Future of Voice AI in the Collections Industry | Vrinda Gupta of Sequin | Ep. 243

Adam Parks (00:07)
Hello everybody, Adam Parks here with another episode of Receivables Podcast. Today I'm here with a very exciting guest, Vrinda Gupta with Sequin, who's going to talk to us about voice AI, artificial intelligence in general, and boy is this a qualified individual. I had the opportunity to meet Vrinda for the first time at the ACA conference and then spend some time together building our relationship at the RMAI Executive Summit in Vermont recently.

And I was really impressed with the depth of knowledge and understanding of not only the financial services side of the world, but also the use of artificial intelligence and what that's going to look like. So for the thank you so much for coming on today and participating. I appreciate your insights.

Vrinda Gupta (Sequin AI) (00:49)
Thank you so much for having me and thank you for the kind introduction.

Adam Parks (00:53)
Absolutely. So for anyone who has not been as lucky as me to have an opportunity to share a meal together and actually get to know each other, could you tell everyone a little about yourself and how you got to the seat that you're in today?

Vrinda Gupta (Sequin AI) (01:04)
Yes, absolutely. So as context, I started off my career at Visa and I worked there for half a decade building credit cards and managing compliance and audits for top 10 insurers. My main product that I had the opportunity to work on was actually the Chase Sapphire Reserve credit card. And I loved that project. It was incredible to work on such a genre defining product. But when I went to apply for the product I helped create, I was denied.

Adam Parks (01:34)
you

Vrinda Gupta (Sequin AI) (01:34)
And I told that story a million times now and it still doesn't get old. ⁓ And so, you know, after that, I just realized there's a lot of opportunity in our financial system to build. after many incredible years at decided to leave Visa, get my MBA and then become an entrepreneur. ⁓ One of the pieces that was really interesting when I was at Visa on the compliance side is, you know, top 10 issuers, but all financial entities are quite

Vrinda Gupta (Sequin AI) (02:02)
heavily regulated and I do believe that's for the best. Ultimately, these are touching real consumers and touching real consumers' lives. And my first company was actually a NeoBank geared towards women. As I experienced my own denial there and thought there was a lot of opportunity. And running a NeoBank was quite interesting also from a compliance standpoint because every single word matters in marketing materials.

Vrinda Gupta (Sequin AI) (02:28)
And we actually had members of the FDIC reviewing our marketing materials. And I think being on the other side of the table gave me a really good view as to how important compliance is, how important compliance is, and the importance of every single word and also the interpretation of it all. And so my second company now is Sequin AI, and we're building AI voice agents. And how we got here was, As I was evaluating all my experiences across Visa, building on a NeoBank, we actually had great success going through Y Combinator and raising from really high profile investors, including members of the Schwab family and executives at OpenAI. And my co-founder and CTO and I have been building together for a while. We looked at the intersection of a heavily regulated industry and the incredible innovations that are happening in AI, especially it relates to AI voice agents. And we got really excited about the challenge around building in, building in a very heavily regulated industry like collections, but also being able to have this very human empathetic conversation as well. And we looked around and you know, we thought that we've actually been on all sides of the table here. We come from this heavily regulated industry.

Vrinda Gupta (Sequin AI) (03:51)
I've managed audits, we've been on the side, on the other side of regulators. And you know, this is just, it's a really big space, right? And ultimately, I believe if you're able to bring down the cost of loan servicing, then ultimately that brings down the cost of credit for everyone. And so it's a very exciting mission to be working on. And it's led me to really get excited about all of the things that are going on in AI.

And the last thing I'll say here is, is I really appreciate this platform because this is moving so quickly, right? And being able to keep on top of what's going on and really understand it, I think is quite challenging even for us who are building in this space. And so being able to educate and understand, okay, how exactly does this all work under the hood is a really exciting opportunity for me here.

Well, the AI race has just begun. We're at the very beginning of what's happening. I think people might envision that we're further along than we are because there were so many products that came early to the market that were pitching things that I don't think were really, you they were pitching two years ago products that I don't think have really been available to the last three months in terms of the reality of actually creating those. So I've always been or I have been a little hesitant to start pushing down the AI path because it didn't feel like we were really there yet in terms of being able to speak and we've identified six main use cases for debt collection companies to roll out artificial intelligence. And to me, workflow analysis, data appending, negotiation, some of these other things seem to be, we call it lower hanging fruit, more back office things was just where I felt more comfortable starting to roll it out.

But over the last probably three months, maybe since like June or July of 2025, I've really started to go, okay, like, wow, these are starting to sound real because things that I listened to at the beginning of the year, let's call it December of last year, January of this year, just weren't at the same quality level that we're starting to hear and see today.

I mean, how do you feel about this pace starting to move forward? Is this a realistic trajectory? Are we going to start looking at AI in the same terms that we've looked at? Let's call it the data storage curve over the last 20 years. Where it's exponentially on the rise year over year. Like what's your expectation over the coming 12 to 18 months?

Vrinda Gupta (Sequin AI) (06:12)
It's incredible the pace of innovation, right? And as you mentioned, what's been possible in the last few months here was not possible last year this time, or maybe it was possible, but it was cost prohibitive, right? And so it's so incredible to see the pace and the fact that GPT-5 just launched a few days ago. I mean, that's so exciting, right? And you see these meaningful performance improvements that are game changing for industries, right? And are changing the way that industries work overnight.

Vrinda Gupta (Sequin AI) (06:42)
That's not just a phrase, right? It's truly overnight. Something is launched and now things are markedly better. And so the AI technology is something that I'm extremely bullish on and it's not a fad, it's here to stay and it's real and it's working too. So I'm just extremely excited about the pace of technology here. And I think it is more a matter not of.

Vrinda Gupta (Sequin AI) (07:04)
if you will be adopting AI, but it's more when you will be adopting AI, right? And where on the adoption curve do

Adam Parks (07:12)
I think we've only got probably 18 months before you start really getting left behind and 18 months I think is a really generous time period. I mean there's even some really simple things. I know when we were sitting at the, we're sitting at dinner at the RMAI Executive Summit, we were talking a little bit about debt buyers and the number one complaint that debt buyers have on Google in terms of their Google reviews is that people didn't respond fast enough, they didn't get a return to their voicemail or they didn't answer the phones.

Adam Parks (07:38)
And when I said to you, like what we really need is just the ability to direct the consumer to the right servicer, whether it be a collection agency, a law firm, you kind of looked at me like, wow, like it can't possibly be that simple. But I do think that some of the functions that we're looking to achieve as an industry that can make a truly marked difference in being able to get those consumers to the right places and reduce the amount of friction between I found you and I paid you is really what we're looking for as an industry.

Vrinda Gupta (Sequin AI) (08:08)
Yeah, for sure.

Adam Parks (08:09)
I you it can be that simple.

Vrinda Gupta (Sequin AI) (08:12)
Yeah, no, I remember that conversation, Adam. I think it's, you know, it makes sense, right? All this stuff is moving so quickly. And I appreciated that conversation, right? In some of the context where, you know, where are folks today and where are they looking to be as well? And especially in just a fast moving environment, I can understand and also appreciate a natural level of skepticism, right? Where it's like, is this going to work? And having to completely...

Vrinda Gupta (Sequin AI) (08:37)
change your business operations within 18 months. I mean, that's so quick, right? And so I think it is interesting to see where we're at as an industry, but also what the potential is to move and also the requirement to move.

Adam Parks (08:53)
I think those FinTech customers, those that have borrowed their money online are most likely to want to interact with a bot or anything other than a live person. Do you think that we'll start to see a generational shift or that the, let's say the other generations, the boomers are going to start actively engaging with the technology at a faster rate than maybe they started engaging with the internet, for example?

