WEBVTT
00:00:06.559 --> 00:00:07.519
Hello, everybody.
00:00:07.599 --> 00:00:11.919
Adam Parks here with another episode of Receivables Podcast.
00:00:12.080 --> 00:00:49.600
Today I'm here with a great guest and a great topic here with Nir Laznick from Cedric, here to talk to me about well, keeping an eye on the bots, keeping the guardrails around those uh let's call it shiny new voice AI bots that everybody's talking about and how to build that governance guardrail framework around those tools because I still struggle with understanding why we would ever watch our phone calls differently, whether they were generated by a human or they were being generated through an AI technology.
00:00:49.759 --> 00:00:53.520
So, Nir, thank you so much for coming on, sharing your insights with me today.
00:00:53.679 --> 00:00:55.280
I look forward to learning from you.
00:00:55.759 --> 00:00:57.280
Thank you for hosting me, Adam.
00:00:57.439 --> 00:00:59.920
It's been a while and we've been trying to plan it.
00:01:00.000 --> 00:01:02.000
So I'm excited to already have it.
00:01:02.719 --> 00:01:04.560
Well, I appreciate you joining me today.
00:01:04.799 --> 00:01:13.040
Nir, for anyone who has not been as lucky as me to get to know you through the years, can you tell everyone a little about yourself and how you got to the seat that you're in today?
00:01:13.439 --> 00:01:17.040
Yes, uh, I always like to start with a personal introduction.
00:01:17.359 --> 00:01:22.000
So, Nier, I'm a father, always comes first.
00:01:22.400 --> 00:01:26.640
After that, after my daddy's dead, I'm an information system engineer.
00:01:26.799 --> 00:01:40.319
And a few years ago, myself and my talented co-founder, which is one of the leading data scientists back in Tel Aviv, identified a few macro trends that are now taking a very big part in the story of what we do.
00:01:40.560 --> 00:01:56.400
And the first one was the fragmentation of the financial services sector, which basically got us to the conclusion that financial services, in specific uh consumer finance, are coming in different shapes and forms.
00:01:56.640 --> 00:02:06.239
Alongside the traditional banking, we now have fintechs, we have embedded finance that creates like different shapes and forms of those consumer products.
00:02:06.480 --> 00:02:22.159
The other one was the increasing complexity around compliance, and together we realized that there is a disproportional pressure being made on compliance that was already slowing down business, and we wanted to unlock and to turn compliance into a business enabler.
00:02:22.319 --> 00:02:27.280
So, with that conclusion in mind, we accepted the first AI compliance officer.
00:02:27.439 --> 00:02:42.319
We called our Cedric, and we are here really to enable financial institutions to grow faster without breaking things, and on the collection side is really to unlock productivity without the burden and the risk that of the compliance that is involved.
00:02:42.960 --> 00:02:47.520
Well, prior to these kinds of tools, we maybe listened to 2% of the calls.
00:02:47.759 --> 00:02:59.680
And you know, that was the sample size that we were able to get to using live folks to listen to those calls, and through the deployment of artificial intelligence, we've seen that increase to 100%.
00:03:00.080 --> 00:03:14.960
And now the 2% that's being reviewed is the 2% of exceptions that were found systematically, which has changed the face of compliance for the debt collection and for the let's call call center customer service environments as well.
00:03:15.120 --> 00:03:19.520
And you've been on the leading edge of creating that technology through the years.
00:03:19.599 --> 00:03:24.240
It's been interesting to see how this has, you know, started to play out.
00:03:24.800 --> 00:03:42.800
Now, where you guys are even beyond the debt collection space and looking at the creditor space as well, where the debt is actually originated, what are some of the biggest concerns that these organizations have related to the deployment of artificial intelligence in general?
00:03:43.759 --> 00:03:45.039
It's a great question.
00:03:45.199 --> 00:03:57.759
And uh I think that um if we really observe on what's happening right now in the market, we are currently in the middle of the second phase of adoption of AI.
00:03:58.159 --> 00:04:10.639
The first one included a lot of pressure coming from board members to introduce AI solutions to increase productivity and to basically leverage those capabilities.
00:04:10.960 --> 00:04:16.879
The problem with that first adoption cycle is that the technology was not mature enough.
00:04:17.199 --> 00:04:23.519
We saw a very large number of POCs with a very high, above 90% failure.
00:04:23.680 --> 00:04:39.759
Now, what we're seeing in the second phase of adoption, which we're in the middle right now, is that the technology actually matured into a posture that it can actually deliver production large scale in different areas of the business.
00:04:40.000 --> 00:04:47.920
But the problem that we sometimes see, especially in the more conservative institutions, is that it's being blocked still.
00:04:48.240 --> 00:04:51.759
And in many cases, that would be by compliance.
00:04:51.920 --> 00:05:00.800
And one of the reasons for that being that way is that compliance needs to make sure that they have the right control on those solutions.
00:05:00.959 --> 00:05:12.399
They do want to support the business, but if they don't know enough and if they have blind spots, they rather just stop rather than taking the risk if they cannot quantify it.
00:05:12.639 --> 00:05:38.319
One of the things that we are trying to do is to really rethink compliance as the enabler of the business by enabling all of that supervision and oversight across all of the different channels, including some of those innovative ones, like the AI solutions, that's really matured into a stage where they can serve the different operations, even the conservative ones.
00:05:39.600 --> 00:05:46.959
So as you think about the type of technology that you're deploying for those organizations and the things that they're trying to accomplish now.
00:05:47.120 --> 00:06:00.720
So you've you've mastered your craft on those live human calls, and now these financial institutions are pushing and saying, we want to start using bots because we think it's going to service consumers in this way or that way.
00:06:02.399 --> 00:06:19.519
How has the have you had some of those conversations with them around how your tool set creates guardrails not just for those live callers, but also for let's call it the new shiny object in the room, the voice AI outbound call.
00:06:20.560 --> 00:06:21.360
Of course.
00:06:21.519 --> 00:06:27.279
And maybe we can travel back in time, probably two or three years together.
00:06:27.519 --> 00:06:39.759
And Cedric was in a crossroad when we realized that from a technology perspective, the gap between what we do right now to actually enable those as a first line is actually not that big.
00:06:40.000 --> 00:06:47.360
But the reason why we've decided to stay on the compliance side is that when you think about risk and compliance, it has to be a separate line of defense.
00:06:47.600 --> 00:06:57.600
Just like a very experienced collector cannot claim that they know everything and they don't need to be monitored because they've been doing it for a while.
00:06:57.759 --> 00:07:06.000
In a similar fashion, even the most compliance-oriented AI solution cannot just evaluate itself.
00:07:06.079 --> 00:07:14.480
You need to have a separate line of defense that can orchestrate the different channels, including those Gen AI related ones.
00:07:14.800 --> 00:07:40.800
So, what Cedric provides is an orchestration layer where you can basically take all of your policies that includes all the different complexities, jurisdiction considerations, different product nuances, and so forth, and to basically codify them into the control system that is basically serving as the infrastructure for any future agency workflow that you want to run on top of it.
00:07:40.959 --> 00:07:48.399
So that can be listening to voice calls, that can be listening to chat threads or email threads and so forth.
00:07:48.639 --> 00:07:55.360
And in other verticals, we actually support different mediums as well, which are less relevant to the count recovery space.
00:07:55.439 --> 00:08:08.000
But this is part of the stack of enabling financial institutions or a regulated ones to basically follow detailed, complex policies in different mediums where they interact with their consumers.
00:08:08.560 --> 00:08:21.600
So you've actually pivoted slightly from being a let's say call monitoring or call compliance quality type operation to looking at all channels, both written and spoken.
00:08:22.399 --> 00:08:24.959
Yes, we extended the offering.
00:08:26.560 --> 00:08:45.600
And we always like to take liked to take a very broad approach about what we do, like thinking about the entire workflow in specific verticals, so we can really represent all the nuances and the complexities of this vertical and really deeply understand the persona that we're serving.
00:08:45.759 --> 00:08:51.440
In that case, that's the compliance and the risk, and to really help them to take over the entire workflow.
00:08:51.600 --> 00:09:03.360
And while phone calls used to be the channel that took over most of the volume, we now see omnichannel and the number of channels that is only increasing to more and more.
00:09:03.519 --> 00:09:07.759
So that was a natural expansion for the product offering, basically.
00:09:08.159 --> 00:09:12.080
What makes me think about how people are stacking bots into workflows now?
00:09:12.240 --> 00:09:18.879
And rather than trying to be everything to everybody, you said, let me just be the master of this piece.
00:09:19.039 --> 00:09:24.720
But that piece carries across all touch points with the consumer in all channels.
00:09:24.960 --> 00:09:28.639
Because if you're doing the written, the email, the text messages, right?
00:09:28.720 --> 00:09:31.519
God knows what kind of communication we're going to come up with in the future.
00:09:31.679 --> 00:09:44.320
But you know, looking at how we've focused over the last couple of years, it seems like you guys have kind of well positioned yourselves to be that check and balance system against those communications.
00:09:45.039 --> 00:09:57.919
Have you started looking, and I'm sure you're doing it with let's call it the pre-fixed consumer journeys, but how does that technology play into uh let's say on the fly generated content to consumers?
00:09:58.080 --> 00:10:03.440
Is it watching it the same way that it was watching the live human, or how does that change over time?
00:10:04.080 --> 00:10:05.519
It's a great question.
00:10:05.679 --> 00:10:09.679
Uh, you did mention a comment that was spot on at the beginning.
00:10:10.080 --> 00:10:16.399
And basically, this is my philosophy about venture building, right?
00:10:16.559 --> 00:10:23.759
Especially especially in an era where the tech mode cannot be the only mode that you're having for your business.
00:10:24.480 --> 00:10:37.759
And you know, in the past you had like many horizontal solutions that served well many organizations, where today it's actually getting a lot of challenge in the build versus buy discussion that you have.
00:10:38.000 --> 00:10:52.320
But when you go deep in a specific vertical, in a complex problem that requires you to have domain expertise alongside your tech acumen as a firm, this is where you create a lot of value that's hard to replicate.
00:10:52.559 --> 00:11:07.519
And when we get partners on board, they're not only subscribing to the solution, they are subscribing to the roadmap, they are subscribing to the talent, they are subscribing to their potential to stay at the edge of technology every time.
00:11:08.159 --> 00:11:10.240
Um so yeah.
