Fintech One-On-One

Castellum CEO argues AI compliance agents require specialized data harnesses to achieve 99% accuracy over raw LLMs

32 min Listen to the original
Financial Crime ComplianceArtificial IntelligenceRegulatory TechnologySanctions ScreeningFintech

Summary

Peter Piatetsky, CEO of Castellum.ai, explains that his company provides AI agents for financial crime compliance that integrate with existing screening systems rather than replacing them. He claims that without a specialized 'harness' of data and testing, raw Large Language Models (LLMs) achieve only 70% accuracy, whereas Castellum's approach reaches 99%. Piatetsky notes that banks over-generate alerts due to regulatory pressure, creating false positives that AI can filter out. He predicts that Level 1 compliance work will be fully automated within two to three years as criminals also adopt AI for scams.

Episode sections

  1. 0:10 – 6:23

    Introduction and Peter Piatetsky's Background

    Peter Piatetsky, CEO of Castellum.ai, discusses his background as an intelligence analyst in Afghanistan and at the US Treasury Department, leading to the founding of Castellum in 2019 to solve compliance data access issues.

  2. 6:23 – 9:01

    Castellum's Evolution from Data to AI Agents

    Castellum shifted from being a primary source data company to offering modular AI agents for alert resolution to avoid the high change management costs of replacing entire compliance systems.

  3. 9:01 – 16:33

    The Problem with Alert Volume and AI Accuracy

    Piatetsky argues that banks over-generate alerts to avoid regulatory fines, drowning analysts in false positives, and claims AI agents can remove over 95% of these alerts.

  4. 16:33 – 18:18

    The 'Harness' vs. Raw LLM Debate

    Piatetsky distinguishes between using raw LLMs and using them within a 'harness' of specialized data and infrastructure, asserting the harness is critical for reliability in financial crime compliance.

  5. 18:18 – 21:17

    Benchmark Results: 70% vs. 99% Accuracy

    Castellum tested nine AI models, finding 70% accuracy without a harness and 99% with it, warning that a 30% error rate is unacceptable for compliance officers.

  6. 21:17 – 25:46

    Integration Strategy and Market Position

    Castellum integrates agents into existing systems like LexisNexis Bridger and Ferkasoft, targeting mid-market banks and fintechs with a 'water filter' approach.

  7. 25:46 – 32:25

    Criminal Use of AI and Future Outlook

    Piatetsky notes criminals are using AI to scale scams 10-20x and predicts Level 1 and Level 2 compliance work will be fully automated within two to three years.

Claims worth verifying

Top AI models were accurate only 70% of the time without a harness versus 99% of the time with a Castellum harness when asked to read policies, conduct screening, create an alert, and adjudicate.

Peter Piatetsky ·18:18

How to verify: Request Castellum's 'FinCrime agent benchmark' report or the 'Agentic screening market map' which cites this research; verify by replicating the test with 9 models using and not using the specified harness.

Using just Claude without a harness results in a 30% error rate (3 out of 10 alerts wrong), which is unacceptable for compliance officers.

Peter Piatetsky ·18:42

How to verify: Verify by testing Claude (Anthropic) on a standardized set of compliance alerts without specialized prompting/harnessing and measuring error rates against ground truth adjudications.

Castellum was founded in 2019.

Peter Piatetsky ·1:30

How to verify: Check corporate registration records or Castellum.ai's official 'About Us' page for founding date.

Castellum is a primary source for sanctions, PEP, and adverse media data, not outsourcing data collection.

Peter Piatetsky ·7:01

How to verify: Review Castellum's technical documentation or data sourcing methodology whitepapers to confirm direct publisher relationships and lack of third-party data aggregation.

AI agents can remove 95%+ of alerts accurately for clients.

Peter Piatetsky ·9:55

How to verify: Request client case studies or performance metrics from Castellum detailing alert reduction rates post-implementation.

Peter Piatetsky worked as an intelligence analyst in Afghanistan and later at the US Treasury Department (OFAC) on Iran and Russia.

Peter Piatetsky ·1:12

How to verify: Check Peter Piatetsky's LinkedIn profile or professional biography on Castellum.ai for employment history.

Castellum has integrations with LexisNexis Bridger, Ferkasoft, TRM, and Hummingbird.

Peter Piatetsky ·21:17

How to verify: Check Castellum's partner page or press releases for announcements of these specific integrations.

Criminal scam volume has increased 10-20x due to AI usage.

Peter Piatetsky ·26:46

How to verify: Review industry reports from financial crime monitoring firms (e.g., Chainalysis, Elliptic, or government reports like FinCEN) on AI-driven fraud trends.

Level 1 compliance work will be fully automated within two to three years.