I mean, think what's interesting is a lot of this is kind of a step function, right? So you're able to interact with these technologies and they work, right? And so one of the pieces that I've been talking a lot about has been, you know, when you are interacting with an IVR, right? And it's saying, you know, okay, tell me a little bit about your problem. And it's just listening for keywords, right? And I think you put it really well, Adam, in terms of if-then statements, right? And the reason that you're smashing the talk to a human button is that this isn't actually solving your problem. And what I find really fascinating with AI technology is that this is actually, think of your most competent customer support agent. Hey, think it's your product person that actually made the technology is able to answer this question for right? And so the level of support that you're getting is actually higher, much higher quality than maybe your average, you know, support staff. And the other piece too is, you know, especially with collections conversations, those are sensitive, right? And so sometimes chatting with an impartial AI might actually be a bit of a safer space than talking to another human as well. So I think there is some general trends here, but also some industry-specific trends that make this especially compelling.

Well, when it comes to chat and it comes to voice AI, I think one of the big concerns that the industry has had has been hallucinations. We've all we've all played games with ChatGPT and tried to provide entirely too much context to the little tiny context window that it's capable of truly understanding. And as we were kind of preparing for for the podcast, we were talking about.

Vrinda Gupta (Sequin AI) (10:45)
you

Adam Parks (10:57)
the realities of hallucinations. Talk to our audience about what are hallucinations and what's the realistic approach to limiting them because elimination doesn't seem to be an option.

Vrinda Gupta (Sequin AI) (11:11)
Yes, and Adam, is it helpful to maybe define some of the, I know we said context window and yes, yes, okay, sounds good. Yeah, my mom's a teacher and so think I have some teacher blood in me so I'm always like, okay, let's define and let's go. So I mean 101, right? LLM, large language model, right? And essentially what these things are is it's essentially a lot of text that is brought together and trained on. And so it's kind of,

Adam Parks (11:15)
Yeah, let's hit some definitions here too. ⁓

Vrinda Gupta (Sequin AI) (11:41)
Probabilistic right and so based off of you know all of this information How might a conversation go given some context right and so the way these LLMs are trained is they're actually trained to please and They actually really want to answer your question right, but the challenges happen
One, in terms of context windows, as you just mentioned, Adam. And a definition of a context window is basically, you don't want the conversation to remember every single thing that's happening, because that's just going to take a really long time. And also, it might pull from the wrong part of the conversation. So it puts you at risk, right? The second piece is in terms of what you need to think about in terms of how these things are trained is they also aren't going to know anything they don't know, right? So if there's something.

Vrinda Gupta (Sequin AI) (12:29)
specific that it needs to know in order to please, it's just going to say, probably this is the answer, right? And that's basically a hallucination is, hey, you know, based off of what we know, this is what we think the answer is going to be. but sometimes that's just nonsense or doesn't make sense or it's pulling from a wrong part of the conversation. And so, you hallucination generally

Vrinda Gupta (Sequin AI) (12:52)
It's funny, right, when we're all playing on chat cheeky tea, but when it comes to a heavily regulated industry, just not okay, right? It can't, it's not funny. It's putting you at regulatory risk and it's real, right? And I think anyone that comes to you and says, we are eliminating hallucinations. I would, you know, really be cautious of that. Cause it's not possible to eliminate hallucinations. In just that.

Adam Parks (12:58)
It's not as funny.

Vrinda Gupta (Sequin AI) (13:18)
These are probabilities, right? It's creating probabilistic conversations. And in any probability, there's a margin of error, right? And so that's going to happen here as well. So there's a variety of ways to be thinking about how exactly to minimize those. And we can go more into detail. But Adam, is there anything on a 101 that you think we should define before going into that?

Adam Parks (13:38)
So I just one thing I want to go back to in the context windows and I want to try and provide maybe an illustration to what a context window actually is. So from, you know, from my perspective, if I've got if I'm trying to write a summary of multiple reports and I load 20 reports to a ChatGPT thread, and I start asking questions, it can't understand all of that context. It's really good at short sprints. It's really bad at long marathons.

Vrinda Gupta (Sequin AI) (13:47)
Yeah.

Adam Parks (14:07)
And so when you're communicating with these models, being able to limit that context window or at least understand that the context of the response is limited, think is an important, it's important for people to understand what that is. And it took me some time to understand why it kept forgetting what I wanted it to understand. And the more that I started reading about that, I think the easier it became for me to, kind of see how big or how small the window was within specific models. just something, I don't know if there's an easier way to explain that. Any other addition to that?

Vrinda Gupta (Sequin AI) (14:43)
No, I think that's great. I think you explained it well. And one of the things that GPT-5 has actually improved upon in its performance update is its ability to lengthen that context window. And so it can just get faster in terms of pulling from different parts of the conversation. So I think you explained that well.

Adam Parks (15:01)
I mean, that's ultimately where we're going to. But as we started planning for this, we started talking about kind of five critical questions for evaluating AI vendors. And the reason I thought this was such an interesting conversation, and I want to say it was at the eight, I it really started at RMAI. And then at ACA, I want to say I was approached by no less than 15 different vendors that were walking around with a phone and saying, like, listen to my demo. And it starts to become so much white noise to it that you have to start asking yourself, okay, how am I going to weed things out? How am I going to get a little bit more direct and understand? How can I evaluate this vendor's understanding of my real needs? Even going back to December and January, I was approached by a lot of companies from outside the United States that were going to take their AI voice and they were going to bring it to the United States but my first two questions are always the same. And then I want to go into the five that you and I talked about, but my first two are always the same. Who's your who's your attorney? And who's helping to manage compliance? Like, do you have an understanding of it? Because I think when you take a group of technical folks, and you drop them into the debt collection industry, they assume that the rules are clean, clear cut, and easy to understand. When in reality, it's a giant web of items from a state, a federal level, even some local jurisdictions and what that means. And as you and I started talking, it became clear that you had an understanding of financial services compliance, which I think is a massive piece of the puzzle here in order for us to be able to safely deploy this type of technology and to have the level of, it's called depth of understanding that's necessary. But then the questions start to come in. So like once we've gotten past like, is this, not even necessarily is this AI or not, but like, do you have an understanding of the compliance framework that's required? Then we start talking about things like, are you using generative AI for responses? Or is this a conversational IVR? And you had mentioned a little bit about how the conversational IR is really looking for those keywords, but talk to me a little about why those things are different and what somebody should ultimately look for when you're trying to find the needle in the haystack here, when you're trying to find like a vendor that actually can assist with your needs and cut out some of that white noise. How do you identify whether this is an agent AI response or if this is a conversational IVR system?

Vrinda Gupta (Sequin AI) (17:35)
Yeah, I mean, think these are great, Adam, right? And I think I'm still looking forward to getting into this because, you know, it is challenging to understand what to ask, right? And, you especially as you're listening to a demo, a demo is not where you're going to get.

you know, the most interesting pieces of information. And so I think it's quite important to be able to understand not only what questions to ask, but also what to expect, right? So I love these questions. So yeah, so I think starting off, I mean, we defined LLMs and generative AI, right? And

Yes.

Vrinda Gupta (Sequin AI) (18:08)
I think question number one is, I mean, how long have you been doing this? Because if folks come to you and they say, I've been doing this for two years, that is actually a bit of a red flag because this technology didn't exist two years ago.

Vrinda Gupta (Sequin AI) (18:22)
I think, know, just understanding if someone says that they've been doing this two years ago or even beyond the past six months, likely what they are doing is closer to machine learning, right? And machine learning is great, but it has some really key limitations, especially as it relates to collections use cases. So the difference between machine learning or let's say a conversational IVR, right? And basically a conversational IVR is more of

Adam Parks (18:22)
Exactly.