00:11:11.440 --> 00:11:12.480
No, it's interesting.
00:11:12.559 --> 00:11:15.120
Uh, I I like the approach to it.
00:11:15.279 --> 00:11:22.399
Do you think the primary fear of these financial institutions is the regulatory risk, the brand risk?
00:11:22.559 --> 00:11:31.919
Is it operational or is everything just blended together into one layer of anxiety for these compliance professionals?
00:11:33.360 --> 00:11:54.559
It's a great question, and it's actually an interesting point in time to ask it because like we all understand that the climate around the what is risk, how regulation is being enforced these days with the attempt dismantle of the CFPB, and um maybe the only constant which is change right now in that landscape.
00:11:54.960 --> 00:12:06.080
But um, yeah, I think that look, we need to remember that risk and being conservative is rooted in the DNA of some of those firms.
00:12:06.320 --> 00:12:13.279
And even if the macro changes right now, everyone understands that this is a string, right?
00:12:13.679 --> 00:12:16.320
It's it can be flexed, but it can be loosened.
00:12:16.399 --> 00:12:18.879
Nobody thinks that regulation is gone forever.
00:12:19.039 --> 00:12:20.799
We understand that it's changing.
00:12:20.960 --> 00:12:23.759
We also understand that it's not the same.
00:12:24.000 --> 00:12:34.559
We have the different attorney generals that are behaving slightly differently, which actually add a lot another layer of complexity for some jurisdictions, for some states.
00:12:34.799 --> 00:12:44.960
So one thing is clear: you do need to have a more flexible system that you can orchestrate either to loosen or to stress in your controls.
00:12:45.279 --> 00:12:51.039
So this is something that Cedric provides, and not just Cedric, but other platforms as well.
00:12:51.279 --> 00:13:00.320
And uh I think that it's a hard question because it is rooted in the DNA of those organizations.
00:13:00.879 --> 00:13:18.480
Some would take those opportunities to maybe loosen some of the controls, some would understand that the change management has so much friction and potential risk that you don't know what would be the perspective of the regulator for whatever control change management that you went through throughout that time.
00:13:19.200 --> 00:13:22.080
I think each one of those comes in ebbs and flows.
00:13:22.399 --> 00:13:38.000
So when we think about regulatory risk, you know, I think who's in power in the US government and how powerful is the CFPB at that moment in time, who's the director, all plays into or all signals that we look at to measure the level of regulatory risk.
00:13:38.240 --> 00:13:53.519
At the same time, I think we've got privacy concerns and patchworks of information, patchworks of requirements that are starting to bloom around the country that are requiring us to take new steps that haven't been taken before, right?
00:13:53.600 --> 00:14:06.639
To organize ourselves in different ways and being on the forefront of a lot of that uh requirement and staying, let's say, staying on that wave is a full-time job in and of itself.
00:14:06.879 --> 00:14:36.960
As a reformed uh compliance professional myself, you know, I know how much time, effort, and energy went into maintaining that level of sophistication flexibility to be able to deploy the different solutions as these requirements changed, and the and really the the reactionary time periods to be able to respond to these changes in a reasonable time period, because you're not always given a reasonable time period um when something new is being deployed.
00:14:37.120 --> 00:14:50.480
Um, so we start thinking about how the technology has changed over time, and now we're starting to talk about not just getting a human on the phone, but getting this bot on the phone.
00:14:50.639 --> 00:15:01.519
And it reminds me of the movie casino where you know the pit boss is watching the dealer, and the you know, the floor manager is watching the pit boss, and the eye in the sky is watching us all.
00:15:01.679 --> 00:15:10.639
And I think it's a great analogy for what you guys have been able to do, kind of circling up all of these solutions together.
00:15:10.879 --> 00:15:17.279
And you know, what has been your experience as you've been able to put some guardrails in place?
00:15:17.440 --> 00:15:25.200
Is that helping to enable those financial institutions to take that calculated risk and try something new?
00:15:25.360 --> 00:15:34.399
Or are they still maintaining kind of those tight reins and not venturing into the new technology that's available?
00:15:35.360 --> 00:15:42.879
Look, what I can only testify from uh you know our portfolio, uh I believe it does represent a trend.
00:15:43.519 --> 00:15:50.639
And we are actually for us, it's a backwind, right?
00:15:50.879 --> 00:16:02.240
Because, like I mentioned before, we are in the middle of the second adoption cycle where the technology matured into a state where it can actually deliver, not only to promise.
00:16:02.720 --> 00:16:16.799
And one of the things that we're seeing again is that there is a friction coming from compliance that doesn't feel comfortable to introduce and to roll out some of those programs to larger scales because of the blind spots.
00:16:17.200 --> 00:16:27.919
So we did see how Cedric basically plays a role in enabling those programs moving from a pilot stage into production large-scale programs.
00:16:28.080 --> 00:16:31.360
So, to your question, the answer is definitely yes.
00:16:32.240 --> 00:16:57.039
So, from your perspective, uh, when organizations are rolling out a new piece of AI technology, are they generally talking to the financial institution and saying, well, look, you know, we our our bot is watching our other bot, or how comfortable do they become as a service provider, knowing that there's an overwatch of their technology from another compliance tool set?
00:16:57.200 --> 00:17:04.559
What did what do those relationships start to look like and does it enable them to actively forward?
00:17:05.119 --> 00:17:18.319
Are you asking about the relationship between the creditor and the so from from your perspective, when you roll out, so uh uh a financial institution is going to roll out this new tool set?
00:17:18.480 --> 00:17:26.720
It's been my experience that a lot of organizations are saying, hey, I have this AI bot and I've got this judge LLM that's gonna watch the bot, right?
00:17:26.799 --> 00:17:35.680
But like you had said at the beginning, that generally is falling somewhat within the same environment and not necessarily a third-party check and balance system.
00:17:36.000 --> 00:17:45.359
What has that relationship been like for you guys coming in as kind of a compliance umbrella that's sitting over not only their voice bot but their judge LLM?
00:17:45.440 --> 00:17:49.119
And what how has that worked in reality?
00:17:49.279 --> 00:17:52.640
Like I could see it conceptually, but like what's the reality of that situation?
00:17:52.880 --> 00:17:58.559
Yeah, so let's map out the different players in that in that uh ecosystem that you just uh outlined.
00:17:58.799 --> 00:18:02.240
So we would have the AICX solution.
00:18:02.559 --> 00:18:06.880
CERUC would be an enabler for a production rollout for them.
00:18:07.279 --> 00:18:19.680
We have the agency that would potentially like to get better allocation, and by demonstrating better compliance posture in high confidence, that would help them to do so.
00:18:19.920 --> 00:18:32.319
For the creditor, it would give them the comfort that a solution that is trained to do that specific job, not a generic AI solution that is also doing compliance.
00:18:32.480 --> 00:18:43.039
This is not something that you would just trust a generic solution, just like you probably won't trust a generic LLM to review the contract before you buy a house, right?
00:18:43.119 --> 00:18:52.559
You want to make sure that it's nailed and that you're not missing a bit because you are basically jeopardizing your firm by going after a generic solution.
00:18:52.720 --> 00:18:54.559
So you want to feel confident.
00:18:54.880 --> 00:19:07.680
So, yeah, if we think about the voice AI solution, this is an enabler for them, for the agency that would give them more confidence and hopefully get them better business with the creditor.
00:19:08.000 --> 00:19:14.319
And for the creditor, that reduces the exposure for them from both regulatory but also brand.
00:19:14.480 --> 00:19:26.400
As you climb up the ladder, the larger creditors sometimes cares way more about brand and compliance rather than the collection rate, and they want to make sure that they are protected.
00:19:27.359 --> 00:19:28.000
Sure, right?
00:19:28.319 --> 00:19:37.759
Debt collection is an afterthought for an originating creditor or a major financial institution, they're making their money on lending money, and this is a byproduct of it.
00:19:38.079 --> 00:19:51.599
But you do get me thinking here as I as I start working my way through, you know, is there a fundamental difference between watching an AI bot and monitoring its activity versus monitoring a live collector?
00:19:51.680 --> 00:19:59.759
Is there anything fundamentally different between those two scenarios, or are we basically taking treating everything as equal?
00:20:00.960 --> 00:20:03.680
Yeah, it's not it's not the same for sure.
00:20:03.920 --> 00:20:15.599
First off, if done right, and it's not always the case, you do have a good sense of control on the Gen AI side, right?
00:20:16.160 --> 00:20:26.160
There are really good firms that are doing the right thing, they are educated about compliance, and the initial output would generally be good, right?
00:20:26.559 --> 00:20:30.000
The other side, which is the consumer, you don't get to control.
00:20:30.480 --> 00:20:44.079
Sometimes it can drag the conversation to edge cases that were not trained, or just impossible to be trained at the technology stage right now, and this is where the risk would potentially surface.
00:20:44.400 --> 00:20:51.680
So basically, those nuances, those edge cases, and this is where it's going to be treated differently.
00:20:52.000 --> 00:21:02.319
The amount of control that you're gonna have on one side versus the other would also be a difference between a human to human versus an AI agent to a human.
00:21:02.480 --> 00:21:09.119
And there are other considerations that are more technical, but high-level speaking, there is a difference.
00:21:10.799 --> 00:21:16.079
So as we think about the differences there, it's you know, from the conversations that I've been having.
00:21:17.359 --> 00:21:31.359
For those of us that use Chat GPT, Claude, right, like the general kind of use case, LLMs across the space, we've all been subject to and seen hallucinations, degradation of a session, right?
00:21:31.440 --> 00:21:41.119
The canary in the coal mine, there's all different kinds of ways that those of us that engineer on a regular basis are kind of building those things out and trying to understand that degregation.
00:21:41.440 --> 00:22:11.359
From your perspective, as some as as having Overwatch for all of these different types of technologies, is that a real risk in the AI voice space, or have we come far enough over the last, let's say, two years or so to where the hallucinations are are really Minimal to almost non-existent in that voice space, or are we still seeing slips of the models versus ironing out of the uh campaigns and programs?
00:22:13.119 --> 00:22:17.039
There is no straight answer to that, but there is a trade-off, right?
00:22:17.200 --> 00:22:25.759
Okay, if you would narrow the degrees of freedom of your solution, you would potentially have less hallucinations.
00:22:26.720 --> 00:22:35.519
But that the trade-off would be that they won't be able to cover some of the edge cases, and then you're gonna have to escalate maybe a higher percentage to human.