Peter Piatetsky ·28:37

How to verify: Monitor industry adoption rates of AI alert resolution tools and job market trends for Level 1 compliance analysts over the next 36 months.

Castellum's agents undergo a 10-day training period involving open-book and closed-book tests.

Peter Piatetsky ·12:35

How to verify: Review Castellum's implementation methodology documentation or sales collateral describing the onboarding process.

Do your own research

  • Verify Castellum's 99% accuracy claim by reviewing their published benchmark methodology and the specific nine AI models tested.
  • Check if LexisNexis Bridger and Ferkasoft officially list Castellum as an integrated partner or technology provider.
  • Investigate regulatory guidance on 'not enough alerts' versus 'too many alerts' to understand the true incentive structure for banks.
  • Research the current adoption rate of AI agents in Level 1 compliance roles among mid-market banks and fintechs.
  • Review Castellum's data sourcing practices to confirm they collect sanctions and PEP data directly from publishers.
  • Analyze the cost-benefit analysis of integrating Castellum's modular agents versus replacing existing compliance infrastructure.
Full transcriptexpand
0:10
Unidentified speaker: We tested nine different models, both with a castellum harness and with now, and we found that the top AM models, when they were asked to read policies and procedures, conduct screening, create an alert, and adjudicate an alert, they were accurate only 70% of the time without a harness, versus 99% of the time with a harness.
0:33
Unidentified speaker: So in other words, kind of for a compliance program, if you're using just Claude, you're at high risk of getting 3 out of 10 of your alerts wrong, which is kind of what gets any compliance officer fired immediately.
0:48
Unidentified speaker: This is the FinTech 1-on-1 podcast, the show for FinTech enthusiasts looking to better understand the leaders shaping fintech and banking today.
0:56
Unidentified speaker: My name is Peter Renton, and since 2013, I've been conducting in-depth interviews with fintech founders and banking executives.
1:04
Unidentified speaker: Today on the show, I'm delighted to welcome Peter Pietetsky, the CEO and co-founder of Castellum.ai.
1:12
Unidentified speaker: Peter began his career as an intelligence analyst in Afghanistan.
1:16
Unidentified speaker: He then moved to the US Treasury Department where he worked at OFAC on Iran, and later became the lead Russia analyst.
1:24
Unidentified speaker: After that, he ran US compliance from major agent banks, before founding Castellan in 2019.
1:30
Unidentified speaker: In our conversation we talk about why Castellan chose to be a primary source of sanctions, F and Adverse Media Data, and how it grew into building AI agents that resolve compliance loads.
1:43
Unidentified speaker: Peter walks through what actually happens when the load pilot explains why the real debate today is building maintain the supply, and shares benchmark results showing how much an LLM planet improves accuracy and consistency.
1:58
Unidentified speaker: We also discuss how the tilt then it's happened that criminals are using AI at scale and its prediction for L2 and L1 compliance work.
2:07
Unidentified speaker: Now, let's get on with the show.
2:18
Unidentified speaker: Welcome to the podcast, Peter.
2:20
Unidentified speaker: Thank you, Peter.
2:21
Unidentified speaker: I'm gonna enjoy saying that.
2:23
Unidentified speaker: Yes, indeed.
2:24
Unidentified speaker: It's quite nice to have another Peter on the show.
2:26
Unidentified speaker: So let's kick it off by giving listeners a little bit of background.
2:30
Unidentified speaker: I know you've had an interesting career, spent quite a bit of time in government before starting your your own company here.
2:36
Unidentified speaker: So give us some of the highlights of your career to date.
2:40
Unidentified speaker: My first real job was as an intelligence analyst in Afghanistan.
2:45
Unidentified speaker: That's a heck of a first job.
2:47
Unidentified speaker: It really is.
2:48
Unidentified speaker: It I think confused me as to what most jobs actually are, but that that was my first like W-2.
2:55
Unidentified speaker: And I really wanted to do national security and like fight the bad guys, and went to Afghanistan and it was supporting a special forces unit there, and it was unclear to me kind of what our why we were really there.
3:09
Unidentified speaker: Nobody could really tell me.
3:10
Unidentified speaker: And I think you'll you'll hear a similar story from anyone that was in Afghanistan, at least from from the US side.
3:17
Unidentified speaker: And I read about sanctions, and I was like, okay, you know what?
3:19
Unidentified speaker: This sounds great.
3:20
Unidentified speaker: You get to fight bad guys, and it's very focused.
3:25
Unidentified speaker: I mean, that was one of my challenges with our mission in Afghanistan was that it wasn't sure are we are we fighting terrorism?
3:32
Unidentified speaker: Are we building democracy?
3:33
Unidentified speaker: Are we training national security forces?
3:35
Unidentified speaker: Are we doing anti-drug trafficking?
3:37
Unidentified speaker: Like there was not one clear mission that we could win at.
3:39
Unidentified speaker: And so with sanctions, I thought, okay, here is something very clear.
3:43
Unidentified speaker: You know, there are, let's say, narco traffickers or proliferators or or governments or terrorist financiers, they all either need money or use money.