Vrinda Gupta (Sequin AI) (18:51)
pre-created scripts and it's not really understanding the context, right? And so if I say something, it's just going to pull from a specific script and a template and it can say that in a pretty human, realistic sounding voice. But what you're not getting is actually the context and the understanding and the empathy and the ability to negotiate, right? Because if you're just reading from a script, how are you supposed to provide that empathy outside of something generic? How are you supposed to negotiate based off of someone's specific circumstance?

Vrinda Gupta (Sequin AI) (19:21)
So that's what's really fascinating about generative AI is it's understanding the context, it's able to have empathy, and it also is understanding the outcome that you want and can really kind of probabilistically go towards that outcome based off what someone is saying. So the main differences is...

they might be, if someone's using machine learning or if they are, you know, leveraging a conversational IVR, they might still be leveraging LLMs, but their responses are not being generated using an LLM. And again, that's really important to reach the business outcome, which is you want to have a collections outcome and you want to have a promise to pay and you want to make sure you're in compliance.

So if I distill that down, what I hear is that the agentic AI responses have context to the discussion itself and the conversational IVRs are if and then statements. So it's a decision tree. It heard the keyword operators, now it's going to drop me into this operator queue or it heard this keyword payment plan and now it's dropping me into a payment plan scripted response. And you can make that, I I sound real right now, right? But I'm.

Vrinda Gupta (Sequin AI) (20:19)
Yes.

Yes.

Hahaha

Adam Parks (20:31)
clearly speaking into a microphone. I think it works pretty much the same way. It can play those prerecorded messages in a series that makes more sense, but it is still IVR driven. It's still press one for this, press two for that. Or you can say one, but that doesn't make it aogenic AI, right? Just because I said one and you now moved me down a different if and then process. But would you say that adaptability...

Vrinda Gupta (Sequin AI) (20:48)
Exactly.

Adam Parks (20:57)
to the conversation is also a differentiator there. I mean, how complex do these if and then statements become and how close to a similar type, it's called user experience, could you come up with using that conversational IVR?

Vrinda Gupta (Sequin AI) (21:11)
conversational IVR can be pretty complex, right? Let's say you put in a million keywords, right? But still it's not going to really, it's not really gonna understand the context of the conversation. For example, if I say, you know what, I lost my job, but actually I think I'm gonna be employed in the next six months, a conversational IVR can't really take that and run with it.

Adam Parks (21:19)
It sounds useful.

Vrinda Gupta (Sequin AI) (21:34)
But generative AI can say, okay, yeah, well, in six months, let's put you on a payment plan. Let's do, really adapt to that situation. And that's what leads to a really good outcome. And especially with the negotiation and with collections, right? You need to be empathetic. And so really understanding that person and their circumstance when they share, I lost my job, being able to reference that specifically versus just jumping to, okay, when can you pay, right? That's not.

You know, that's not going to lead to a good outcome, right? So I think just being able to have that human empathy, having the context, it's a difference, maybe a good analog to this is it's you and I speaking right now, Adam, is you're listening to me, I'm listening to you, and we're riffing off each other. And that's why this conversation is interesting. But if I only had a preset, you know, kind of script that I could say, and you can only respond to that preset script with prescriptive things on your end.

Vrinda Gupta (Sequin AI) (22:26)
I don't think anyone would be listening to this conversation. So I think that's a good analog.

Adam Parks (22:29)
No, that's a really good point.

I really like that point in terms of how that balances. When it comes to the conversational IVRs, I think that's just the technology that we were looking at, let's call it a couple of years ago, and it's gotten, let's say it came out five, six years ago, and it's gotten better and better and better. But it hit that Peter principle in terms of, I think it's become more frustrating from a consumer experience than it has become a benefit.

certain airlines were using that technology and I just could not speak to it now that some of them have moved towards the agent at AI experience. Now I can have again more context to the discussion about my need and their ability to respond to that or to execute on a task on their side has improved pretty dramatically, but it's beyond just the conversational IVR. I think it's also that ability to execute on something.

on your side to send out that payment reminder to write to trigger the signals that are happening from a voice is that communication methodology, but now it needs to feed the signals to the other aspects of your it's called technology and communication channels. How does that have you seen that as a as a usable force as you've started to deploy a genetic AI within the financial services space?

Vrinda Gupta (Sequin AI) (23:23)
Yes, take the action.

you

That's what's really incredible here, Adam. And you and I talked about this, think, at dinner at RMAI, right? Where one thing that's interesting is historically this industry has been quite fragmented, right? Where there's your dialers of the world, then there's your human servicers, then there's your payment processors. There's specific companies for voicemail drops, right? And everything is quite fragmented. And I think what is incredible and truly disruptive, right? That disruptive word has

Adam Parks (23:56)
Mm.

Vrinda Gupta (Sequin AI) (24:14)
and I think overused, but I think this is the time where it is and we're guilty. We're sitting right in the middle of San Francisco right now, but I don't think it is overused in this case, right? Because what you're talking about.

Adam Parks (24:16)
Overused. It's very Silicon Valley.

Vrinda Gupta (Sequin AI) (24:30)
And in terms of the agenda capabilities of this means that it can take actions and complete multiple actions. Right. So now when you think about voice, say I technology like hours, for example, you don't need five different services to make this work. Right. We'll do the call. So you don't need a dialer. We'll actually service the call and have, you know, that human-like interaction, we will transcribe and put everything together and we'll also facilitate the painting, right? And it was interesting when I first kind of entered this industry having conversations with collections experts like you and your peers, right? That was really incredible where they were like, so we don't need the dialer. And I'm like, yeah, all of that is just all together. And I think that overall context is what makes this so compelling and the ability to not only

Vrinda Gupta (Sequin AI) (25:19)
you know, have human-like interactions. So you're reducing the burden on your human agents, but also the ability to take those actions is just, it's so incredible and even takes it a step beyond kind of your SaaS companies, right? Because you're having kind of the human capability and the action capability in one.

Adam Parks (25:36)
I mean, generally those conversational IVRs are just going to push to another live agent to take that next live action, being able to move those signals across your siloed technology stack and activate the whatever trigger is necessary for next action, I think is where the future starts to lie. But then we start asking ourselves kind of our next question here, which was how do you evaluate and test AI conversations from a quality and a compliance standpoint? It's all fun and games.

Vrinda Gupta (Sequin AI) (25:55)
Yeah.

Adam Parks (26:05)
the gets involved. So really kind of a two-part question here, but we have to look at two different things. Sitting on your side of the equation as a technology provider, we have to look at how are you going to train this and how are you going to make sure that it's staying on the rails? And then from a buyer's guide perspective, what questions do they need to be asking? And even what questions should they be asking in a live demo to demonstrate the breadth of the product in front

Vrinda Gupta (Sequin AI) (26:31)
I think this is a great question, Adam, because again, know, listening to a demo, it's going to sound great, Everyone's demos. Yes. Exactly. That's a good way to put it.

Adam Parks (26:37)
You make a demo sound however you want it to sound. That's why they call it a demo.

Vrinda Gupta (Sequin AI) (26:45)
Yeah, this is the crown jewel of all these AI companies is basically how are you evaluating a conversation? Right? So of course, the front edge, the prompting is really important because ultimately that's what's going to guide the conversation. but also how are you evaluating those conversations and how are you evaluating them for three things we say, right? The first is how do you make sure the conversation itself went well, right? Is this interrupting someone? That's a big thing.

Vrinda Gupta (Sequin AI) (27:14)
right, interaction handling. As I was interviewing collections experts, I asked, actually, what's your number one tip for me? And are there, is there any kind of secret sauce we can put into our AI? And they said, just listen, right? And, and interrupting and interruption handling is a big piece of voice AI, right? And so I think just understanding, you know, how did that call go?