00:22:35.920 --> 00:22:44.079
On the other side, you can increase the degrees of freedom, and you cannot completely um avoid having hallucinations.
00:22:44.240 --> 00:22:57.839
There are a lot of best practices, and I would leave those best practices to the ones that are implying them for the Gen AI solutions that can really improve and get this technology again to a very high maturity.
00:22:57.920 --> 00:23:04.880
And this is why we do see large-scale adoptions taking place right now for AICX solutions.
00:23:05.039 --> 00:23:13.359
Like the tech is there, they they were able to like reduce the amount of hallucinations to a minimal amount.
00:23:14.640 --> 00:23:28.160
So, as we think about monitoring calls and solutions for the financial institutions, how much are they monitoring their downstream placement?
00:23:28.319 --> 00:23:43.519
So the financial institution either sells the accounts, they place the accounts with a collection agency, a law firm, whatever the case or strategy may be, how much of that oversight is being carried down that stream over time?
00:23:43.759 --> 00:23:59.680
And it was uh it was a work it was an individual from a major bank who recently mentioned to me uh at one of the conferences that the better that they got with AI, the more difficult the debt collector's life was going to become because the low-hanging fruit wasn't going to be there.
00:23:59.839 --> 00:24:05.200
They were gonna work out those solutions, they were gonna be able to kind of resolve for that.
00:24:05.519 --> 00:24:18.079
But as you're looking at it from a monitoring perspective, and you're thinking about as these accounts start to flow into the collections lifecycle, how much of that monitoring continues to flow downstream?
00:24:18.160 --> 00:24:24.000
And what are those what are the expectations being set by those financial institutions today?
00:24:24.480 --> 00:24:26.720
Look, I think that the bar is just higher.
00:24:26.960 --> 00:24:36.400
I mean, if you could avoid um patterns and the cycle from event to remediation used to take months, sometimes quarters.
00:24:36.640 --> 00:24:41.599
Now the expectation is to adopt and to close the feedback loop much faster.
00:24:41.839 --> 00:24:53.359
I can tell you that the similar trend of um the sample-based to 100% and exception-based based on risk trend that you had mentioned before.
00:24:53.599 --> 00:24:56.000
This is the exact same trend that we see.
00:24:56.079 --> 00:25:01.119
Like I can refer to one of our maybe larger banks that we're serving.
00:25:01.359 --> 00:25:09.920
We basically held them to oversee 100% of the communication of all of the outsource agencies that are serving them.
00:25:10.319 --> 00:25:10.799
Right?
00:25:11.039 --> 00:25:12.559
It wasn't the case.
00:25:12.799 --> 00:25:16.400
It used to be probably less than 1% in the past.
00:25:17.279 --> 00:25:22.400
It used to be submit these 12 call recordings, and now you're saying it's 100%.
00:25:22.960 --> 00:25:26.400
Is that a direct connection or are they kicking over those calls every month?
00:25:26.480 --> 00:25:30.000
Or now you're just directly connected into that process from a monitoring?
00:25:31.119 --> 00:25:33.759
It would vary between one agency to another.
00:25:33.920 --> 00:25:43.920
Like one agency can be a direct integration into the dialers, another can be just an SFTP that they're gonna push uh the data towards Cedric, and we're gonna analyze that.
00:25:44.559 --> 00:25:45.119
Interesting.
00:25:45.279 --> 00:26:11.920
So, regardless of the uh let's say the technology connection between that creditor and that service provider, you've got a mechanism to hook into that for that monitoring, which I would think from a uh even from an agency perspective would be preferential because now they don't have to manage this process so much as just kind of open the doors and allow you to see you know what you need to see, which which makes sense.
00:26:12.160 --> 00:26:15.839
Um true, although it's like redefine the relationship, right?
00:26:16.400 --> 00:26:16.720
Sure.
00:26:17.680 --> 00:26:18.799
To some extent.
00:26:18.960 --> 00:26:22.480
I mean, you are very exposed, transparent.
00:26:22.880 --> 00:26:40.079
We always believed in the films that we're serving and and their intention, but uh it's slightly harder when everything is flowing that fast and the expectation to fix is also uh basically being redefined as part of this uh change in the process.
00:26:40.559 --> 00:26:45.359
Well, reaction and response times change as information moves faster, right?
00:26:45.440 --> 00:27:10.240
If if I'm sending you 12 calls and you're reviewing them and sending them back later this month, and now I can, you know, instigate or I can begin my retraining of whatever individual and kind of work my way through it, but that's like a month-long cycle, and you're saying that the feedback cycle and the expectation to implement is being shrunk, I'm assuming, to hours at this point based on the flow of information and capabilities, if not real time.
00:27:12.319 --> 00:27:27.680
How much of this is is starting to happen from a real-time perspective when we think about call monitoring and and keeping an eye on the agents or the bots, you know, how much of the the compliance is being done as let's say part of a process and how much of it is being done like in the moment.
00:27:28.880 --> 00:27:44.160
So if we think about different types of controls in compliance, and the way we structured it is that you can have preventive, detective, or corrective controls.
00:27:44.480 --> 00:27:59.920
The real time that you're referring to would be considered as preventive control because we would connect into your system of record, into your GRC or the policy instance that we uploaded alongside the live stream of the interaction.
00:28:00.079 --> 00:28:07.759
We will ingest that and fuse it into the model in real time and basically suggest those remediation live.
00:28:08.160 --> 00:28:20.079
But back to our previous example, we didn't see any live integration or expectation from the creditor to have live preventive monitoring on the outsourced agencies.
00:28:20.240 --> 00:28:24.720
This is something that we didn't see to date, although it's technically possible.
00:28:25.039 --> 00:28:31.359
But on the other side, whether it's the agency themselves as a client, we see it a lot.
00:28:31.440 --> 00:28:40.960
And usually that's going to be a fusion of productivity, increased collection rate alongside compliance and not compliance as a standalone.
00:28:41.279 --> 00:28:47.359
Similar trend we see for the creditor when they are doing it for their internal servicing teams, right?
00:28:47.920 --> 00:28:58.480
Whether it's for the performing stage or the delinquent stage, before they're placing it with their first or third party partners, they are using those real-time solutions.
00:28:58.960 --> 00:29:00.559
That are actively using those solutions.
00:29:00.720 --> 00:29:01.519
That's interesting.
00:29:01.759 --> 00:29:09.599
Uh, you know, when we start thinking about how the organizations are coming together and how quickly things are changing right now.
00:29:09.759 --> 00:29:19.039
I mean, even just in the last couple of years, as people have started to deploy artificial intelligence in all kinds of new and interesting places around their organization.
00:29:19.440 --> 00:29:28.079
What's the most surprising use case that you've seen for an organization to deploy a new piece of technology that you've then had to go monitor?
00:29:29.359 --> 00:29:34.079
In general, not any any curveballs that you've received.
00:29:34.160 --> 00:29:39.680
And I mean, I've I've received a hundred curveballs, I feel like, in the last 12 to 18 months.
00:29:39.839 --> 00:29:47.279
But has anybody actually put a piece of technology in front of you to where you had to say, all right, I gotta I gotta think through how I'm actively gonna monitor this?
00:29:47.359 --> 00:29:53.039
Or have you been able to get everything into let's call it the buckets that you're comfortable with?
00:29:54.640 --> 00:29:57.279
Um it's a great question.
00:29:57.440 --> 00:30:02.799
Um let me think if I have a an exotic example to share.
00:30:07.279 --> 00:30:19.119
I think one of the most exotics that uh maybe it's similar crossroads that we had with uh with the the call monitoring piece.
00:30:19.599 --> 00:30:24.720
Cedric also has other other offering related to marketing compliance.
00:30:24.880 --> 00:30:37.680
So we basically take over the entire uh consumer journey starting from the marketing and ending up with servicing after underwriting and then delinquency, collection, and and the rest we know.
00:30:37.920 --> 00:30:49.680
And one of the things that we've been asked recently is instead of just identifying uh what's wrong in the promotion or the ad, can we already fix it?
00:30:50.000 --> 00:31:02.240
Now, theoretically we can, but uh again, are we the one creating the marketing in a compliant way, or are we the one doing the compliance oversight for that uh content creation?
00:31:02.400 --> 00:31:12.319
And we made a similar decision of sticking to our practice of being that separate separate line of defense, basically reflecting the compliance.
00:31:13.039 --> 00:31:17.680
That's a really interesting use case on the marketing side leading into it, right?
00:31:17.839 --> 00:31:27.839
Finding that consumer, getting them in, getting them under written application process, um, that whole thing, and thinking about the compliance guardrails that are necessary there.
00:31:28.000 --> 00:31:34.000
And I guess that I mean it really is similar from a mechanical standpoint to collecting.
00:31:34.160 --> 00:31:35.039
It's not that, right?
00:31:35.119 --> 00:31:39.440
Like loan servicing and collecting from a mechanics perspective is not that different.
00:31:39.680 --> 00:31:44.799
Intent and drive and motivation and all that is is is where the differentiation falls in.
00:31:44.960 --> 00:31:53.359
But that's an interesting use case, and not even that crazy or exotic, but it makes a lot of sense as to why you would move into that space.
00:31:53.519 --> 00:32:00.000
Again, you're already servicing that financial institution, you're already providing that compliance team with that oversight.
00:32:00.160 --> 00:32:05.599
So adding another channel in which you're monitoring, I think makes a lot of sense.
00:32:05.680 --> 00:32:09.839
It's you know, it's economies of scope, it's the same reason McDonald's sells breakfast, right?
00:32:09.920 --> 00:32:15.759
They already had a restaurant and they already had a kitchen and everything else that they needed to be able to do that.
00:32:15.920 --> 00:32:19.200
That's an interesting, it's an interesting approach.
00:32:19.519 --> 00:32:36.480
Now, as you as you look at how you are engaging with those financial institutions, what's the biggest misconception that they have about monitoring their compliance using artificial intelligence?
00:32:37.039 --> 00:32:47.359
Like, what's the biggest challenge that you have to overcome in their mind to get them comfortable with this being the right solution for their financial institution?
00:32:49.200 --> 00:33:00.319
My answer would be consistent between what we see in those larger FIs, smaller FIs, and regulators alike, regulators alike.
00:33:00.799 --> 00:33:14.720
There is a very common bias and misconception about comparing AI to perfection, where suddenly people forget that you need to compare AI to current state.
