinar that we did, this is going to be the norm. How you use it is going to come down to your clients, your competitiveness, all these things, your regulatory oversight, what states are you in, what states aren't you in, all those kinds of things. so look at all the opportunities. Just don't get tunnel vision and think it's got to be one thing And one thing only, voice AI.

Adam Parks (43:25)
So Mike, we had a question come across in the comments. I don't know if you have an answer for this, but they were asking, what's the pickup rate of the RPC in the US versus international clients? And do you see a big difference in how the US is answering the phone versus how it's being picked up in other countries?

Mike Walsh (43:41)
It's a great question. I think different geographies definitely have different RPCs, right? Like, so we're in APAC. I mentioned Australia, UK. It's but I would tell you that. It's basically like, let's say it's 3.6 in Australia versus 3.7. It's always consistent in that geography, right? Like the US, the data is probably very similar to what you're seeing. I don't think, like Tim said, no one's going to answer because it's a it, a machine, what they might get with the uptick might be for right parties is I got a tech text message and my data is a little convoluted because it's a omnichannel solution centralized. Right. So I'm sending an email, a text, and I might generate a call from one of those to my machine. And people do prefer a lot of times to choose the machine over the human. They can always say human, right? But what we're seeing is. I always, I think when Tim and I worked together, always said, nobody wants to talk to you about, no one wants to talk to a stranger about how they can't pay their bills. Like who's going to, nobody wants to talk about how much they own student loan debt to people that they're friends, right? Like nobody that those conversations never happen. You never had a dinner party and you're like, Hey, I gotta, I gotta pay, you know, the university this much money this month or the government this, it just doesn't happen. So I think.

Adam Parks (44:53)
Yeah.

Mike Walsh (45:11)
I think that that might go up in the future as people prefer to self-serve. But I don't, that's something I'll look into, Adam and follow-up because I do wonder, like, I don't have all the, like, I have specific clients in my head that I can talk to, but I need to look more. So that is a great question.

But I think it's probably consistent with what agencies see. It's the age of the paper, the data, the quality. We do a lot of like bucket one stuff. That stuff, that answers a lot. But is that what the number you're looking at? Probably not unless you're a creditor. But I can look at like fresh charge off and give you a number and then you can kind of figure out from there. But I think it's interesting.

Adam Parks (45:54)
I'll have to put that into one of the upcoming pieces. Cause it feels like I definitely understand the point of like, we need to look at the data specific to a particular subset and compare apples to apples and not apples to Mac trucks, right? Cause they're just not the same thing. But one of the things that, you know, as we kind of come into our final 10 minutes here, one of the things that I saw in the data this year was an increase in the percentage of organizations that said that their AI investment was.

Mike Walsh (46:07)
You

Tim Collins (46:07)
Right.

Adam Parks (46:21)
exceeding their expectations. The numbers will come out at the RMAI conference when we release the final report, but we definitely saw an uptick in that. I guess my question to you is, do you think that that might be impacted by speed to market? Because if you're building something internally, which is a lower percentage of companies, you're going to have delay, delay, delay, delay. It's not going to get out there as fast. And I'm sure that that's going to have an impact on whether or not you think this investment is meeting or exceeding your expectations. But do you think that the deployment timeline for an in-house solution versus an outsourced solution is playing a role in the level of financial satisfaction that organizations are seeing?

Mike Walsh (47:03)
I think that makes sense to me, right? Like, so if I'm gonna hire EXL to be my AI company, right? I can get that up and running in two months, right? So if I'm going to build my own centralized virtual agent slash two way SMS machine, I'm playing the long game. I don't want to pay a monthly per account fee to EXL That's where I'm going to save down the road. And I think that's part of the strategy of why people do it that way. But the fact that they're happy with the results,

Tim Collins (47:14)
and

Mike Walsh (47:36)
What happens with AI is it's very predictable. It shows up for work, it's on time, it says what it's supposed to say. part of what we do is we map out, here's the lift you should get, here's the lift.

I'm not surprised by that number. As long as you have the right vendor for yourself, it fits what you're doing. It should be very predictable. Adam, you've always, I've always stolen your line about, you know, if they're asking you for 70,000 hours of tapes, maybe you're building the product for them. But, you know, don't know how that's legal, Tim, to share that. But,

Mike Walsh (48:15)
As long as it's somebody who's done it before and you know, they should be able to predict a lift for you depending on what you're allowing. Is it old? it fresh? Is it tersh? Is it primes? Is seconds? Like is it know, utility versus credit card? You know, there's all these factors. it healthcare? We know our models will tell us what we should get based on what you do. Like are you currently texting? Are you currently emailing? You tell us that, we tell you boom. So I'm not surprised by that number at all, but I think it's part of the strategy. The strategy is am I gonna try to save money in the long term or am I ready to go? And you can always do combos, right? You could get a product now and just build internally, you know, and see if it's worth it.

Adam Parks (49:03)
stopgap the current problem and build

Tim Collins (49:04)
Yeah. Yeah.

Adam Parks (49:06)
Something out that you want to build out. I go back to the question of why, because now I got these maintenance issues. Now I got this upfront investment that I got to make not only to roll somebody in, but when we're measuring ROI, right, there's only three things that matter in a business, the cash flow, the timing of the cash flow and the risk associated with that cash flow. So if I'm looking at the ROI of me investing in AI technology, the timeline, in which I'm going to start to see the cash flow based on that investment and the risk associated with that investment are two of the main things that I'm looking at. Am I going to make more money? When am I going to make more money? And how much risk am I taking on to make more money? I can break down any business into those three areas.