Vrinda Gupta (Sequin AI) (27:34)
The second thing you want to look at is, what's the business outcome of the call? Did it get a promise to pay? Did you actually receive a payment? Whatever else? Did someone say cease and desist? And is it capturing that action? It's all of those kind of business outcome pieces. And then the third piece is from the regulatory side. And as you were talking about previously, there's federal regulations, there's state regulations, there's district by district regulations. So just making sure that the conversation is not putting you at risk. And now as we talked about the beginning of this, of our conversation here, every single word matters, right? And the order of that word matters as well. So making sure that your same engineers that are kind of creating your product, you want them to have a really intimate understanding of the regs and not exactly, not even just what the regs are, but how are the regs interpreted, right? Are they looking at, you know, cases that are recent cases that are coming out? Are the members of these associations right to understand, okay, this

Yeah.

Vrinda Gupta (Sequin AI) (28:32)
is not only what the reg is, but how is it being interpreted? How is it specific to your specific asset class, to your demographic, right? And then how are you taking those learnings and then putting those intentionally back into your original prompt? And I'll add one thing here, Adam, that's really, it's a question I get frequently and I just feel the need to share it on this platform is the question I get so much is, well,

Adam Parks (28:48)
Please.

Vrinda Gupta (Sequin AI) (29:00)
I just want this to self-learn and I just want this to be out the box and I don't want to deal with it. I don't want to put any more resources into it. And the reality is I shudder at that question as someone who just has a lot of experience, the compliance side of this, because these things again want to please, right? And so you don't know what this is putting back into your...

Vrinda Gupta (Sequin AI) (29:24)
into your original prompts, right? And so whichever vendor you work with, you want to make sure and feel very comfortable that they have an understanding of how this all works, what the regs are and what each sentence needs in context in order to ensure that they can flag items for you, get your perspective through your specific use case, and then make sure that they're improving their prompts from your specific use case. So those are kind of the three high level pieces.

Adam Parks (29:26)
Yeah.

Vrinda Gupta (Sequin AI) (29:53)
We can go deeper in terms of, I know we talked about judge LLMs and those pieces, but...

Adam Parks (29:57)
Yeah, that's actually that was my next question, right was, you

that was an interesting concept, because I had heard the concept previously of, well, this model watches that model. And that one, that model is going to watch this one over here. And I never heard the term judge LLM before, which makes all the sense in the world. Could you talk to the audience about what that actually means and put it into the context of voice AI?

Vrinda Gupta (Sequin AI) (30:06)
Yeah.

Yes.

Yeah, for sure. So again, all of these AI companies, we are building evaluation suites, right? Because how do you know the conversation went well? And so we're looking at before a conversation ever happens, what is your test week, right? And how are you testing these conversations? What are questions you should be asking? Right. And I think, you know, really low-hanging fruit questions are just making sure, you know, if this says, don't call me, don't text me, right? Are you a human?

Vrinda Gupta (Sequin AI) (30:49)
especially

for outbound calls to the CPA, like how exactly is that all working, right? Making sure you are comfortable with that and your regulatory counsel is comfortable with that, right? During the conversation, what are your stop gaps, right? What are the escalating, someone says lawyer, right, needs to escalate, right? So what are those conversation paths? But also if the conversation is really going off the rails,

Adam Parks (31:08)
Yeah. Sure.

Vrinda Gupta (Sequin AI) (31:16)
What's a kill switch, right? How can you just gracefully end that conversation, say, actually I'm pass this to a manager or whatever else, right? And then post the conversation, that's where you can have a lot of fun, right? And fun in that you can get compliant, I guess is maybe the right way to say, you what's fun here. And some of the kind of industry jargon that I'll share is judge LLMs, right? And essentially what a judge LLM is, is exactly that. It's a different LLM.

Adam Parks (31:31)
Yeah. ⁓

Vrinda Gupta (Sequin AI) (31:44)
that's actually trained to evaluate the conversation. And you may have multiple different judge LLMs. Again, as I mentioned, one can be for how the conversation flow went, the second could be on the regs, and the third could be on the business outcomes. And you will kind of have these little mini models and that will evaluate the conversation for you.

Vrinda Gupta (Sequin AI) (32:07)
And then it'll kind of surface insights for you and help you understand, okay, hey, is the way that something was said, was that maybe not the best way to say it and next time make sure you don't do that. And it's just kind of replicating what you do with the human agent where you have a new collector coming on board, you have a trainer, you're spot checking some of these calls. But what's great about the Judge LLM is you're spot checking every single call across all of these different verticals.

Vrinda Gupta (Sequin AI) (32:33)
which is quite powerful, especially regulated industries like this one.

Adam Parks (32:36)
That was one of the first moving AI products into the space was the ability to start listening to 100 % of the calls and be able to feed forward to those live compliance professionals the specific calls with the highest probability of a need of their manual review. So I think that's something that the industry has become somewhat comfortable with. But the idea of the model watching the model or multiple models watching what's happening here, I think can help to increase the level of comfort.

Vrinda Gupta (Sequin AI) (32:42)
Yeah.

Adam Parks (33:04)
that organizations have in deploying this type of technology at scale and in the wild because it's all fun and games when we're playing in the sandbox. But as we start to roll things out at scale, having these additional risk mitigation tools, knowing that hallucinations can never be zero, but we can put these types of things in place to continue to constrict those risks and mitigate our risks in those challenges.

Vrinda Gupta (Sequin AI) (33:10)
Yes.

Adam Parks (33:30)
One of the things that I've started to recognize early on is voice AI has started to move into the space that there seems to be two business models that organizations are using to deploy this type of technology into the space. And so we have some organizations like yours that are software providers and you're coming in, you're providing software to an agency, a bank, debt buyer, whatever the case is, and you're rolling that new.

product out as a software provider. On the flip side, we've seen organizations that have either acquired agencies or have started to license themselves as an entity, as a debt collection agency, and they're going to be a virtual, let's call it debt collection agency or virtual call center. I mean, those two types of models seem very, very different.

What brought you to choose the path that you chose for your organization as a software?

Vrinda Gupta (Sequin AI) (34:22)
This is such a, it's an astute question and can only come from someone who's really looking at the landscape deeply. So I appreciate this one, Adam. So, you know, this actually steps, sorry, were you gonna say something?

Adam Parks (34:33)
No,

no, go ahead.

Vrinda Gupta (Sequin AI) (34:35)
So the reason that we are taking the software first approach actually stems from my experience at Visa. And that's more on the origination side, right? And what's really interesting is, so I worked on the U.S. consumer credit team. However, Visa had a variety of different geography specific credit groups because it's quite different across geographies, right? And even when you think about Asia, there are three different groups of credit, of consumer credit.

Vrinda Gupta (Sequin AI) (35:02)
within Asia, right? Because each of geographies differ a lot. Cultural attitudes towards lending, know, the way that these products work, the way the products have to work, right? They're all quite different and quite specific. And I think what is very interesting about AI in general is the ability to hyper-personalize to specific use cases, right? So whether that is to geography, whether that is to your specific demographic, whether that is your specific asset class, whether that is, you know, days past due, right, whatever, or maybe, you know, you've been subject to regulation in the past and there's something that, you know, you really need to be on top of, or maybe there's new regulation that's come in that's affecting your asset class, right? We all know that's going on. And so, you know, being able to be a software provider allows you to hyper personalize and really take the lead from whoever you're working with right, whether that's a creditor, whether that's a servicer, right? You can say, hey, let me take your best practices for your very specific use case. And, you know, it's not really a one size fits all. And I think that also allows, you know, to go very deep in collections, but also be able to go live.

Ultimately, there's so much opportunity in terms of lending in general, from the rich nation all the way to the collection charge off site as well. And so that's quite exciting too.