00:33:14.960 --> 00:33:20.400
And current state in many cases is not that shiny and not that good.
00:33:20.799 --> 00:33:25.920
Even the most fancy compliance and risk programs are missing a lot.
00:33:26.079 --> 00:33:42.400
And sometimes you would see regulators, you would see the modern risk governance teams in those large organizations picking to those mistakes that the AI, and then when you're trying to ask them for the benchmarks of humans, you hear silence.
00:33:43.119 --> 00:34:01.759
Sometimes they're very proud of those programs, but they are missing the same analytical approach to the existing operating system, and they're trying to compare or to find those like issues with the AI, like it's um it's something that you need to prove.
00:34:01.920 --> 00:34:07.279
AI is not perfect, but so does human, and so does the processes that we have set in the past.
00:34:08.079 --> 00:34:18.800
We need to think about this implementation of the new solutions as an opportunity because there is a risk of implementing it, but there is a bigger risk of not implementing them.
00:34:18.960 --> 00:34:25.920
And you need to do it in a thoughtful way and to take an analytical approach about each individual use case and workflow.
00:34:26.000 --> 00:34:39.360
And if you're really thinking about risk in a risk length, you would get to probably different answers that sometimes people get to just by you know trying to compare AI to perfection.
00:34:39.920 --> 00:34:40.960
That's really interesting.
00:34:41.119 --> 00:34:47.280
Instead of comparing a Ford to a Ford, they're comparing a Ford to a Ferrari and saying, I don't know why it doesn't go so fast.
00:34:48.079 --> 00:34:48.719
It's interesting.
00:35:14.559 --> 00:35:20.320
Compared to where it was six months before that, I mean, we are moving faster than I've ever seen anything move.
00:35:20.400 --> 00:35:28.159
I mean, we're faster than the uh the data storage growth curve right now, and that's exponential growth since 2009.
00:35:28.559 --> 00:35:37.679
Um, and we're seeing all of that happen so fast that they have to start shutting down models because they're entirely too powerful at this point, according to the US government.
00:35:37.920 --> 00:35:47.039
Um, um and I'm kind of curious as you've gone through these processes, have you seen the technology start to outpace?
00:35:47.119 --> 00:35:55.280
And and how do you, as a as a leader of an AI company, stop yourself from chasing the next shiny object?
00:35:55.519 --> 00:36:02.800
How have you stayed so tried and true to the object the original objective of the organization?
00:36:04.000 --> 00:36:18.639
First off, I think that uh it's one of the hardest challenges, which is decide what's not to do, especially when you have such a talented group around you that can basically chase any crazy idea and potentially even make it.
00:36:18.880 --> 00:36:22.239
But if you try to chase them all, you probably won't.
00:36:22.559 --> 00:36:34.159
I think that one of our observations early on, and we had to give a lot of pride and realize that some of the hard work that we did is not longer our IP.
00:36:34.320 --> 00:36:44.239
And one of our observations led us to that conclusion that like AI is probably outpacing any other trend that we've seen in history.
00:36:44.480 --> 00:37:01.199
We won't be able to bid it, we won't be able to beat those foundation models, and we need to create an observability layer that basically orchestrate the best solution for every problem that we're trying to solve.
00:37:01.440 --> 00:37:05.679
For some, it's gonna be proprietary models that we retrained.
00:37:06.000 --> 00:37:22.320
For others, that's gonna be only on the prompt layer with the perfect uh foundation model that comes on top that would always take into consideration latency, accuracy, and the economic factor as well, right?
00:37:22.880 --> 00:37:41.119
So we basically decided to take an approach that we would be able to introduce those uh latest and greatest technology into production in a safe way rather than trying to beat the race by developing everything ourselves all the time.
00:37:41.360 --> 00:37:47.360
So it and it was a change in our approach uh a few years ago after we started.
00:37:47.920 --> 00:37:54.000
Well, that's a change in your organizational DNA to be able to say, hey, I'm not going to chase that next thing.
00:37:54.159 --> 00:38:00.480
And look, I've even struggled with it in the marketing firm and saying, okay, so is perplexity the right solution now?
00:38:00.559 --> 00:38:01.280
Oh, is it Claude?
00:38:01.360 --> 00:38:02.239
Is it Chat GPT?
00:38:02.320 --> 00:38:04.480
And what I've realized is there is no right solution.
00:38:04.639 --> 00:38:13.599
And it's about the ability to carry context across the solutions based on the use case of each individual person within the organization.
00:38:13.840 --> 00:38:34.719
And it wasn't until recently that I figured out kind of how that contextual uh carryover could actually function, how it could work, and how I could use that to more finitely empower the people around me to leverage the tools, even though they may not have that same depth of prompt engineering experience.
00:38:34.880 --> 00:38:35.119
Right?
00:38:35.199 --> 00:38:36.719
Because I'm not an engineer by trade.
00:38:36.880 --> 00:38:43.280
I just sat at the computer for three, four hundred hours until I could make it do the things that I wanted it to do.
00:38:44.000 --> 00:38:48.239
I'm not suggesting that everybody take that approach, but for me it had to be brute force, right?
00:38:48.400 --> 00:38:55.760
I'm sure for you it was probably a little bit uh I'd say more graceful than than uh than me running into the wall.
00:38:56.239 --> 00:38:56.880
We share that.
00:38:57.039 --> 00:39:01.280
We share that the brute force we it's either or brute force it.
00:39:01.440 --> 00:39:03.039
Those are the two only options.
00:39:03.280 --> 00:39:03.519
Right?
00:39:03.679 --> 00:39:10.159
I if I can't find it, if I can't finesse it, I have to brute force it, and I'm gonna figure it out one way or the other.
00:39:10.320 --> 00:39:22.480
Um, and you know, I went from I'm almost embarrassed to tell you, but I went from 2024 basically telling my organization do not use artificial intelligence like in the marketing process, like that's bad.
00:39:22.719 --> 00:39:33.599
You know, I frown upon this to the beginning of 2025, rolling out a series of tools to help us better understand the clients, build better profiles, find that voice, right?
00:39:33.679 --> 00:39:44.639
And to really be able to leverage some of these tools in new and creative ways, because I've realized that if you ask a chat GPT to give you something, it gives you trash.
00:39:44.800 --> 00:39:57.840
If you ask it the right way and you ask it to ask you questions about the intent, and you take a structured and thoughtful approach to the instructions that you're giving to the machine, you can get a thoughtful response back.
00:39:58.079 --> 00:40:03.840
But it took me a long time to realize that that was the reality of it.
00:40:04.079 --> 00:40:11.840
So, uh kind of a new question for you, Nir, that I want to start asking folks as we you know conduct this receivables podcast series.
00:40:12.000 --> 00:40:13.840
What's your go-to model these days?
00:40:14.000 --> 00:40:21.360
Where are you going as an executive, not in your deployment across your organization, but as another from one executive to another?
00:40:21.599 --> 00:40:27.039
What's the kind of go-to models for you these days for your kind of day-to-day?
00:40:28.480 --> 00:40:29.760
Mostly cloud.
00:40:30.400 --> 00:40:34.000
That would be like for 90% of the tasks.
00:40:34.320 --> 00:40:39.440
Sometimes when it involves visuals, I'll do Gemini, but mostly cloud.
00:40:39.679 --> 00:40:40.559
Mostly cloud.
00:40:40.719 --> 00:40:40.960
Yeah.
00:40:41.119 --> 00:40:41.440
Fair.
00:40:41.679 --> 00:40:43.199
Simple, straightforward.
00:40:43.840 --> 00:40:58.800
I look, and I feel like you were one of those early movers on Claude because you were talking to me about this like over a year ago, and I'm just starting to kind of move in that direction since co-work came out and trying to become more comfortable, confident with that, you know, tool set.
00:40:58.960 --> 00:41:04.320
Although my partner will keep me out of Claude code at any expense so that I don't break things.
00:41:04.480 --> 00:41:18.719
So I've started playing around with perplexity computer to be able to visualize some of the things that are in my head without torturing the developers and then be able to organize that in a way that I can put it back in their hands and they can do what they need to do with it to turn it into reality.
00:41:18.880 --> 00:41:21.519
Um, like you, I'm surrounded by some great people.
00:41:21.679 --> 00:41:25.920
They just have to kind of manage me a little bit so that I do the right things for them.
00:41:29.119 --> 00:41:31.920
Nir, this was a really incredible discussion.
00:41:32.000 --> 00:41:34.960
Is there I learned a lot from this talk today.
00:41:35.039 --> 00:41:40.079
Is there anything that we didn't cover that you wanted to include in our conversation today?
00:41:40.480 --> 00:41:44.000
Of course, but we need to have like another series just for that.
00:41:44.559 --> 00:41:46.639
You know, and that actually sounds like a good idea.
00:41:46.719 --> 00:41:48.400
We should probably talk about that.
00:41:48.960 --> 00:41:51.440
I feel like we're gonna have to do this more often.
00:41:52.559 --> 00:41:56.400
We definitely got the the gist and I enjoyed it.
00:41:56.559 --> 00:41:58.960
Thank you for uh for hosting me today.
00:41:59.039 --> 00:42:02.559
Uh it was a lot of fun, and uh, I got to learn a lot as well.
00:42:03.440 --> 00:42:05.599
Well, I really do appreciate your insights.
00:42:05.760 --> 00:42:11.599
For those of you that are watching, if you have additional questions you'd like to ask Nier or myself, you can leave those in the comments on LinkedIn and YouTube.
00:42:11.679 --> 00:42:12.960
We'll be responding to those.
00:42:13.119 --> 00:42:17.280
Or if you have additional topics you'd like to see us discuss, you can leave those in the comments below as well.
00:42:17.360 --> 00:42:23.679
And hopefully I can get Nier back at least one more time to help me continue to create great content for a great industry.
00:42:23.920 --> 00:42:25.679
But until next time, Nier, thank you so much.
00:42:25.760 --> 00:42:27.519
I really appreciate all your insights.
00:42:28.159 --> 00:42:29.039
Thank you, Adam.
00:42:30.000 --> 00:42:31.280
And thank you everybody for watching.
00:42:31.360 --> 00:42:32.719
We appreciate your time and attention.
00:42:32.880 --> 00:42:33.920
We'll see you all again soon.
00:42:34.159 --> 00:42:34.800
Bye, everyone.