3:53
Unidentified speaker: So they either need money or they want money.
3:55
Unidentified speaker: Like narco traffickers, they do it for money.
3:57
Unidentified speaker: Terrorists generally don't do terrorism for money, but they need it.
4:02
Unidentified speaker: And so having that really narrow scope of the work, I think, was really energizing and motivating because I felt like I could measure success.
4:11
Unidentified speaker: Money frozen, aircraft impounded, like these are in many cases directly translates to lives saved.
4:18
Unidentified speaker: So worked in Afghanistan, then got a job with the Treasury Department first at OFAC, the sanctions office.
4:24
Unidentified speaker: And then Putin innovated, I was working on Iran, then Putin invaded Ukraine the first time, and they were like, hey, who who here speaks Russian?
4:34
Unidentified speaker: It's like me.
4:35
Unidentified speaker: So I they were like, okay, you are now lead Russia analyst and worked on uh Russian Eastern Ukraine, uh, Russia, Ukraine, uh, Eastern Europe, and then during the JCPOA back to Iran, but kind of in all of my government work, I found it very difficult to run investigations.
4:55
Unidentified speaker: And I I found it kind of very backwards that the US and other governments are spending hundreds of millions of dollars to identify bad actors and sanction them and say, hey, here, general public, here are these people that are very risky.
5:09
Unidentified speaker: But then actually finding that information was incredibly difficult, even for me while I was in the government.
5:15
Unidentified speaker: And so the origin of Costellum is very much so from my work in in government and my frustrations there.
5:23
Unidentified speaker: After I went to uh banking, ran US compliance for a major Asian bank.
5:30
Unidentified speaker: And I think it's kind of seeing how compliance technology was used across the private sector really convinced me that there was there was a lot of room to improve in terms of speed, in terms of accuracy, in terms of ease of usability.
5:45
Unidentified speaker: And you know, now uh here we are.
5:47
Unidentified speaker: Costellum is uh seven years old, and we have top four US bank, top three global payments provider, top five global crypto exchange, and we're really excited to be solving financial crime problems for clients.
6:01
Unidentified speaker: Right, right.
6:02
Unidentified speaker: So when you started, I mean you started started 2019, I believe.
6:06
Unidentified speaker: So that was before the AI revolution was upon us.
6:09
Unidentified speaker: And you know, you've certainly um, you know, embraced that.
6:14
Unidentified speaker: So tell us a little bit about what you were doing when you started and how that's evolved as we now live in a world of AI agents.
6:23
Unidentified speaker: So we we started as a as a data company, and we initially started with sanctions data.
6:30
Unidentified speaker: And then we you know would go to clients and they they would say, Okay, this is cool, but we need politically exposed persons data.
6:38
Unidentified speaker: And we added politically exposed persons data.
6:41
Unidentified speaker: And then you know, we went back to clients and they said, Okay, this is great, but we also need adverse media.
6:47
Unidentified speaker: So we went to add adverse media.
6:49
Unidentified speaker: And I think a this in initially felt like it was a mistake, but is now turning out to have been you know very fortuitous move.
6:56
Unidentified speaker: Is we did not outsource our data collection at any point.
7:01
Unidentified speaker: So we are a primary source.
7:04
Unidentified speaker: Same global sanctions that ourselves we build our own pep database.
7:08
Unidentified speaker: We get media directly from publishers, and then we turn it into adverse media.
7:13
Unidentified speaker: So we separate, you know, let's say Russian oligarch donates money to charity from Russian oligarch involved in corruption with construction, for example, and then identify who all the bad actors are, the company.
7:27
Unidentified speaker: So we are a primary data source.
7:30
Unidentified speaker: And I think what we learned is that the compliance data market is actually really extremely difficult to work in because we might go to folks and say, hey, our software and our data is gonna lead to 80, 90 percent efficiency improvements of no loss of accuracy, and people would say that's amazing, but then there's change management costs.
7:54
Unidentified speaker: And so it'd be like, okay, well, you just need to integrate and it's this easy, and then that's when people just start ghosting you.
8:00
Unidentified speaker: And so we decided to take the company in a modular direction.
8:04
Unidentified speaker: So instead of saying that, hey, you have to buy data, screening, case management, and agents, we separated out the modules, and that's really been huge for the company in terms of meeting clients where where they are, making it a lot easier to get the product.
8:19
Unidentified speaker: And the product that is really taking off right now is AI agents for alert resolution.
8:25
Unidentified speaker: And I think it's partially because we have narrowed down our focus, like some specifically to AI agents for sanctions, perhaps address media alerts, you know, level one, which is resolving the alert with the information available, and then level two, which is kind of going outside of the alert and conducting additional due diligence.