Look, I think it's interesting to see how these business models have started to evolve. I understand that somebody might want to be a collection agency in order to do their proof of concept and whether or not that's actually lubricating the deployment of new technology versus having to ask internal IT resources for assistance. But as the systems of record have become more API driven, I feel like that should be easier and easier for the technology departments to interconnect these different systems.

I questioned the difficulty in being able to interpret the incoming signals for next action, but I don't see it becoming any easier in doing it through an agency model versus doing it through being a technology provider, because I would imagine that the level of integration you're able to complete as a technology provider is significantly different than if you're stuck running on the same communication and data rails as every other agency in their stack.

Vrinda Gupta (Sequin AI) (37:24)
you

That's really exciting, right? And going back to kind of what we talking about previously with kind of the state of the current state is things aren't quite fragmented, right? And I think that's how you do get a better view of an individual, a borrower, another business is to be able to kind of see across the board, you know, what are the communications? What's working? What is your tech stack, right? And being able to evolve. And the other piece I'll add to, Adam, and this is just being able

Vrinda Gupta (Sequin AI) (37:53)
AI native as well and being flexible with the technology, there's so much changing, right? As we mentioned, GPT-5 just launched a few days ago, right? And so being able to evolve with these models is quite exciting. And one tidbit I'll share, and I think we'll probably go into this with speech to text and text to speech. So maybe I'll kind of put on this a little, but it is very exciting when you are built

Yeah.

Vrinda Gupta (Sequin AI) (38:18)
as a technology first company because you're able to take advantage of all the amazing and rapidly evolving technology that's in this space to create a best in class solution and keep on evolving.

As we've looked at the questions related to what should I be asking of an AI vendor, one of the things that we were talking about was does your team have experience in the US financial regulations? And one of the things that I don't think we pointed out as we talked about it earlier was the B2B versus the B2C environments. And I think there's an interesting challenge here as I've talked with organizations that were coming in from outside the United States and they said, you know, we're going to go to consumer debt collection because

If I'm not mistaken, the United States represents roughly 25 % of the entire global market of consumer debt collection, but the regulation comes with that. Whereas in the B2B environment, higher balances. I question how AI will start to impact that particular marketplace, but with a significantly lower barrier to entry, I can imagine that some organizations are starting to look down that path. What drew a view towards the B to C focal point for your organization.

Vrinda Gupta (Sequin AI) (39:28)
Yeah, I mean, I think that's the most heavily regulated part, right? And I shared my experience across the space, right? And being on the Visa side, running audits for top 10 banks, my own Neobank and having literal members of the FDIC looking at my marketing material, right? I think, you

Adam Parks (39:33)
Absolutely is.

you

Vrinda Gupta (Sequin AI) (39:47)
B2B regulations are, even in the US, are significantly lesser than direct-to-consumer work. And again, I think that's for the best, and I take this very seriously. And ultimately, the financial industry is the backbone of our society, right? And so being able to uphold what is intended and just protect consumers, I think is just quite important. And so I think going from, you know, Being able to say, hey, we've done this for B2B use cases, but then saying, OK, we're just going to take that exact same engine and put that towards consumers, I think is honestly irresponsible. And again, would question, ultimately, that could really put your business at risk. But I would not take that approach just for the risks and for how nuanced this all is.

Adam Parks (40:40)
I think there's entirely too much risk in trying to take a B2B tool and say, well, this is B2C now, just like I think that there's a lot of risk in taking something and be like, well, this worked in Kenya, it'll work in the United States. That's not how this works. Like there's a differentiator there, the understanding and the same reason that I always ask the first same two questions of a new AI vendor, who's your lawyer, who's your compliance folks, like which teams are you working with to help you understand this marketplace? And when I had that conversation with you, we went,

Vrinda Gupta (Sequin AI) (40:51)
No.

Adam Parks (41:08)
way down the rabbit hole. There's others who have had that conversation with you. They're like, what do I need a lawyer for? I'm an AI company. Whoa, whoa, whoa, whoa, whoa, hold on, hold on. Well, that's, so that starts, not only is that scary in terms of their understanding of the marketplace, but now how are they gonna be training those conversations? What's their understanding of data privacy from the United States perspective? Where is the data being housed and stored in all of those things? But.

Vrinda Gupta (Sequin AI) (41:13)
my gosh, scary. Scary, scary.

Adam Parks (41:31)
I think the next question that you and I had talked about over dinner in terms of starting to make these determinations was how are your AI conversations being trained? What guardrails are you putting in place to actually make these conversations compliant?

Vrinda Gupta (Sequin AI) (41:45)
Yes, yeah, for sure. I mean, think that's where a lot of the understanding of the space and understanding of the regs comes in, but also...

Vrinda Gupta (Sequin AI) (41:55)
I think an empathy for your clients, right? And making sure you're listening to, okay, a lot of, you what we talked about yesterday, Adam, in prep for this call is, you know, a lot of the regulations are kind of evolving around AI as well. And different companies may have different views on this and different councils might have different views on this as well. And so making sure you're flexible, making sure you're listening. I know as a...

Adam Parks (42:06)
Yes.

Vrinda Gupta (Sequin AI) (42:17)
You know, I've been on the business side. It feels very tempting to say, okay, I want to just have something that's out the box. That's going to work that I don't really need. It's going to self-learn, right? That's where this all comes from. And I get it. Everyone's busy with their own work, but I think the reality. ⁓ great. Yes. And it shouldn't exist. I think that's maybe the,

Adam Parks (42:31)
I would love the silver bullet, but it don't exist.

Vrinda Gupta (Sequin AI) (42:38)
nuance here is not that this is a limitation of the technology. I think it's just quite important to understand end to end how these things are trained and how they work. And a few pieces that I share with folks as, you know, kind of a one-on-one is again, going back to that conversational IVR machine learning, right? This isn't a predefined script. So the way that you basically train an AI is you're giving it scenarios.

And you're basically saying, okay, these are some common scenarios. These are how, you know, conversation should work. When it comes to collections or something that's as heavily regulated, the order matters, right? You better not be, you know, disclosing any information about that debt until you authenticate someone, right? So when you're asking questions, are you doing kind of the single shot prompt or do you have a workflow or a state machine where it's actually not even exposing that personal information?

until this authentication step is being passed, right? So it's quite important to ask those questions and feel really comfortable with the way that these are being built because the best way to ensure that you're not out of compliance is not even exposing the information to the AI in the first place, right? So that's something you can do in advance, for example.

Let's

talk about that in a little more detail. So I'm going to try and simplify this a little bit for the audience. And you tell me if I'm on point for this description. So similar to the way in which a consumer goes to a portal, they have to provide three pieces of information, whatever the flavor is. And once I've got these three pieces of information, now I can log into the portal, I can see my balance, I can see the other things, you're saying it's taking the same approach, or the model would be taking the same approach in terms of

Vrinda Gupta (Sequin AI) (44:01)
Yeah.

Yes. Okay, sounds good.

Adam Parks (44:26)
data availability. And one of the great ways to restrict its ability to respond to something is to restrict its ability to even see the underlying data set until it has achieved the objective before it, meaning authentication is objective one. Once we've been authenticated, then we can talk about balances or we can talk about original creditor or whatever the question may be. But that's that almost that paywall that sits in between. Can't log into the account without a username and password. Can't see the balance.

Vrinda Gupta (Sequin AI) (44:50)
That's a great way to, yes. Love it.

That's great. Love it.

Adam Parks (44:56)
Okay,

well I just want to make sure that I simplify it because when we start using some of the terminology like workflows and versus state machines, I want to make sure that our audience has a clear understanding of really what we're trying to accomplish here. We're restricting the visibility of the data to that model so that it can not talk about the balance. It can't make a mistake and let the balance slip if it doesn't know the balance to begin with.