00:00:06.559 --> 00:00:07.519
Hello, everybody.
00:00:07.599 --> 00:00:11.919
Adam Parks here with another episode of Receivables Podcast.
00:00:12.080 --> 00:00:49.600
Today I'm here with a great guest and a great topic here with Nir Laznick from Cedric, here to talk to me about well, keeping an eye on the bots, keeping the guardrails around those uh let's call it shiny new voice AI bots that everybody's talking about and how to build that governance guardrail framework around those tools because I still struggle with understanding why we would ever watch our phone calls differently, whether they were generated by a human or they were being generated through an AI technology.
00:00:49.759 --> 00:00:53.520
So, Nir, thank you so much for coming on, sharing your insights with me today.
00:00:53.679 --> 00:00:55.280
I look forward to learning from you.
00:00:55.759 --> 00:00:57.280
Thank you for hosting me, Adam.
00:00:57.439 --> 00:00:59.920
It's been a while and we've been trying to plan it.
00:01:00.000 --> 00:01:02.000
So I'm excited to already have it.
00:01:02.719 --> 00:01:04.560
Well, I appreciate you joining me today.
00:01:04.799 --> 00:01:13.040
Nir, for anyone who has not been as lucky as me to get to know you through the years, can you tell everyone a little about yourself and how you got to the seat that you're in today?
00:01:13.439 --> 00:01:17.040
Yes, uh, I always like to start with a personal introduction.
00:01:17.359 --> 00:01:22.000
So, Nier, I'm a father, always comes first.
00:01:22.400 --> 00:01:26.640
After that, after my daddy's dead, I'm an information system engineer.
00:01:26.799 --> 00:01:40.319
And a few years ago, myself and my talented co-founder, which is one of the leading data scientists back in Tel Aviv, identified a few macro trends that are now taking a very big part in the story of what we do.
00:01:40.560 --> 00:01:56.400
And the first one was the fragmentation of the financial services sector, which basically got us to the conclusion that financial services, in specific uh consumer finance, are coming in different shapes and forms.
00:01:56.640 --> 00:02:06.239
Alongside the traditional banking, we now have fintechs, we have embedded finance that creates like different shapes and forms of those consumer products.
00:02:06.480 --> 00:02:22.159
The other one was the increasing complexity around compliance, and together we realized that there is a disproportional pressure being made on compliance that was already slowing down business, and we wanted to unlock and to turn compliance into a business enabler.
00:02:22.319 --> 00:02:27.280
So, with that conclusion in mind, we accepted the first AI compliance officer.
00:02:27.439 --> 00:02:42.319
We called our Cedric, and we are here really to enable financial institutions to grow faster without breaking things, and on the collection side is really to unlock productivity without the burden and the risk that of the compliance that is involved.
00:02:42.960 --> 00:02:47.520
Well, prior to these kinds of tools, we maybe listened to 2% of the calls.
00:02:47.759 --> 00:02:59.680
And you know, that was the sample size that we were able to get to using live folks to listen to those calls, and through the deployment of artificial intelligence, we've seen that increase to 100%.
00:03:00.080 --> 00:03:14.960
And now the 2% that's being reviewed is the 2% of exceptions that were found systematically, which has changed the face of compliance for the debt collection and for the let's call call center customer service environments as well.
00:03:15.120 --> 00:03:19.520
And you've been on the leading edge of creating that technology through the years.
00:03:19.599 --> 00:03:24.240
It's been interesting to see how this has, you know, started to play out.
00:03:24.800 --> 00:03:42.800
Now, where you guys are even beyond the debt collection space and looking at the creditor space as well, where the debt is actually originated, what are some of the biggest concerns that these organizations have related to the deployment of artificial intelligence in general?
00:03:43.759 --> 00:03:45.039
It's a great question.
00:03:45.199 --> 00:03:57.759
And uh I think that um if we really observe on what's happening right now in the market, we are currently in the middle of the second phase of adoption of AI.
00:03:58.159 --> 00:04:10.639
The first one included a lot of pressure coming from board members to introduce AI solutions to increase productivity and to basically leverage those capabilities.
00:04:10.960 --> 00:04:16.879
The problem with that first adoption cycle is that the technology was not mature enough.
00:04:17.199 --> 00:04:23.519
We saw a very large number of POCs with a very high, above 90% failure.
00:04:23.680 --> 00:04:39.759
Now, what we're seeing in the second phase of adoption, which we're in the middle right now, is that the technology actually matured into a posture that it can actually deliver production large scale in different areas of the business.
00:04:40.000 --> 00:04:47.920
But the problem that we sometimes see, especially in the more conservative institutions, is that it's being blocked still.
00:04:48.240 --> 00:04:51.759
And in many cases, that would be by compliance.
00:04:51.920 --> 00:05:00.800
And one of the reasons for that being that way is that compliance needs to make sure that they have the right control on those solutions.
00:05:00.959 --> 00:05:12.399
They do want to support the business, but if they don't know enough and if they have blind spots, they rather just stop rather than taking the risk if they cannot quantify it.
00:05:12.639 --> 00:05:38.319
One of the things that we are trying to do is to really rethink compliance as the enabler of the business by enabling all of that supervision and oversight across all of the different channels, including some of those innovative ones, like the AI solutions, that's really matured into a stage where they can serve the different operations, even the conservative ones.
00:05:39.600 --> 00:05:46.959
So as you think about the type of technology that you're deploying for those organizations and the things that they're trying to accomplish now.
00:05:47.120 --> 00:06:00.720
So you've you've mastered your craft on those live human calls, and now these financial institutions are pushing and saying, we want to start using bots because we think it's going to service consumers in this way or that way.
00:06:02.399 --> 00:06:19.519
How has the have you had some of those conversations with them around how your tool set creates guardrails not just for those live callers, but also for let's call it the new shiny object in the room, the voice AI outbound call.
00:06:20.560 --> 00:06:21.360
Of course.
00:06:21.519 --> 00:06:27.279
And maybe we can travel back in time, probably two or three years together.
00:06:27.519 --> 00:06:39.759
And Cedric was in a crossroad when we realized that from a technology perspective, the gap between what we do right now to actually enable those as a first line is actually not that big.
00:06:40.000 --> 00:06:47.360
But the reason why we've decided to stay on the compliance side is that when you think about risk and compliance, it has to be a separate line of defense.
00:06:47.600 --> 00:06:57.600
Just like a very experienced collector cannot claim that they know everything and they don't need to be monitored because they've been doing it for a while.
00:06:57.759 --> 00:07:06.000
In a similar fashion, even the most compliance-oriented AI solution cannot just evaluate itself.
00:07:06.079 --> 00:07:14.480
You need to have a separate line of defense that can orchestrate the different channels, including those Gen AI related ones.
00:07:14.800 --> 00:07:40.800
So, what Cedric provides is an orchestration layer where you can basically take all of your policies that includes all the different complexities, jurisdiction considerations, different product nuances, and so forth, and to basically codify them into the control system that is basically serving as the infrastructure for any future agency workflow that you want to run on top of it.
00:07:40.959 --> 00:07:48.399
So that can be listening to voice calls, that can be listening to chat threads or email threads and so forth.
00:07:48.639 --> 00:07:55.360
And in other verticals, we actually support different mediums as well, which are less relevant to the count recovery space.
00:07:55.439 --> 00:08:08.000
But this is part of the stack of enabling financial institutions or a regulated ones to basically follow detailed, complex policies in different mediums where they interact with their consumers.
00:08:08.560 --> 00:08:21.600
So you've actually pivoted slightly from being a let's say call monitoring or call compliance quality type operation to looking at all channels, both written and spoken.
00:08:22.399 --> 00:08:24.959
Yes, we extended the offering.
00:08:26.560 --> 00:08:45.600
And we always like to take liked to take a very broad approach about what we do, like thinking about the entire workflow in specific verticals, so we can really represent all the nuances and the complexities of this vertical and really deeply understand the persona that we're serving.
00:08:45.759 --> 00:08:51.440
In that case, that's the compliance and the risk, and to really help them to take over the entire workflow.
00:08:51.600 --> 00:09:03.360
And while phone calls used to be the channel that took over most of the volume, we now see omnichannel and the number of channels that is only increasing to more and more.
00:09:03.519 --> 00:09:07.759
So that was a natural expansion for the product offering, basically.
00:09:08.159 --> 00:09:12.080
What makes me think about how people are stacking bots into workflows now?
00:09:12.240 --> 00:09:18.879
And rather than trying to be everything to everybody, you said, let me just be the master of this piece.
00:09:19.039 --> 00:09:24.720
But that piece carries across all touch points with the consumer in all channels.
00:09:24.960 --> 00:09:28.639
Because if you're doing the written, the email, the text messages, right?
00:09:28.720 --> 00:09:31.519
God knows what kind of communication we're going to come up with in the future.
00:09:31.679 --> 00:09:44.320
But you know, looking at how we've focused over the last couple of years, it seems like you guys have kind of well positioned yourselves to be that check and balance system against those communications.
00:09:45.039 --> 00:09:57.919
Have you started looking, and I'm sure you're doing it with let's call it the pre-fixed consumer journeys, but how does that technology play into uh let's say on the fly generated content to consumers?
00:09:58.080 --> 00:10:03.440
Is it watching it the same way that it was watching the live human, or how does that change over time?
00:10:04.080 --> 00:10:05.519
It's a great question.
00:10:05.679 --> 00:10:09.679
Uh, you did mention a comment that was spot on at the beginning.
00:10:10.080 --> 00:10:16.399
And basically, this is my philosophy about venture building, right?
00:10:16.559 --> 00:10:23.759
Especially especially in an era where the tech mode cannot be the only mode that you're having for your business.
00:10:24.480 --> 00:10:37.759
And you know, in the past you had like many horizontal solutions that served well many organizations, where today it's actually getting a lot of challenge in the build versus buy discussion that you have.
00:10:38.000 --> 00:10:52.320
But when you go deep in a specific vertical, in a complex problem that requires you to have domain expertise alongside your tech acumen as a firm, this is where you create a lot of value that's hard to replicate.
00:10:52.559 --> 00:11:07.519
And when we get partners on board, they're not only subscribing to the solution, they are subscribing to the roadmap, they are subscribing to the talent, they are subscribing to their potential to stay at the edge of technology every time.
00:11:08.159 --> 00:11:10.240
Um so yeah.