8:47
Unidentified speaker: And um it's been frankly like really, really exciting.
8:51
Unidentified speaker: When we founded the company, we thought, okay, this is something that we could do and it's possible, but now actually seeing our products do this is incredible.
9:01
Unidentified speaker: And I think it's there's an actual sea change coming in the world of of financial crime because no large organization actually knows what is going on inside of it simply due to the volume.
9:15
Unidentified speaker: And you have all of these systems that are designed to tee up alerts, but because organizations get fined for not enough alerts, not for too many alerts, when systems are configured, you know, program managers generally configure it to over-generate alerts because it they're trying to not get fined.
9:33
Unidentified speaker: Um, oftentimes more so than like, of course, they want to catch financial crime, but really what the biggest thing they're trying to avoid is being yelled at by the regulator, and the regulator telling them you're not trying hard enough.
9:42
Unidentified speaker: And so, as we all know, like investigators, analysts are drowning in mostly useless alerts.
9:48
Unidentified speaker: And so a lot of actual financial crime is not getting caught because of quite simple like volume issues.
9:55
Unidentified speaker: And I mean, AI agents are going to, at least for our clients, are already allowing them to remove, in some cases, 95 plus percent of alerts accurately, which is really important because inaccurately reducing workload is useless.
10:10
Unidentified speaker: But yeah, it's really allowing them to spend a ton of time on alerts that really matter.
10:15
Unidentified speaker: And that is phenomenal.
10:17
Unidentified speaker: So let's dig into that just for a little bit.
10:19
Unidentified speaker: Can you take us through the process when an alert fires at one of your clients?
10:27
Unidentified speaker: What actually happens then and how is sort of the AI agent involved and where is the human involved and when does that happen?
10:36
Unidentified speaker: Some of our clients use Costellum for data and screening, and then they use our agents, and other clients use different screening systems, uh, like let's say TRM or Bright Lexus Nexus Bridger or or Hummingbird with integrations with those and happy to talk about that more later.
10:52
Unidentified speaker: But let's say an alert is generated either in our system or in another system.
10:56
Unidentified speaker: And so what happens is that alert then goes to uh via API to our Arbiter endpoint.
11:04
Unidentified speaker: Arbiter is our family of AI agents.
11:07
Unidentified speaker: And the first thing is the alert gets routed to either a sanctions alert clearing agent, a PEP alert clearing agent, or an adverse media alert clearing agent.
11:16
Unidentified speaker: So because we started as a data and screening service company, we have a lot of IP and knowledge around how alerts are resolved.
11:26
Unidentified speaker: And you know, there are very different risk tolerances between the same alert that matches on a sanctions name is going to have a much higher threshold for clearing it than the same alert but the matches on adverse media.
11:40
Unidentified speaker: Uh so before we review our agent reviews any alerts, our agent ingests the client's policies and procedures.
11:47
Unidentified speaker: And sometimes these are 13 pages, other times they're 145 pages.
11:53
Unidentified speaker: And usually before we even sign a contract to the client, we go through a full proof of concept where our agent learns the client's policies and procedures, then does an open book test that's alerts with adjudications, and then we can see, hey, is our agent hitting 100?
12:10
Unidentified speaker: Uh, interestingly enough, our agent almost always identifies cases where the human analysts whose work we are kind of being graded against have made mistakes that don't correspond with the policies and procedures.
12:22
Unidentified speaker: So even during kind of the POC phase, we're identifying improvements, but we our agent ingests the policies and procedures, our agent takes the open book test, and then our agent takes a closed book test, which is alerts without the adjudications.
12:35
Unidentified speaker: So we go through this training period, which we do in 10 business days, and then the alerts go to an already trained agent.
12:43
Unidentified speaker: And so if it's a sanctions, PAPS Adverse Media Alert gets routed that way, and then each one of those agents has dozens of sub-agents.
12:52
Unidentified speaker: So there's, for example, a name normalization.
12:55
Unidentified speaker: There's a text normalization agent that will clean up the alert if it's kind of dirty and unreadable or might have a typo in it.
13:02
Unidentified speaker: Then there's a name normalization agent.
13:04
Unidentified speaker: Sounds really silly, but for example, is Le Thong Wei written in Mandarin characters the same as Lee Thong Wei written in English?
13:13
Unidentified speaker: Is Mohammed, let's say, Bastrukayev in Russian the same as Mohammed Bastrukayev spelled differently in English?
13:20
Unidentified speaker: So kind of difficult transliteration phonetic analysis, location analysis, contextual analysis, severity analysis.
13:28
Unidentified speaker: So we do not have we do not take an alert and just send an alert to an LLM and then say, hey, what do you think?
13:37
Unidentified speaker: That leads to very low accuracy and very low consistency.