Vrinda Gupta (Sequin AI) (45:18)
Exactly. Yes. And I think the other piece too, Adam, in terms of, you again, when it comes to, I'm always going to go back to compliance here, right, is, if there are specific

Vrinda Gupta (Sequin AI) (45:29)
responses that need to be said a certain way, even within the prompting, you can say, no, if this comes up, read this basically exactly the way it needs to be said, right? And so there is kind of an art and a science in all of this that requires a deep level of understanding of what is important and what is not important to say very specific ways, right? But this is different from using any other sort of technical provider because your traditional, let's say your traditional SaaS company

Vrinda Gupta (Sequin AI) (45:56)
They are coding, right? But it doesn't require a lot of understanding, right? It's going to be a PM somewhere over here that kind of tells the engineers, okay, this is exactly what you need to build. And they build it, right? There's not much context.

Vrinda Gupta (Sequin AI) (46:08)
The difference in building AI companies is that the engineers need to understand the business and the regulatory context around all this because it's all prompt engineering. And so you're writing prompts in certain ways to ensure that it's meeting certain outcomes and certain requirements. And so your engineers and your entire team needs to understand the regulatory context. And I think that kind of comes back to, know, when folks are overseas and might not understand the regulatory landscape in the U.S., that's

Vrinda Gupta (Sequin AI) (46:37)
where it kind of puts you a bit at risk in these conversations.

Adam Parks (46:41)
And

if there's nobody looking over your shoulder, the level of risk that you're creating is significantly higher. So we're you're kind of putting people into a weak position on day one. Now, the other thing that we've talked about, and I've definitely seen in terms of how this technology has evolved over the last, let's call it six to eight months, has been the reduction in latency as you listen to these demos, as you participated in playing with these tools.

Vrinda Gupta (Sequin AI) (46:47)
Yes. ⁓

Adam Parks (47:08)
in 2024, there were these big long gaps between my statement and the response that was coming back. So that latency has to come from somewhere. Talk to me about how you've been able to reduce that latency to such a level to where these conversations now sound significantly more natural, almost to the point of falling into the uncanny valley of artificial intelligence.

Vrinda Gupta (Sequin AI) (47:32)
Yes.

Yeah, for sure. I mean, I love this conversation, right? Because especially with voice AI, the latency is quite important because anything over half a second, essentially of a pause, starts to give you pause, right? It's like, hey, did they hear me? Did something happen? Did they drop, right? And so having that consistency kind of makes you fail the Turing Test a bit where you're not going to think that you're having a human realistic conversation. So having the latency low is quite important.

Vrinda Gupta (Sequin AI) (48:02)
and there are different ways to do that. But before I get into that, should we talk about text to speech here? Is this the moment? This is the moment.

Adam Parks (48:09)
Yes, I think this is how I think this is the moment that we really start talking about like that, that reduction in latency for the future comes from an ability to translate differently. So you explained this to me at dinner, I thought this was very interesting. So I'm please help help help our audience understand.

Vrinda Gupta (Sequin AI) (48:20)
Yes.

So it's good, you and me have a bottle of wine and talking about latency is a good time. My idea of a good time. Okay, so one thing that is amazing about AI voice agents is the fact that the latency is quite low because of what is happening under the hood. And the difference between Adam, you and I having a conversation is that we're just listening to each other and we're processing that real time, right, in our brains.

Adam Parks (48:30)
Yeah.

Vrinda Gupta (Sequin AI) (48:55)
What's actually happening today with LLMs, the way that voice AI works, is it's actually not understanding the speech, it's understanding the text. So it's converting the speech into text, putting it into the brain, which is your LLM, that's actually taking the text, and then spinning text back out and then converting that back into speech.

And what's amazing is that this is actually happening pretty quickly. Imagine if everything you said to me, I need to write down and then I need to process it and I need to write it back and then I need to say it back. That would take us a really long time, but these are doing it quickly, right? And there's a lot of ways to optimize that. And we'll talk about that, right? The benefit of being kind of AI native and tech first is there's a lot of innovation happening. And so one of the interesting innovations, lot of work that's being done right now is actually on speech to speech technology and it exists, it's not good. again, you know, maybe tomorrow it'll get good, right? With the pace of how this is all going. But today, if anyone's selling you speech to speech technology, I'd be quite wary of that because it's not great at interruption handling, right? It's also, there's a lot of limitations around it right now.

Vrinda Gupta (Sequin AI) (50:12)
But when it does get good, which I think it will soon, it's pretty incredible what it can do. Think about from a collection standpoint, think about some of the recs. And if you are trying to make sure you're not harassing someone, a lot of the time that person might perceive themself to be harassed. And the words that they're saying might not really be indicating how they're feeling. They might be saying something sarcastic. There might be something in their tone.

Vrinda Gupta (Sequin AI) (50:40)
So when you actually have the speech, exactly. Yeah, part of the context is the tone, right? If you say something sarcastic, it's not just looking at the keywords. And so today, most of what's happening is it's looking at the keywords and then it's assessing risk, right? Based off of what's being said. But in the future and shortly here, we'll be able to get the speech-to-speech technology good enough where it can really understand the tone and context and make sure you're not violating any of the recs, which I think is...

Adam Parks (50:41)
The context.

Vrinda Gupta (Sequin AI) (51:08)
quite fascinating. So what's going on in terms of how this all works? going back to your original question in terms of how you can decrease latency in conversations, a lot of that comes back to the context windows, which is essentially how much of the conversation it's remembering, right? And so limiting those context windows is helpful, right? So you don't have to go back to, know, a ton of different conversations, but also the way that these are built,

Vrinda Gupta (Sequin AI) (51:38)
you can decide, you can turn up or turn down how quickly you want it to respond. And faster is not better, right? Because faster can mean that you're cutting people off, which is a problem, right? And we've heard that a lot with, you know, some of, even the demos, right? ⁓ Where it just, it's very eager to please. And so it's going to respond very quickly and that can cut someone off, right? And so there's different things you can do in the conversation, but there's also,

Vrinda Gupta (Sequin AI) (52:05)
in a ways that this is actually, conversations are created and prompted that can reduce latency.

Adam Parks (52:10)
When you say that the the models are eager to please, I think it was one of my attorneys who reminded me that ChatGPT is the equivalent of a drunk frat boy. It's going to be wildly confident in everything that it says, regardless of the truth behind its statement. And that I mean, that kind of changed my mind frame a little bit. So I just wanted to throw that one out there. I thought that might make you laugh a little. ⁓ The other thing is when you talk about the

Vrinda Gupta (Sequin AI) (52:21)
Ha ha ha!

my gosh.

I love it.

Adam Parks (52:36)
speech to text and then bringing that text back to speech to me, it reminds me of when my wife and I first met before she spoke English. So in her mind, everything that I said to her in English, she then needed to translate to Portuguese, think about that response, retranslate it back to English and then respond to me again. And I think about it in very much the same way versus being fluent in a language and then being able to listen in English, respond in English, right? Process it in English. And I do think

Vrinda Gupta (Sequin AI) (52:58)
Yes.

Adam Parks (53:05)
beyond even latency, I think that's where the context starts to get better and better. It may not be the increase of the context window, but it's understanding of the other elements of human context. Because in that same communication model with my wife, so much of our communication was hand-based. It was hand signals or motions are pointing to, and how much of our language and communication falls outside of just the words that are being spoken, the tone, the...

Yeah.

Adam Parks (53:34)
pace, the frequency of language, even vocabulary choice. But I think all of those things will start coming together in being able to be processed at not only a faster pace, but a more accurate rating as speech to speech type of technology continues to improve in the not too distant future. I mean, what are you thinking? Like by 2.30, three o'clock today, we should start seeing some of that tech rolling out?

surprise, know, with GPT-5, it's already a big performance improvement, right? And so, you know, there's just these incremental updates that are happening so quickly. mean, I think jokes aside, like, I don't know by the end of the year, early next year, who knows, right? We thought, yeah, these things are happening so quickly, even with GPT-4, right? We're like, oh, okay, maybe that'll be launched. Then it happens so quickly. It's incredible.