00:11:11.440 --> 00:11:12.480
No, it's interesting.
00:11:12.559 --> 00:11:15.120
Uh, I I like the approach to it.
00:11:15.279 --> 00:11:22.399
Do you think the primary fear of these financial institutions is the regulatory risk, the brand risk?
00:11:22.559 --> 00:11:31.919
Is it operational or is everything just blended together into one layer of anxiety for these compliance professionals?
00:11:33.360 --> 00:11:54.559
It's a great question, and it's actually an interesting point in time to ask it because like we all understand that the climate around the what is risk, how regulation is being enforced these days with the attempt dismantle of the CFPB, and um maybe the only constant which is change right now in that landscape.
00:11:54.960 --> 00:12:06.080
But um, yeah, I think that look, we need to remember that risk and being conservative is rooted in the DNA of some of those firms.
00:12:06.320 --> 00:12:13.279
And even if the macro changes right now, everyone understands that this is a string, right?
00:12:13.679 --> 00:12:16.320
It's it can be flexed, but it can be loosened.
00:12:16.399 --> 00:12:18.879
Nobody thinks that regulation is gone forever.
00:12:19.039 --> 00:12:20.799
We understand that it's changing.
00:12:20.960 --> 00:12:23.759
We also understand that it's not the same.
00:12:24.000 --> 00:12:34.559
We have the different attorney generals that are behaving slightly differently, which actually add a lot another layer of complexity for some jurisdictions, for some states.
00:12:34.799 --> 00:12:44.960
So one thing is clear: you do need to have a more flexible system that you can orchestrate either to loosen or to stress in your controls.
00:12:45.279 --> 00:12:51.039
So this is something that Cedric provides, and not just Cedric, but other platforms as well.
00:12:51.279 --> 00:13:00.320
And uh I think that it's a hard question because it is rooted in the DNA of those organizations.
00:13:00.879 --> 00:13:18.480
Some would take those opportunities to maybe loosen some of the controls, some would understand that the change management has so much friction and potential risk that you don't know what would be the perspective of the regulator for whatever control change management that you went through throughout that time.
00:13:19.200 --> 00:13:22.080
I think each one of those comes in ebbs and flows.
00:13:22.399 --> 00:13:38.000
So when we think about regulatory risk, you know, I think who's in power in the US government and how powerful is the CFPB at that moment in time, who's the director, all plays into or all signals that we look at to measure the level of regulatory risk.
00:13:38.240 --> 00:13:53.519
At the same time, I think we've got privacy concerns and patchworks of information, patchworks of requirements that are starting to bloom around the country that are requiring us to take new steps that haven't been taken before, right?
00:13:53.600 --> 00:14:06.639
To organize ourselves in different ways and being on the forefront of a lot of that uh requirement and staying, let's say, staying on that wave is a full-time job in and of itself.
00:14:06.879 --> 00:14:36.960
As a reformed uh compliance professional myself, you know, I know how much time, effort, and energy went into maintaining that level of sophistication flexibility to be able to deploy the different solutions as these requirements changed, and the and really the the reactionary time periods to be able to respond to these changes in a reasonable time period, because you're not always given a reasonable time period um when something new is being deployed.
00:14:37.120 --> 00:14:50.480
Um, so we start thinking about how the technology has changed over time, and now we're starting to talk about not just getting a human on the phone, but getting this bot on the phone.
00:14:50.639 --> 00:15:01.519
And it reminds me of the movie casino where you know the pit boss is watching the dealer, and the you know, the floor manager is watching the pit boss, and the eye in the sky is watching us all.
00:15:01.679 --> 00:15:10.639
And I think it's a great analogy for what you guys have been able to do, kind of circling up all of these solutions together.
00:15:10.879 --> 00:15:17.279
And you know, what has been your experience as you've been able to put some guardrails in place?
00:15:17.440 --> 00:15:25.200
Is that helping to enable those financial institutions to take that calculated risk and try something new?
00:15:25.360 --> 00:15:34.399
Or are they still maintaining kind of those tight reins and not venturing into the new technology that's available?
00:15:35.360 --> 00:15:42.879
Look, what I can only testify from uh you know our portfolio, uh I believe it does represent a trend.
00:15:43.519 --> 00:15:50.639
And we are actually for us, it's a backwind, right?
00:15:50.879 --> 00:16:02.240
Because, like I mentioned before, we are in the middle of the second adoption cycle where the technology matured into a state where it can actually deliver, not only to promise.
00:16:02.720 --> 00:16:16.799
And one of the things that we're seeing again is that there is a friction coming from compliance that doesn't feel comfortable to introduce and to roll out some of those programs to larger scales because of the blind spots.
00:16:17.200 --> 00:16:27.919
So we did see how Cedric basically plays a role in enabling those programs moving from a pilot stage into production large-scale programs.
00:16:28.080 --> 00:16:31.360
So, to your question, the answer is definitely yes.
00:16:32.240 --> 00:16:57.039
So, from your perspective, uh, when organizations are rolling out a new piece of AI technology, are they generally talking to the financial institution and saying, well, look, you know, we our our bot is watching our other bot, or how comfortable do they become as a service provider, knowing that there's an overwatch of their technology from another compliance tool set?
00:16:57.200 --> 00:17:04.559
What did what do those relationships start to look like and does it enable them to actively forward?
00:17:05.119 --> 00:17:18.319
Are you asking about the relationship between the creditor and the so from from your perspective, when you roll out, so uh uh a financial institution is going to roll out this new tool set?
00:17:18.480 --> 00:17:26.720
It's been my experience that a lot of organizations are saying, hey, I have this AI bot and I've got this judge LLM that's gonna watch the bot, right?
00:17:26.799 --> 00:17:35.680
But like you had said at the beginning, that generally is falling somewhat within the same environment and not necessarily a third-party check and balance system.
00:17:36.000 --> 00:17:45.359
What has that relationship been like for you guys coming in as kind of a compliance umbrella that's sitting over not only their voice bot but their judge LLM?
00:17:45.440 --> 00:17:49.119
And what how has that worked in reality?
00:17:49.279 --> 00:17:52.640
Like I could see it conceptually, but like what's the reality of that situation?
00:17:52.880 --> 00:17:58.559
Yeah, so let's map out the different players in that in that uh ecosystem that you just uh outlined.
00:17:58.799 --> 00:18:02.240
So we would have the AICX solution.
00:18:02.559 --> 00:18:06.880
CERUC would be an enabler for a production rollout for them.
00:18:07.279 --> 00:18:19.680
We have the agency that would potentially like to get better allocation, and by demonstrating better compliance posture in high confidence, that would help them to do so.
00:18:19.920 --> 00:18:32.319
For the creditor, it would give them the comfort that a solution that is trained to do that specific job, not a generic AI solution that is also doing compliance.
00:18:32.480 --> 00:18:43.039
This is not something that you would just trust a generic solution, just like you probably won't trust a generic LLM to review the contract before you buy a house, right?
00:18:43.119 --> 00:18:52.559
You want to make sure that it's nailed and that you're not missing a bit because you are basically jeopardizing your firm by going after a generic solution.
00:18:52.720 --> 00:18:54.559
So you want to feel confident.
00:18:54.880 --> 00:19:07.680
So, yeah, if we think about the voice AI solution, this is an enabler for them, for the agency that would give them more confidence and hopefully get them better business with the creditor.
00:19:08.000 --> 00:19:14.319
And for the creditor, that reduces the exposure for them from both regulatory but also brand.
00:19:14.480 --> 00:19:26.400
As you climb up the ladder, the larger creditors sometimes cares way more about brand and compliance rather than the collection rate, and they want to make sure that they are protected.
00:19:27.359 --> 00:19:28.000
Sure, right?
00:19:28.319 --> 00:19:37.759
Debt collection is an afterthought for an originating creditor or a major financial institution, they're making their money on lending money, and this is a byproduct of it.
00:19:38.079 --> 00:19:51.599
But you do get me thinking here as I as I start working my way through, you know, is there a fundamental difference between watching an AI bot and monitoring its activity versus monitoring a live collector?
00:19:51.680 --> 00:19:59.759
Is there anything fundamentally different between those two scenarios, or are we basically taking treating everything as equal?
00:20:00.960 --> 00:20:03.680
Yeah, it's not it's not the same for sure.
00:20:03.920 --> 00:20:15.599
First off, if done right, and it's not always the case, you do have a good sense of control on the Gen AI side, right?
00:20:16.160 --> 00:20:26.160
There are really good firms that are doing the right thing, they are educated about compliance, and the initial output would generally be good, right?
00:20:26.559 --> 00:20:30.000
The other side, which is the consumer, you don't get to control.
00:20:30.480 --> 00:20:44.079
Sometimes it can drag the conversation to edge cases that were not trained, or just impossible to be trained at the technology stage right now, and this is where the risk would potentially surface.
00:20:44.400 --> 00:20:51.680
So basically, those nuances, those edge cases, and this is where it's going to be treated differently.
00:20:52.000 --> 00:21:02.319
The amount of control that you're gonna have on one side versus the other would also be a difference between a human to human versus an AI agent to a human.
00:21:02.480 --> 00:21:09.119
And there are other considerations that are more technical, but high-level speaking, there is a difference.
00:21:10.799 --> 00:21:16.079
So as we think about the differences there, it's you know, from the conversations that I've been having.
00:21:17.359 --> 00:21:31.359
For those of us that use Chat GPT, Claude, right, like the general kind of use case, LLMs across the space, we've all been subject to and seen hallucinations, degradation of a session, right?
00:21:31.440 --> 00:21:41.119
The canary in the coal mine, there's all different kinds of ways that those of us that engineer on a regular basis are kind of building those things out and trying to understand that degregation.
00:21:41.440 --> 00:22:11.359
From your perspective, as some as as having Overwatch for all of these different types of technologies, is that a real risk in the AI voice space, or have we come far enough over the last, let's say, two years or so to where the hallucinations are are really Minimal to almost non-existent in that voice space, or are we still seeing slips of the models versus ironing out of the uh campaigns and programs?
00:22:13.119 --> 00:22:17.039
There is no straight answer to that, but there is a trade-off, right?
00:22:17.200 --> 00:22:25.759
Okay, if you would narrow the degrees of freedom of your solution, you would potentially have less hallucinations.
00:22:26.720 --> 00:22:35.519
But that the trade-off would be that they won't be able to cover some of the edge cases, and then you're gonna have to escalate maybe a higher percentage to human.