13:45
Unidentified speaker: And that in our space, that's that's unacceptable.
13:47
Unidentified speaker: So that's a short answer.
13:48
Unidentified speaker: I know that was like five minutes, but the agent learns the policy and procedures, the agent gets tested, and then when alerts get sent to the agent, we really break down the alert into um as many different components and evaluate each one of those separately and then come back with an answer.
14:04
Unidentified speaker: So as you're talking, it seems to me that every single financial institution needs this technology.
14:11
Unidentified speaker: But I'm curious about who is your market?
14:15
Unidentified speaker: I mean, who are you focused on?
14:17
Unidentified speaker: Are you focused just you know on the sponsor banks, on the fintechs, on the computer lead banks, on everybody?
14:22
Unidentified speaker: What's the market?
14:24
Unidentified speaker: We focus on US-based banks and fintechs.
14:28
Unidentified speaker: We are primarily trying to work with mid-market banks and fintechs.
14:35
Unidentified speaker: But in in practice, we have seen a fair amount of interest from tier ones.
14:41
Unidentified speaker: And so that is we're not going to say no to you know top 10 global bank or top four US bank knocking on our door.
14:48
Unidentified speaker: So it's essentially in practice, we work like mid-market and up.
14:52
Unidentified speaker: But what we've seen is in one year, the space has gone from AI agents are not reliable enough to do this to I need AI agents to do this now.
15:03
Unidentified speaker: And the only question is do I have budget for it this year or do I need budget for it next year?
15:09
Unidentified speaker: And we have seen this from folks that were AI skeptics.
15:13
Unidentified speaker: We have seen this from folks that have laid off people and rehired them.
15:17
Unidentified speaker: And we have seen this from people that have been kind of very bullish on AI use cases all along.
15:22
Unidentified speaker: And I think what's part of what is AI tools have gotten exponentially better.
15:27
Unidentified speaker: But also we've been able to really show, and some other companies in the market as well, we've been able to show that, hey, agents can clear alerts reliably.
15:37
Unidentified speaker: Right.
15:37
Unidentified speaker: That is the key because I think that was the knock early on on the some of the AI technologies you said and and hallucinations with LLMs has made it a lot of people think that there is a reliability issue.
15:51
Unidentified speaker: So when you're talking with potential clients, is that is that reliability piece central or is that debate kind of falling by the wayside now?
16:02
Unidentified speaker: The reliability debate has actually fallen, is gone.
16:07
Unidentified speaker: We have very few conversations where people say, I wouldn't trust AI to do this.
16:14
Unidentified speaker: And if someone says that they don't trust AI to do this, we say that's totally fine.
16:18
Unidentified speaker: We're gonna go talk to the next person because we have so much demand right now, which is you know a great place to be.
16:25
Unidentified speaker: And the question that we'd have much more often is why should I not use Claude?
16:31
Unidentified speaker: And I think that's a great question.
16:33
Unidentified speaker: And it's a very reasonable and smart question.
16:35
Unidentified speaker: And I think we also, you know, whether it's Claude or or Astor or Copilot, a lot of institutions have you know enterprise AI accounts with really large budgets, and they're saying, look, why don't we just build it?
16:48
Unidentified speaker: And so the question that we mostly have is what we used to call build versus buy, but what is much more accurate to be called build and maintain versus buy.
16:59
Unidentified speaker: And so we have done a fair amount of really detailed data-driven AI research showing what it is like to use an LLM and what it is like to use an LLM with a harness.
17:13
Unidentified speaker: A harness is a combination of infrastructure, data, regression testing, QAQC, specialized instructions, and more that is put around an LLM to make it purpose built.
17:27
Unidentified speaker: And for those that follow uh Formula One, there's a couple of companies that are engine manufacturers, like Mercedes manufactures engines for itself, but also for other cars, for other teams.
17:39
Unidentified speaker: But then the teams that buy the engines from Mercedes, they still have varying results than Mercedes, and they will put on different tires, they have a different driver, different wings, brakes, etc.
17:52
Unidentified speaker: So using you know uh anthropic or open AI as kind of the engine, everyone should be using an engine.
18:00
Unidentified speaker: But then the harness around it is actually what makes the difference.
18:03
Unidentified speaker: And so for financial crime compliance, that harness includes specialized data, specialized training, specialized regression testing, built-in QAQC, and the results are enormously different.
18:18
Unidentified speaker: So we tested nine different models, both with a Costellum harness and without, and we found that the top AI models, when they were asked to read policies and procedures, conduct screening, create an alert, and adjudicate an alert, they were accurate only 70% of the time without a harness versus 99% of the time with a harness.
18:42