Yeah, look, five coming out. mean, four came out at the beginning of this year. Five is already out. I mean, we saw 4.5 for deep research. We're seeing these other models that are coming out, but that constant evolving of the technology. mean, really, the only exponential curve I've ever seen like this before is the data storage curve since 2009 with there was 2007 or nine with the advent of the iPhone, right? Like the amount of data that the world creates is exponentially rising year over year.

This year we start rolling out 200 megapixel phones. So how is that going to impact, right? The volume of worthless pictures and videos that we're now storing in the cloud and how much more storage is necessary. I start thinking about this in very much the same way. I realize that those aren't really, you know, tied together. Although I think data storage and processing power and artificial intelligence do all tie together, but that's the only time I've ever seen something that was that exciting, that was that exponential.

Vrinda Gupta (Sequin AI) (55:01)
Yeah, yeah.

Adam Parks (55:19)
and now to not move into artificial intelligence today is like not moving into the internet for your business in the late 90s. It's either, like Barnes and Noble regrets not moving in that direction sooner. I mean, there's a lot of organizations, I think, that regret not making those moves and, know, the future. So as I like to say, the future is now, but looking into your crystal ball and looking into the future, what do you think happens over the next 12 months?

I mean, 12 months ago, I don't know if I could have predicted, you know, the place that we're at, right? And I think, I love what you're talking about with Barnes and Nobles, right? Because I think it's a really great analog to the moment we're in as an industry right now where, you know, it's not some of those companies, it's not even that they were like, you know, we need to do this, it's just a matter of when. Many of them are outright dismissive of the technology, right? And saying, hey, what, we don't need to be on the internet, right? And I think we have a graveyard of companies that felt that way, right? They're like, well, we're gonna win by actually not doing that. And I think the reality is, that's...

Adam Parks (56:10)
Blockbuster.

Vrinda Gupta (Sequin AI) (56:23)
We don't want anyone listening to this to be in that graveyard of companies, right? And so I think it's quite important to really understand and dig in and ask the right questions. But also, Adam, maybe at this point in the conversation, a few red flags we've called out, right, is one, we never hallucinate. There is a 0 % chance that this is going to say something wrong and nothing.

Adam Parks (56:40)
Please.

Vrinda Gupta (Sequin AI) (56:51)
from a compliance standpoint, that can happen to you when you're using AI. If someone said that, would run, I'd end the call and be like, gotta go, right? You've got another call, gotta go. I think the second one is not having that healthy respect for US consumer regulations, and especially when it comes to different asset class by asset class and just saying, this worked.

Adam Parks (56:57)
But...

Vrinda Gupta (Sequin AI) (57:16)
overseas for a different use case, it's gonna work here to believe us, like it was great for their business. I would just be very, very, I would be very cautious of, you know, anyone that's just saying, yes, this is gonna solve all your problems. Right, you want someone with that healthy sense of skepticism to understand where the technology is and what its limitations are. And I think the third piece is just having a respect for your specific business as well.

Vrinda Gupta (Sequin AI) (57:41)
understanding that the technology needs to adapt to your business. So anyone that's saying this is ready to go out the box and it's not going to require anything from you, I would be pretty skeptical of that as well because you do need to and want to be involved, right? And making some of the decisions of how exactly are we getting the insights from this and training it back in. This is going to be an ongoing process that you want to be involved in. it's not...

Vrinda Gupta (Sequin AI) (58:08)
you know, one and done. And I think that's why a lot of creditors we talked to are interested in working with vendors like us, right? Where it's, you know, this is going to be consistently evolving in terms of the learnings, in terms of the regulatory landscape, in terms of what AI can do. And so you can't just say, okay, this is six month project. We've put our engineers on this and now hands off and this is going to work forever, right? It's not, and it shouldn't, right? Cause you're going to be outdated.

I think you need to look for those partners that are on that growth curve, meaning that the partners that are going to not sell you a product today, but are going to work with you to evolve that product over time, because that's where that's what we're going to need. Because like you said, we can't we can't even predict what's going to happen in the next 12 months. I mean, debt collection technology has been rather predictable for the last 25 years. Now, all of a sudden, everything has become unpredictable. We don't really know what's going to roll out next, but

Vrinda Gupta (Sequin AI) (58:46)
Yes.

Adam Parks (59:02)
we tend not to be first movers in terms of deploying technology. And I think Reg F opened the window a little bit here and hopefully we'll see a door get opened at some point so we can really start moving into new technology. But when you look at the rule sets that are built for our industry, they were written in the 1970s. Yes, we got a fairly recent update, but even when that update in Reg F rolled out in 2022,

That also did not contemplate the use of artificial intelligence in a variety of ways. And I think a lot of what we're talking about deploying today wasn't even a pipe dream three years ago. And so how are they supposed to foresee that into the future and what that regulation is going to look like? So I think it's important as an industry that we look at regulating and managing ourselves, putting those judge LLMs and things in place to help us understand.

the massive volume of communications because the whole world is not going to live in a template forever. And as we start moving out of those templates, as we start moving out of those conversational IVRs and turning ourselves into new and interesting technology, I think that's what the consumer is looking for.

Vrinda Gupta (Sequin AI) (59:57)
Yes.

Adam Parks (1:00:06)
especially over the next 10 years, that evolution of comfort level, just look at the comfort level that consumers have found over the last 10 years in using subscription models versus paying or buying for, you buying things outright. I don't even think you can walk into AT&T and buy a phone anymore. You have to put it on the installment plan and then you can pay it off at the end of the first month, which is the weirdest thing I've ever heard in my life. But it talks to me about the comfort level that these consumers are having.

Adam Parks (1:00:34)
that preference set has only been over the last five to 10 years. And it really started off with Adobe, Netflix, and a few others that really grew that subscription base. Now it's the foundation of your general consumer's economic life cycle. So what is that gonna look like? And what kind of changes might we expect to start seeing in terms of the comfort levels of engaging and interacting with artificial intelligence?

These are really big questions, Adam. I think ultimately, looking at collections, you have to look at the lenders and the originations and the different products that are out there. And it's quite exciting what's been going on over the last few decades in FinTech. And there are these alternative lending models. And most of the MPLs today aren't reporting to credit euros. So what does that actually mean when it comes to collections and charge-offs?

I get really excited about the different innovations and ways to access capital, but that also means that everyone in the industry is kind of trying to understand how these products work, right? Whether that is the originators themselves being like, hey, what's our risk model be around this? And they're constantly tweaking that, right?

whether it's the credit bureaus who are looking at this and saying, does this specific construct, does that work? Whether it's the network saying, hey, does this work within our existing structures? So I think a lot is changing, which is exciting. And even looking at Gen Z lending and borrowing behaviors, how they're looking at credit, how they're looking at BNPL, all of these different asset classes really mirror the way that our society and especially geography by geography is moving. So I think it's quite interesting. And I think that's why on the collection side, really understanding, what was the motivation around the originations of this type of product? It's also going to differentiate in terms of what your collection strategy is. So I think there's a lot happening in this space, which is very interesting from an innovation standpoint.

Vrinda Gupta (Sequin AI) (1:02:34)
from a regulation standpoint, from technology companies like ours, from that standpoint, enabling different things as well. So I think it's an especially exciting time, but coming from an industry where, as you mentioned, things have been pretty consistent over the last few decades, this is an inflection point. And I'd be curious, Adam, you're so well connected, everyone loves you. What is the perception right now in the marketplace? Are people...

Vrinda Gupta (Sequin AI) (1:03:02)
excited about that change? Are they scared? Like how are people feeling right now, especially folks who have been in an industry that hasn't, you know, needed much change over the last few decades?