00:22:35.920 --> 00:22:44.079
On the other side, you can increase the degrees of freedom, and you cannot completely um avoid having hallucinations.
00:22:44.240 --> 00:22:57.839
There are a lot of best practices, and I would leave those best practices to the ones that are implying them for the Gen AI solutions that can really improve and get this technology again to a very high maturity.
00:22:57.920 --> 00:23:04.880
And this is why we do see large-scale adoptions taking place right now for AICX solutions.
00:23:05.039 --> 00:23:13.359
Like the tech is there, they they were able to like reduce the amount of hallucinations to a minimal amount.
00:23:14.640 --> 00:23:28.160
So, as we think about monitoring calls and solutions for the financial institutions, how much are they monitoring their downstream placement?
00:23:28.319 --> 00:23:43.519
So the financial institution either sells the accounts, they place the accounts with a collection agency, a law firm, whatever the case or strategy may be, how much of that oversight is being carried down that stream over time?
00:23:43.759 --> 00:23:59.680
And it was uh it was a work it was an individual from a major bank who recently mentioned to me uh at one of the conferences that the better that they got with AI, the more difficult the debt collector's life was going to become because the low-hanging fruit wasn't going to be there.
00:23:59.839 --> 00:24:05.200
They were gonna work out those solutions, they were gonna be able to kind of resolve for that.
00:24:05.519 --> 00:24:18.079
But as you're looking at it from a monitoring perspective, and you're thinking about as these accounts start to flow into the collections lifecycle, how much of that monitoring continues to flow downstream?
00:24:18.160 --> 00:24:24.000
And what are those what are the expectations being set by those financial institutions today?
00:24:24.480 --> 00:24:26.720
Look, I think that the bar is just higher.
00:24:26.960 --> 00:24:36.400
I mean, if you could avoid um patterns and the cycle from event to remediation used to take months, sometimes quarters.
00:24:36.640 --> 00:24:41.599
Now the expectation is to adopt and to close the feedback loop much faster.
00:24:41.839 --> 00:24:53.359
I can tell you that the similar trend of um the sample-based to 100% and exception-based based on risk trend that you had mentioned before.
00:24:53.599 --> 00:24:56.000
This is the exact same trend that we see.
00:24:56.079 --> 00:25:01.119
Like I can refer to one of our maybe larger banks that we're serving.
00:25:01.359 --> 00:25:09.920
We basically held them to oversee 100% of the communication of all of the outsource agencies that are serving them.
00:25:10.319 --> 00:25:10.799
Right?
00:25:11.039 --> 00:25:12.559
It wasn't the case.
00:25:12.799 --> 00:25:16.400
It used to be probably less than 1% in the past.
00:25:17.279 --> 00:25:22.400
It used to be submit these 12 call recordings, and now you're saying it's 100%.
00:25:22.960 --> 00:25:26.400
Is that a direct connection or are they kicking over those calls every month?
00:25:26.480 --> 00:25:30.000
Or now you're just directly connected into that process from a monitoring?
00:25:31.119 --> 00:25:33.759
It would vary between one agency to another.
00:25:33.920 --> 00:25:43.920
Like one agency can be a direct integration into the dialers, another can be just an SFTP that they're gonna push uh the data towards Cedric, and we're gonna analyze that.
00:25:44.559 --> 00:25:45.119
Interesting.
00:25:45.279 --> 00:26:11.920
So, regardless of the uh let's say the technology connection between that creditor and that service provider, you've got a mechanism to hook into that for that monitoring, which I would think from a uh even from an agency perspective would be preferential because now they don't have to manage this process so much as just kind of open the doors and allow you to see you know what you need to see, which which makes sense.
00:26:12.160 --> 00:26:15.839
Um true, although it's like redefine the relationship, right?
00:26:16.400 --> 00:26:16.720
Sure.
00:26:17.680 --> 00:26:18.799
To some extent.
00:26:18.960 --> 00:26:22.480
I mean, you are very exposed, transparent.
00:26:22.880 --> 00:26:40.079
We always believed in the films that we're serving and and their intention, but uh it's slightly harder when everything is flowing that fast and the expectation to fix is also uh basically being redefined as part of this uh change in the process.
00:26:40.559 --> 00:26:45.359
Well, reaction and response times change as information moves faster, right?
00:26:45.440 --> 00:27:10.240
If if I'm sending you 12 calls and you're reviewing them and sending them back later this month, and now I can, you know, instigate or I can begin my retraining of whatever individual and kind of work my way through it, but that's like a month-long cycle, and you're saying that the feedback cycle and the expectation to implement is being shrunk, I'm assuming, to hours at this point based on the flow of information and capabilities, if not real time.
00:27:12.319 --> 00:27:27.680
How much of this is is starting to happen from a real-time perspective when we think about call monitoring and and keeping an eye on the agents or the bots, you know, how much of the the compliance is being done as let's say part of a process and how much of it is being done like in the moment.
00:27:28.880 --> 00:27:44.160
So if we think about different types of controls in compliance, and the way we structured it is that you can have preventive, detective, or corrective controls.
00:27:44.480 --> 00:27:59.920
The real time that you're referring to would be considered as preventive control because we would connect into your system of record, into your GRC or the policy instance that we uploaded alongside the live stream of the interaction.
00:28:00.079 --> 00:28:07.759
We will ingest that and fuse it into the model in real time and basically suggest those remediation live.
00:28:08.160 --> 00:28:20.079
But back to our previous example, we didn't see any live integration or expectation from the creditor to have live preventive monitoring on the outsourced agencies.
00:28:20.240 --> 00:28:24.720
This is something that we didn't see to date, although it's technically possible.
00:28:25.039 --> 00:28:31.359
But on the other side, whether it's the agency themselves as a client, we see it a lot.
00:28:31.440 --> 00:28:40.960
And usually that's going to be a fusion of productivity, increased collection rate alongside compliance and not compliance as a standalone.
00:28:41.279 --> 00:28:47.359
Similar trend we see for the creditor when they are doing it for their internal servicing teams, right?
00:28:47.920 --> 00:28:58.480
Whether it's for the performing stage or the delinquent stage, before they're placing it with their first or third party partners, they are using those real-time solutions.
00:28:58.960 --> 00:29:00.559
That are actively using those solutions.
00:29:00.720 --> 00:29:01.519
That's interesting.
00:29:01.759 --> 00:29:09.599
Uh, you know, when we start thinking about how the organizations are coming together and how quickly things are changing right now.
00:29:09.759 --> 00:29:19.039
I mean, even just in the last couple of years, as people have started to deploy artificial intelligence in all kinds of new and interesting places around their organization.
00:29:19.440 --> 00:29:28.079
What's the most surprising use case that you've seen for an organization to deploy a new piece of technology that you've then had to go monitor?
00:29:29.359 --> 00:29:34.079
In general, not any any curveballs that you've received.
00:29:34.160 --> 00:29:39.680
And I mean, I've I've received a hundred curveballs, I feel like, in the last 12 to 18 months.
00:29:39.839 --> 00:29:47.279
But has anybody actually put a piece of technology in front of you to where you had to say, all right, I gotta I gotta think through how I'm actively gonna monitor this?
00:29:47.359 --> 00:29:53.039
Or have you been able to get everything into let's call it the buckets that you're comfortable with?
00:29:54.640 --> 00:29:57.279
Um it's a great question.
00:29:57.440 --> 00:30:02.799
Um let me think if I have a an exotic example to share.
00:30:07.279 --> 00:30:19.119
I think one of the most exotics that uh maybe it's similar crossroads that we had with uh with the the call monitoring piece.
00:30:19.599 --> 00:30:24.720
Cedric also has other other offering related to marketing compliance.
00:30:24.880 --> 00:30:37.680
So we basically take over the entire uh consumer journey starting from the marketing and ending up with servicing after underwriting and then delinquency, collection, and and the rest we know.
00:30:37.920 --> 00:30:49.680
And one of the things that we've been asked recently is instead of just identifying uh what's wrong in the promotion or the ad, can we already fix it?
00:30:50.000 --> 00:31:02.240
Now, theoretically we can, but uh again, are we the one creating the marketing in a compliant way, or are we the one doing the compliance oversight for that uh content creation?
00:31:02.400 --> 00:31:12.319
And we made a similar decision of sticking to our practice of being that separate separate line of defense, basically reflecting the compliance.
00:31:13.039 --> 00:31:17.680
That's a really interesting use case on the marketing side leading into it, right?
00:31:17.839 --> 00:31:27.839
Finding that consumer, getting them in, getting them under written application process, um, that whole thing, and thinking about the compliance guardrails that are necessary there.
00:31:28.000 --> 00:31:34.000
And I guess that I mean it really is similar from a mechanical standpoint to collecting.
00:31:34.160 --> 00:31:35.039
It's not that, right?
00:31:35.119 --> 00:31:39.440
Like loan servicing and collecting from a mechanics perspective is not that different.
00:31:39.680 --> 00:31:44.799
Intent and drive and motivation and all that is is is where the differentiation falls in.
00:31:44.960 --> 00:31:53.359
But that's an interesting use case, and not even that crazy or exotic, but it makes a lot of sense as to why you would move into that space.
00:31:53.519 --> 00:32:00.000
Again, you're already servicing that financial institution, you're already providing that compliance team with that oversight.
00:32:00.160 --> 00:32:05.599
So adding another channel in which you're monitoring, I think makes a lot of sense.
00:32:05.680 --> 00:32:09.839
It's you know, it's economies of scope, it's the same reason McDonald's sells breakfast, right?
00:32:09.920 --> 00:32:15.759
They already had a restaurant and they already had a kitchen and everything else that they needed to be able to do that.
00:32:15.920 --> 00:32:19.200
That's an interesting, it's an interesting approach.
00:32:19.519 --> 00:32:36.480
Now, as you as you look at how you are engaging with those financial institutions, what's the biggest misconception that they have about monitoring their compliance using artificial intelligence?
00:32:37.039 --> 00:32:47.359
Like, what's the biggest challenge that you have to overcome in their mind to get them comfortable with this being the right solution for their financial institution?
00:32:49.200 --> 00:33:00.319
My answer would be consistent between what we see in those larger FIs, smaller FIs, and regulators alike, regulators alike.
00:33:00.799 --> 00:33:14.720
There is a very common bias and misconception about comparing AI to perfection, where suddenly people forget that you need to compare AI to current state.