Unidentified speaker: So, in other words, kind of for a compliance program, if you're using just Claude, you're at high risk of getting three out of ten of your alerts wrong, which is kind of what gets any compliance officer fired immediately.
18:56
Unidentified speaker: Like that's an that's an unacceptable number.
18:58
Unidentified speaker: Like three out of a thousand would get someone in trouble, three out of ten.
19:02
Unidentified speaker: And so we found that the harness greatly improved accuracy, but it also really made consistency close to 100%.
19:12
Unidentified speaker: So if you give an agent the same alert three times, is it going to adjudicate it the same way?
19:17
Unidentified speaker: And if we're just in a LLM by itself, the answer was no, it would not.
19:23
Unidentified speaker: So LLMs are incredibly smart and really powerful, but they absolutely cannot be trusted to just be given a financial crime compliance problem and to solve it.
19:35
Unidentified speaker: They need to understand the policies and procedures of each institution, they need to understand the risk metrics, and the harness is really what does that.
19:43
Unidentified speaker: And so we never say what you know, we work with some very large fintechs that have engineering teams bigger than our whole company.
19:51
Unidentified speaker: So I never I would never say to those people you couldn't build it, but of course they could.
19:56
Unidentified speaker: It's a question of do they want to build it, maintain it?
20:01
Unidentified speaker: Are they going to staff you know their agent with dedicated compliance engineers and dedicated compliance subject matter experts?
20:08
Unidentified speaker: Are they going to do the model validation?
20:10
Unidentified speaker: Are they going to do the regression testing?
20:12
Unidentified speaker: Every time a new model comes out, you know, whatever is going to power the agent on the back end, it has to be retested on hundreds or thousands of cases and alerts.
20:20
Unidentified speaker: And so we've actually found the really interesting kind of allies amongst the general counsel's office where they'll say, like, look, we know my team at you know, named Ding SinTech can build this, but do we want to own the risk?
20:35
Unidentified speaker: Also, it's way more expensive for us to build it and maintain.
20:38
Unidentified speaker: So really the question is build and maintain or buy.
20:43
Unidentified speaker: And that's kind of the debate we're having more often now.
20:46
Unidentified speaker: Everyone has existing sanctioned screening and a whole range of compliance tools at their disposal today that had them for decades in many cases.
20:56
Unidentified speaker: When someone's deploying your technology, what kind of improvements can they expect?
21:03
Unidentified speaker: And I know I know I know you you can you know it's going to be different for every every system out there, but give us some sense of when Castellan comes in, what happens as far as improvements to their systems typically?
21:17
Unidentified speaker: I'll talk more so about our partnership motion.
21:19
Unidentified speaker: So we have our agents are integrated into a number of different systems, including LexisNexis Bridger, LexusNexis Firkasoft.
21:27
Unidentified speaker: We just announced a integration with TRM and we have an integration with Hummingbird.
21:32
Unidentified speaker: And there are some others.
21:33
Unidentified speaker: So when someone adds our agents over the top of an existing screening service, the first thing that they can expect is not to change their existing screening service.
21:44
Unidentified speaker: That's great from a change management perspective.
21:47
Unidentified speaker: And the way I like to think about this is you know if someone comes to you and says you have lead in your pipes in your house, you can't drink that water anymore, and there's a number of solutions.
21:57
Unidentified speaker: And one of the solutions is you rip out all the pipes and you replace them.
22:00
Unidentified speaker: That's very expensive.
22:02
Unidentified speaker: You're gonna have to live in a hotel for a little bit.
22:04
Unidentified speaker: And the Costello approach is hey, just buy this water filter and put this water filter onto your tap, it'll filter everything out.
22:10
Unidentified speaker: You can drink clean water.
22:12
Unidentified speaker: So that is kind of our agent approach with existing screening systems.
22:16
Unidentified speaker: Our agent runs inside of Bridger and inside of Ferkasoft.
22:21
Unidentified speaker: So what that means is clients can continue to use Bridger, Ferkosoft, or other systems that they have in exactly the same way they were.
22:29
Unidentified speaker: No new model validations, no new tuning reports, and then just alerts are handed over to the agent.
22:36
Unidentified speaker: And this is kind of an approach of just saying, let's let's be realistic about what can be done, about budgets, about change management appetite.
22:48
Unidentified speaker: This is a way to focus on on results first.
22:51
Unidentified speaker: Like we're not gonna say that, hey, we're gonna do this massive reimagining of a whole compliance program.
22:56
Unidentified speaker: Usually whoever starts that tends to leave before it's finished.
23:00
Unidentified speaker: We're saying those are like three to five year projects at a minimum.
23:03
Unidentified speaker: But if someone, let's say, has Firkosoft or Bridger and they want to get immediate results, cool.
23:09