Adam Parks (1:03:11)
So that's

an interesting question. In 2023, the TransUnion survey said that I think it was only 24 % of organizations were even going to explore artificial intelligence. Like they were dead set on it. And then it increased pretty significantly. And I want to say it was around 48 % last year said that they were going to start deploying artificial intelligence or they were looking at it.

I was just peeking at the numbers earlier today for the active survey and already starting to see that increase happen. So I think the comfort level is increasing. The biggest challenge that organizations have is have you ever met an IT department that had like additional resources available to take on a new project? No, of course you haven't because there's immediately an accountant there to cut that team in half if they feel like they're overfunded. So I think that's a big part of the challenge is actually being able to

Vrinda Gupta (Sequin AI) (1:03:51)
Never.

Adam Parks (1:04:03)
operationalize and connect to these different systems. Now, as the platforms have become more modular and API driven, I think those opportunities have opened up to a point in which it's a far more realistic opportunity. I think organizations are seeing that. I think they're moving away from the closed systems in order to favor those open API platforms because like we say, the AI race is just beginning. And so if my...

Vrinda Gupta (Sequin AI) (1:04:27)
yeah.

Adam Parks (1:04:28)
software provider is going to tell me that they're the best, you know, they've got the best chatbot, and they've got the best, the best workflow tool and the best this, there is no way that you are the best at every use case of artificial intelligence in this space. That's not realistic, because then I look at organizations like yours that are laser focused on doing one thing really, really well. And they tend to do that one thing really, really well. How well could you do what you're doing now if you also were attacking five other use cases simultaneously and trying to manage the backend database that connects everything. To me, it's about that level of flexibility for the future because we don't know who's gonna win the AI race at this point. We just started.

I feel like there's massive opportunity right now to start deploying these things. I think the more comfortable the consumers are with it and the more ubiquitous the type of technology becomes in non-regulated markets, the faster their comfort level will start to arrive. But I think what's been driving the industry towards these types of technology solutions is an increase in the volume of accounts and a decrease in the liquidity of that same underlying consumer.

And whenever that happens, we have to do more with less. We saw it in 2008, we saw it in 2012, we saw it, you we've seen it through multiple cycles now. And as we're watching that change, the more comfortable the consumer gets with this, the easier everything is going to become. And if we're gonna have to do more with less, I think using smart tools is the way to go. I recently did an interview with a gentleman who has

Vrinda Gupta (Sequin AI) (1:05:39)
Yes.

Adam Parks (1:06:02)
deployed a lot of digital technology into his collection agency. And in response to that, rather than letting people go, he's moved them into first party servicing. And now they're handling a higher volume of accounts, but they're able to put live people on those things that require live people. And their focal point has been on that strategic redirection of making sure that the right person is talking to the right person and that that's driving the...

Vrinda Gupta (Sequin AI) (1:06:19)
Yes.

Adam Parks (1:06:28)
the value for not only the end user, but for the client as well. And because a collection agency is an intermediary between two parties. So how can I simplify that intermediary process as much as possible? How can I deploy my my client level requirements? How can I deploy my portfolio level requirements and make sure that those are being managed appropriately while at the same time meeting the needs of the consumer.

Vrinda Gupta (Sequin AI) (1:06:51)
left this response, right? Because I think, you know, lot of some of the questions I get spicier questions, let's say I get is, Hey, is your goal to eliminate all humans from this process, right? And actually, it's a, you know, maybe that's in the spirit of questions to ask, you know, your potential vendors that you're evaluating, you know, what is their opinion on that, right? I would say personally, I don't think that is the goal, nor do I think it should be the goal, right? And I think to your point, Adam, around reallocating resources to higher value activities, that is the goal here. And I also think there are going to be situations where you really do meet a human on the line, right? If someone's really upset and they're threatening to report you to the CFPB or get their lawyer involved or whatever else, right? It's like, hey, someone should be there to answer that question.

Adam Parks (1:07:24)
Yeah.

Vrinda Gupta (Sequin AI) (1:07:40)
There's no way that an AI is going to ever be able to get a hundred percent of the cases answered because those it's constantly evolving. Right. And so, you know, even if you are able to handle all of those one day, the next day, something might change, whether that's in your product and your procedures or whether that's more in terms of a macro environment, you know, trend, right.

Vrinda Gupta (Sequin AI) (1:08:05)
just are going to be changes as well. so I think, you know, humans are always going to be involved and you humans are the backbone of this business, right? It just more, how are you being intentional with your resources and how are you putting those humans towards, you know, the highest value actions versus others that an AI actually, you know, at the end of the day, like if you are, you know, making collections calls all day, like that is a very hard job. Like it's really hard. And I feel like,

Vrinda Gupta (Sequin AI) (1:08:32)
I'm a pretty nice, patient person, but I don't know. At the end of the day, I might not be my best self in that situation. so there might be just cases where an AI is able to extend the empathy and not really internalize as much. And that's just human versus an AI.

Adam Parks (1:08:48)
I think it also allows the collection agency to put their best foot forward from a staffing perspective. From my understanding, as the individual that I was interviewing kind of moved resources over to first party servicing, a lot of those resources worked hard to try and go back to the third party collections because of the commissions available and the variable compensation that's available in that realm that might not be available on the servicing side.

you

Adam Parks (1:09:15)
as artificial intelligence continues to change the face of our industry over the next six months, Rinda, I'm gonna have to have you come back so that we can continue this conversation and keep our finger on the pulse of not only what's happening in the space, but what's happening with the technology and how that technology can be realistically deployed at this point and how those use cases are going to be modified and changed over the next 12 months.

Vrinda Gupta (Sequin AI) (1:09:26)
Yes. ⁓

so excited to have this conversation again in six months because if we had this conversation even a few days ago, GPT-5 wouldn't have come out, right? So it's, I mean, it's just so exciting the pace of all this stuff. And I do want to, you know, empathize with folks who are trying to evaluate these technologies where it is new, right? And I think maybe, you know, a piece to hammer home here is the answer is not to be afraid and to avoid, right? That is, one way and that's what we're talking about the

Adam Parks (1:09:49)
Yeah

Vrinda Gupta (Sequin AI) (1:10:11)
the graveyard here, right? And what we want is just to be curious and to understand. And I think to take conversations with companies like ours who are building in the space and ask the hard questions, right? And part of how you're evaluating is their ability to explain and tell you, okay, this is actually what to be afraid of is what isn't. And for the things that are scary, how do we mitigate that risk? And really understanding like your existing processes are putting you at risk as well, right?

Vrinda Gupta (Sequin AI) (1:10:41)
And so, you know, this is going to be less than that. But also, yeah, how do we put safeguards around it? So I think just making sure, especially as it's new technology, the teams that you're working with are going to be very critical and helping you stay up to date as to what is the latest and also help you kind of look over your shoulder to make sure that, you know, you're not exposing yourself to any risks that you may not even be aware of.

Adam Parks (1:11:07)
I think.

You've provided a lot of great insights in this discussion. For those of you that are watching, if you have additional questions you'd like to ask Vrinda 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 dig into, you can leave those in the comments below as well. And I'm pretty sure I'm to be able to get her back at least one more time to help me continue to create great content for a great industry. But Vrinda, I really do appreciate all of the insights that you've shared today. This has been a lot of fun and even planning. I feel like we laughed and giggled through most of the planning call.

because there's just so much content for us to cover here in so many great different areas that I think we can further inform the debt collection industry.

Vrinda Gupta (Sequin AI) (1:11:45)
Thanks so much for this platform, Adam. This has been a blast. I really appreciate it.

Adam Parks (1:11:50)
Absolutely. I really do appreciate you joining me today and thank you everybody for watching. We really appreciate your time and attention. We'll see you all again soon. Bye.