Tim Collins (49:46)
No, absolutely agree. I do think, you know, I think a lot of companies were caught off guard, Adam, and maybe that's driving some of these numbers that they were going to have to have this expense because we budget for the year. We're like, this is the budget. You got to stick to the budget. You know, can't move on the budget. And then they realized, okay, co-pilot's not doing it for me. I got to go buy a Claude So, you know, I have co-pilot licenses for everybody. And now I've got, you know, I have my companies on, you know, is now on. know, Claude because they're bigger into the coding or whatever. So you have some of those factors. I do think some of the things that we talked about earlier, the ability to start, implement, execute and continue to do is one of those big issues. So, and with the technology changing so quickly, if we think about what the beginning of 2025 was like from an LLM perspective and where we are now, it is significantly different. And so you may have started building something January 1st got to June when 5.0 came out and said, crap, now you're starting over again, right? Because it's a different model. So you have all those factors that go into it. What Mike talked about is if you wanna try something, because you don't know what's gonna work, let's say you were doing warehouse auto deficiency paper.

I mean, it's been through three agencies and it's just sitting there years now. You there's a bunch of that's still, you can still work that, but maybe that's not right for voice AI. But if it is, and you want to test it, then go find a company and go test it. Don't build it. You know, don't, cause if they, you've done all that investment and it's not working for you. And I think so many companies just said, Hey, we can get a little bit of all of our employees. And then we're just going to put in, you know, energetic new word and jet, AI to be able to do that and said, Now, if your business is based upon relationships, if you remove relationships, there goes your business. So you can build, you can build, yeah, go ahead.

Mike Walsh (51:44)
Yeah. I think that's a really good point Tim. know, we have a Agenic AI and people are like, I walked into a call center and people are like, hey, you guys are the guys with the virtual aid. Am I going to lose my job? Right? And I'm like, no, you're just, your job is going get harder because all the easy calls are going be handled by Agenic. And I'm like, hopefully they'll give you another break. I'm going to tell them to do that. But, To me, do more with the same, not do more with less. Like I think if you map this out right, you're going to handle more of your inventory and there's always going to be those, that's going to discover more edge cases of, wait a minute, I paid this, like, especially if you're in medical, right? Like that's, you're going to need people to do that. You're going to need people for the old lady who says, I won't take my cancer medication. I'll send you the 200 bucks and you're going to get your client on the New York Times. have to, there will always be people. I've heard predictions where all call centers will be agentic AI in five years. I don't believe that. think where I, what I tell people, your career people are going to be there. Your people that are hired just for the two months of the training paycheck and are quitting immediately after training is done. You're not going to have to go through that mess anymore. Right? Like

Adam Parks (53:05)
is one of our biggest expense lines.

Tim Collins (53:07)
Right, yeah, true, very, true. I do think, you know, I think there's a separate.

Mike Walsh (53:07)
Yeah. Nobody hires a friend and class 30. Yeah, and gets 30, right? Like they get five, you know.

Tim Collins (53:16)
Right, yeah, because you always have to over hire, but you know, the AI can help you scale like we talked about earlier. I do think this, I think there's this great opportunity to use AI to help humans be better, right? So if the AI is listening to the call, it can say, you didn't say the mini Miranda, it could pop that up and all of a sudden you could say, oh, by the way, I'm sorry, I'm supposed to tell you this, blah, blah, blah. Or if it's talking about like a medical situation that's, you know, very significant.

You know, it's hard to read everything that happened in the notes and the AI could be whispering that into your ear, popping it up on a screen. So you just made these humans even better. So just imagine, imagine that. And that could be very, very competitive, which gets you the market share that you talked about earlier. So there's all of these things that could be done. We don't have to just run to, yeah, we have a company and I come in every morning and I push one button and we make 10 million bucks. That's not how this piece is going to work.

Mike Walsh (54:07)
you

Adam Parks (54:08)
Well, Tim, you and I talked

Mike Walsh (54:09)
Thank

Adam Parks (54:10)
about on one of the AI hubs about being able to build an entire business, right? And the AI built business, driven business, and I'm not saying that that's not a possibility, but I don't think that that's going to be the reality for the debt collection industry because we exist for the ability to have that touch point, to deploy the compassion and to be able to have those more difficult conversations. I think the more that we can deploy tools to enable us to reach more of those people,

Tim Collins (54:18)
AI first. Exactly.

Adam Parks (54:37)
and to bring them into the fold and to provide the self-service that they're ultimately looking for, right? The better off we're all going to be in the long-term. But gentlemen, I really appreciate you coming on, sharing your insights. I could literally talk about this all day with you guys. I spend my entire day doing this.

Tim Collins (54:52)
Yeah. Agreed. Thanks, Adam. Thanks, Mike.

Mike Walsh (54:54)
We know that we know. Thanks. Yeah, good to see you, Tim.

Adam Parks (54:59)
And that's generally what we do, So for those of you that are watching live or those of you that are watching the replay, if you have additional questions you'd like to ask Tim, Mike or myself, you can leave those in the comments on LinkedIn and YouTube and we'll be responding to those. Or if you have additional topics you'd like to see us discuss, you can leave those in the comments below as well. And I'm willing to bet I can get these gentlemen back at least one more time to help me continue to create great content for a great industry. But until next time, gentlemen, I really do appreciate your insights.

Mike Walsh (55:25)
Thank you, sir.

Tim Collins (55:26)
Thanks, Adam. Thanks, Mike.

Adam Parks (55:28)
And thank you everybody for watching. appreciate your time and attention. We'll see you all again soon. Bye.

Mike Walsh (55:34)
Bye everyone. Bye Tim.