Unidentified speaker: Our agents are already inside, it's like turning it on.
23:12
Unidentified speaker: And that's really our focus right now is on supporting clients through our channels.
23:18
Unidentified speaker: As we were saying before, we hit record.
23:20
Unidentified speaker: I was at Money2020 this week, walking the halls of the show there.
23:25
Unidentified speaker: There's there were some you know AI companies talking about it, um, you know, sanction screening and different alerts that their agents can do.
23:33
Unidentified speaker: So you're not the only game in town.
23:36
Unidentified speaker: And it's difficult, I think, for people to know what's who's better than who's and who's really not.
23:43
Unidentified speaker: So with someone who's listening to this and sort of say, well, how do I evaluate these companies that are all say they're doing something similar?
23:52
Unidentified speaker: How do we tell you guys apart?
23:54
Unidentified speaker: You know, last year I have a FinTech meetup and a bank CEO telling me like literally that.
24:00
Unidentified speaker: I don't know how to tell you guys apart, uh, with some words in there.
24:04
Unidentified speaker: And we have a couple of re resources which I would love if you're able to direct your listeners to.
24:11
Unidentified speaker: We have an agentic screening market map.
24:13
Unidentified speaker: So we have done a bunch of analysis and put together a number of vendors, I think over over a dozen that really focus on this.
24:22
Unidentified speaker: So we're not including you know legacy company that bolts on agents or somebody that is just white labeling like ChatGPT and sending alerts directly to a public unencrypted LLM service.
24:35
Unidentified speaker: It's who is really focusing on on agents.
24:39
Unidentified speaker: And so we encourage everyone to take a look at the Costellum Agenc screening market map.
24:45
Unidentified speaker: That's one that helps people understand like who's in the space.
24:48
Unidentified speaker: You know, obviously someone could say that we're we're biased, but we do put a bunch of our competitors uh up there and say good things about them.
24:55
Unidentified speaker: And I would also say I don't view us other agentic companies as really competing with us.
25:03
Unidentified speaker: I see it as we're all going after the same problems and we're really competing against like BPOs and consulting firms kind of that are just used to out outsource this work somewhere.
25:13
Unidentified speaker: So I mean the agentic screening market map is one, and then we have a FinCrime agent benchmark, which I mentioned earlier, where we actually show research.
25:21
Unidentified speaker: Like, look, this is what our agents have done.
25:24
Unidentified speaker: This is how a LLM without a harness will perform.
25:28
Unidentified speaker: And I Peter, I think you're right.
25:30
Unidentified speaker: The hardest thing right now to do in this space is to stand out.
25:34
Unidentified speaker: Well, I'll I'll make sure we do link to that in the show notes here.
25:37
Unidentified speaker: Okay, so I want to talk about the other side, the financial criminal side of things, because they're they're using AI at scale like like never before.
25:46
Unidentified speaker: They don't have to deal with any compliance uh rules or anything like that.
25:51
Unidentified speaker: What are you seeing on the offensive side that was not there, say two or three years ago, that that where the criminals are taking advantage of the of the latest tools?
26:00
Unidentified speaker: So it's uh horrifying.
26:02
Unidentified speaker: And I mean that's actually one of the reasons that regulators have really come around to AI incident crime is because they are seeing criminals use it extremely well.
26:14
Unidentified speaker: And so regulators are basically initially had a lot of discomfort around the hallucinations and is AI reliable, and now it's just more so like there is a deluge of AI-powered financial crime, and so the only way to really fight back is is to use AI because AI allows you to done well, allows you to scale up your fin crime operations or anti-fin crime operations, but also allows criminals to scale up their fancrime operations.
26:43
Unidentified speaker: So in terms of what what we're seeing, the first is scale.
26:46
Unidentified speaker: So any scam that was occurring beforehand or crime that is a was occurring beforehand is now occurring, you know, at like 10 to 20x in terms of volume.
26:56
Unidentified speaker: I think we've we've all heard about scam farms, like in especially in Southeast Asia, where people are kidnapped, passports taken away, uh, they're told they're going to work somewhere, and then they're put into this call center, like where they have to commit elder fraud.
27:10
Unidentified speaker: And those centers don't need as many people anymore.
27:14
Unidentified speaker: They're able to run the same scams with fewer people.
27:16
Unidentified speaker: So I think one is just volume, two is fake IDs and fake personas.
27:24
Unidentified speaker: So it's just become that much easier to create fake documentation that people can create a deep fake online.
27:32
Unidentified speaker: Like you know, unless you're you're touching someone, you may not know that like that the person that is there is the person who is there.
27:40
Unidentified speaker: And that creates a lot of issues for account opening, wire transfers, theft of funds.
27:47