00:33:14.960 --> 00:33:20.400
And current state in many cases is not that shiny and not that good.
00:33:20.799 --> 00:33:25.920
Even the most fancy compliance and risk programs are missing a lot.
00:33:26.079 --> 00:33:42.400
And sometimes you would see regulators, you would see the modern risk governance teams in those large organizations picking to those mistakes that the AI, and then when you're trying to ask them for the benchmarks of humans, you hear silence.
00:33:43.119 --> 00:34:01.759
Sometimes they're very proud of those programs, but they are missing the same analytical approach to the existing operating system, and they're trying to compare or to find those like issues with the AI, like it's um it's something that you need to prove.
00:34:01.920 --> 00:34:07.279
AI is not perfect, but so does human, and so does the processes that we have set in the past.
00:34:08.079 --> 00:34:18.800
We need to think about this implementation of the new solutions as an opportunity because there is a risk of implementing it, but there is a bigger risk of not implementing them.
00:34:18.960 --> 00:34:25.920
And you need to do it in a thoughtful way and to take an analytical approach about each individual use case and workflow.
00:34:26.000 --> 00:34:39.360
And if you're really thinking about risk in a risk length, you would get to probably different answers that sometimes people get to just by you know trying to compare AI to perfection.
00:34:39.920 --> 00:34:40.960
That's really interesting.
00:34:41.119 --> 00:34:47.280
Instead of comparing a Ford to a Ford, they're comparing a Ford to a Ferrari and saying, I don't know why it doesn't go so fast.
00:34:48.079 --> 00:34:48.719
It's interesting.
00:35:14.559 --> 00:35:20.320
Compared to where it was six months before that, I mean, we are moving faster than I've ever seen anything move.
00:35:20.400 --> 00:35:28.159
I mean, we're faster than the uh the data storage growth curve right now, and that's exponential growth since 2009.
00:35:28.559 --> 00:35:37.679
Um, and we're seeing all of that happen so fast that they have to start shutting down models because they're entirely too powerful at this point, according to the US government.
00:35:37.920 --> 00:35:47.039
Um, um and I'm kind of curious as you've gone through these processes, have you seen the technology start to outpace?
00:35:47.119 --> 00:35:55.280
And and how do you, as a as a leader of an AI company, stop yourself from chasing the next shiny object?
00:35:55.519 --> 00:36:02.800
How have you stayed so tried and true to the object the original objective of the organization?
00:36:04.000 --> 00:36:18.639
First off, I think that uh it's one of the hardest challenges, which is decide what's not to do, especially when you have such a talented group around you that can basically chase any crazy idea and potentially even make it.
00:36:18.880 --> 00:36:22.239
But if you try to chase them all, you probably won't.
00:36:22.559 --> 00:36:34.159
I think that one of our observations early on, and we had to give a lot of pride and realize that some of the hard work that we did is not longer our IP.
00:36:34.320 --> 00:36:44.239
And one of our observations led us to that conclusion that like AI is probably outpacing any other trend that we've seen in history.
00:36:44.480 --> 00:37:01.199
We won't be able to bid it, we won't be able to beat those foundation models, and we need to create an observability layer that basically orchestrate the best solution for every problem that we're trying to solve.
00:37:01.440 --> 00:37:05.679
For some, it's gonna be proprietary models that we retrained.
00:37:06.000 --> 00:37:22.320
For others, that's gonna be only on the prompt layer with the perfect uh foundation model that comes on top that would always take into consideration latency, accuracy, and the economic factor as well, right?
00:37:22.880 --> 00:37:41.119
So we basically decided to take an approach that we would be able to introduce those uh latest and greatest technology into production in a safe way rather than trying to beat the race by developing everything ourselves all the time.
00:37:41.360 --> 00:37:47.360
So it and it was a change in our approach uh a few years ago after we started.
00:37:47.920 --> 00:37:54.000
Well, that's a change in your organizational DNA to be able to say, hey, I'm not going to chase that next thing.
00:37:54.159 --> 00:38:00.480
And look, I've even struggled with it in the marketing firm and saying, okay, so is perplexity the right solution now?
00:38:00.559 --> 00:38:01.280
Oh, is it Claude?
00:38:01.360 --> 00:38:02.239
Is it Chat GPT?
00:38:02.320 --> 00:38:04.480
And what I've realized is there is no right solution.
00:38:04.639 --> 00:38:13.599
And it's about the ability to carry context across the solutions based on the use case of each individual person within the organization.
00:38:13.840 --> 00:38:34.719
And it wasn't until recently that I figured out kind of how that contextual uh carryover could actually function, how it could work, and how I could use that to more finitely empower the people around me to leverage the tools, even though they may not have that same depth of prompt engineering experience.
00:38:34.880 --> 00:38:35.119
Right?
00:38:35.199 --> 00:38:36.719
Because I'm not an engineer by trade.
00:38:36.880 --> 00:38:43.280
I just sat at the computer for three, four hundred hours until I could make it do the things that I wanted it to do.
00:38:44.000 --> 00:38:48.239
I'm not suggesting that everybody take that approach, but for me it had to be brute force, right?
00:38:48.400 --> 00:38:55.760
I'm sure for you it was probably a little bit uh I'd say more graceful than than uh than me running into the wall.
00:38:56.239 --> 00:38:56.880
We share that.
00:38:57.039 --> 00:39:01.280
We share that the brute force we it's either or brute force it.
00:39:01.440 --> 00:39:03.039
Those are the two only options.
00:39:03.280 --> 00:39:03.519
Right?
00:39:03.679 --> 00:39:10.159
I if I can't find it, if I can't finesse it, I have to brute force it, and I'm gonna figure it out one way or the other.
00:39:10.320 --> 00:39:22.480
Um, and you know, I went from I'm almost embarrassed to tell you, but I went from 2024 basically telling my organization do not use artificial intelligence like in the marketing process, like that's bad.
00:39:22.719 --> 00:39:33.599
You know, I frown upon this to the beginning of 2025, rolling out a series of tools to help us better understand the clients, build better profiles, find that voice, right?
00:39:33.679 --> 00:39:44.639
And to really be able to leverage some of these tools in new and creative ways, because I've realized that if you ask a chat GPT to give you something, it gives you trash.
00:39:44.800 --> 00:39:57.840
If you ask it the right way and you ask it to ask you questions about the intent, and you take a structured and thoughtful approach to the instructions that you're giving to the machine, you can get a thoughtful response back.
00:39:58.079 --> 00:40:03.840
But it took me a long time to realize that that was the reality of it.
00:40:04.079 --> 00:40:11.840
So, uh kind of a new question for you, Nir, that I want to start asking folks as we you know conduct this receivables podcast series.
00:40:12.000 --> 00:40:13.840
What's your go-to model these days?
00:40:14.000 --> 00:40:21.360
Where are you going as an executive, not in your deployment across your organization, but as another from one executive to another?
00:40:21.599 --> 00:40:27.039
What's the kind of go-to models for you these days for your kind of day-to-day?
00:40:28.480 --> 00:40:29.760
Mostly cloud.
00:40:30.400 --> 00:40:34.000
That would be like for 90% of the tasks.
00:40:34.320 --> 00:40:39.440
Sometimes when it involves visuals, I'll do Gemini, but mostly cloud.
00:40:39.679 --> 00:40:40.559
Mostly cloud.
00:40:40.719 --> 00:40:40.960
Yeah.
00:40:41.119 --> 00:40:41.440
Fair.
00:40:41.679 --> 00:40:43.199
Simple, straightforward.
00:40:43.840 --> 00:40:58.800
I look, and I feel like you were one of those early movers on Claude because you were talking to me about this like over a year ago, and I'm just starting to kind of move in that direction since co-work came out and trying to become more comfortable, confident with that, you know, tool set.
00:40:58.960 --> 00:41:04.320
Although my partner will keep me out of Claude code at any expense so that I don't break things.
00:41:04.480 --> 00:41:18.719
So I've started playing around with perplexity computer to be able to visualize some of the things that are in my head without torturing the developers and then be able to organize that in a way that I can put it back in their hands and they can do what they need to do with it to turn it into reality.
00:41:18.880 --> 00:41:21.519
Um, like you, I'm surrounded by some great people.
00:41:21.679 --> 00:41:25.920
They just have to kind of manage me a little bit so that I do the right things for them.
00:41:29.119 --> 00:41:31.920
Nir, this was a really incredible discussion.
00:41:32.000 --> 00:41:34.960
Is there I learned a lot from this talk today.
00:41:35.039 --> 00:41:40.079
Is there anything that we didn't cover that you wanted to include in our conversation today?
00:41:40.480 --> 00:41:44.000
Of course, but we need to have like another series just for that.
00:41:44.559 --> 00:41:46.639
You know, and that actually sounds like a good idea.
00:41:46.719 --> 00:41:48.400
We should probably talk about that.
00:41:48.960 --> 00:41:51.440
I feel like we're gonna have to do this more often.
00:41:52.559 --> 00:41:56.400
We definitely got the the gist and I enjoyed it.
00:41:56.559 --> 00:41:58.960
Thank you for uh for hosting me today.
00:41:59.039 --> 00:42:02.559
Uh it was a lot of fun, and uh, I got to learn a lot as well.
00:42:03.440 --> 00:42:05.599
Well, I really do appreciate your insights.
00:42:05.760 --> 00:42:11.599
For those of you that are watching, if you have additional questions you'd like to ask Nier or myself, you can leave those in the comments on LinkedIn and YouTube.
00:42:11.679 --> 00:42:12.960
We'll be responding to those.
00:42:13.119 --> 00:42:17.280
Or if you have additional topics you'd like to see us discuss, you can leave those in the comments below as well.
00:42:17.360 --> 00:42:23.679
And hopefully I can get Nier back at least one more time to help me continue to create great content for a great industry.
00:42:23.920 --> 00:42:25.679
But until next time, Nier, thank you so much.
00:42:25.760 --> 00:42:27.519
I really appreciate all your insights.
00:42:28.159 --> 00:42:29.039
Thank you, Adam.
00:42:30.000 --> 00:42:31.280
And thank you everybody for watching.
00:42:31.360 --> 00:42:32.719
We appreciate your time and attention.
00:42:32.880 --> 00:42:33.920
We'll see you all again soon.
00:42:34.159 --> 00:42:34.800
Bye, everyone.