Unidentified speaker: And so I mean that those are the two primary attack factors that that we're we're seeing.
27:51
Unidentified speaker: One is just, and it's not so much an attack factor, but like all the existing scams are now superpowered in terms of volume.
27:59
Unidentified speaker: And then there's a lot in terms of account opening that is that is being done that is in the financial crime world because people are creating like synthetic personas, complete synthetic personas, using them to open accounts and then commit money laundering, sanctions, evasion, or just fraud.
28:17
Unidentified speaker: Okay, so last question then.
28:19
Unidentified speaker: As you think about the future and how AI agents are gonna uh continue to reshape compliance programs.
28:27
Unidentified speaker: Take us through like where you think, you know, two to three years out, what the landscape will look like, and where does Costellum fit into that landscape?
28:37
Unidentified speaker: Two to three years out, Costellum continues to be the leading provider of AI agents to banks and fintechs in the US and knock on wood.
28:48
Unidentified speaker: But I think where we see the space going is that all work that is known right now as L1 has been fully automated.
28:57
Unidentified speaker: So L1 is an alert is generated, and then an analyst looks at the alert, but they're only looking at like what is in the alert.
29:04
Unidentified speaker: They are not doing external research.
29:06
Unidentified speaker: They're just like Peter Pioteski matched to Peter Pazatsky, false positive, name mismatch.
29:12
Unidentified speaker: And the reason, again, those alerts are generated in the first place is because compliance program owners tend to bias towards let's generate more alerts so we don't get fined.
29:20
Unidentified speaker: Because if we miss something, that's when we really get in trouble.
29:24
Unidentified speaker: And so with agents, you can continue to generate a fair number of alerts, but reviewing them becomes much less expensive.
29:31
Unidentified speaker: So I think L1 as a job does not really exist anymore.
29:37
Unidentified speaker: And most of those analysts are now going to be working on L2 alerts.
29:42
Unidentified speaker: So things where you need to do some external research and analysis.
29:46
Unidentified speaker: Now I think a lot of L2 work is also going to be automated because that is basically taking an L1 alert where there is lack of clarity and then going out to certain external sources to do research, come back and help make a decision.
30:01
Unidentified speaker: So that is something that agents are very good at doing, like gathering data, creating a decision point.
30:07
Unidentified speaker: But I think what this means also is that I do believe we will see, depending on how this technology is implemented, we could see a real true global drop in in financial crime because you would actually see some of the compliance shops at large banks in fintechs become very effective.
30:28
Unidentified speaker: And for someone to be able to say, hey, I actually know what's going through what going on in my bank fully.
30:35
Unidentified speaker: That's something that's never been possible for a large bank.
30:39
Unidentified speaker: Like that's only been possible for a private bank where maybe everyone, the bankers all know know all the clients, similarly to much more what you would see at a credit union, but even credit unions are are trying to grow now.
30:49
Unidentified speaker: And you know, many you have almost all have online account opening and don't know all the members.
30:55
Unidentified speaker: So long story short, uh, Peter, I think we are going to see, you know, near full automation of L1 and L2 roles in the space, but also a corresponding rise in effectiveness in fighting financial crime.
31:10
Unidentified speaker: Well, that that is a an optimistic and a positive future.
31:13
Unidentified speaker: Great place to end.
31:14
Unidentified speaker: Um, Peter, really appreciate you coming on the show.
31:17
Unidentified speaker: I hope uh that your prediction comes true and and everyone is much more effective at uh at fighting financial crime.
31:24
Unidentified speaker: So best of luck and thanks again.
31:26
Unidentified speaker: Thank you, Peter.
31:27
Unidentified speaker: Pleasure to be here.
31:34
Unidentified speaker: I loved Peter's framing of the impact of compliance AI when it comes to large financial institutions.
31:41
Unidentified speaker: He said today that none of them truly know what is going on inside their company.
31:46
Unidentified speaker: Only a small private bank with the bankers know every client has ever had that kind of visibility.
31:51
Unidentified speaker: Peter believes AI agents could finally give a top-tier bank the same understanding, but this could lead to a real drop in financial profit.
32:01
Unidentified speaker: That is a bold prediction, but it changes how we should think about AI and compliance.
32:05
Unidentified speaker: This is not just about cutting the cost of L1 reviews.
32:09
Unidentified speaker: If he is right, compliance teams will finally get to do the job they were always meant to do.
32:15
Unidentified speaker: Anyway, that's it for today's show.
32:17
Unidentified speaker: If you enjoyed this episode, please go ahead and subscribe, tell a friend, or leave a review.
32:22
Unidentified speaker: And thanks so much for listening.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3 and interpreted by Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights