How TBC Bank Uses AI and Lakebase to Stop Money Laundering

Summary

TBC Bank consolidated fragmented AML tools onto the Databricks Data and AI platform, building a unified AI investigation environment that addresses a 95% false positive rate and escalating regulatory requirements including DORA and perpetual KYC.
This video showcases four production use cases: a counterparty AML search engine across OFAC, EU, UN, and open-web adverse media sources; an AI engine comparing invoices, contracts, and customs documents; a transaction purpose scanner; and Genie for natural-language analytics.
The architecture uses Databricks Lakebase to manage agent state and investigation context, with Databricks Apps delivering all capabilities through a single investigative workspace, representing a shift from reactive monitoring to contextual AI-assisted investigation.

How TBC Bank Uses AI and Lakebase to Stop Money Laundering

Anti-money laundering (AML) teams are buried under fragmented tools, scattered data, and manual workflows. Adding more rules has not solved the problem. TBC Bank took a different path, consolidating intelligence on Databricks to create a unified, AI-powered investigation environment. this video shows the production use cases in action: a counterparty AML search engine across sanctions, internal, and external watchlists; an AI engine comparing invoices, contracts, and customs documents; a transaction purpose scanner detecting dual-use and restricted goods; and Genie for natural-language analytics. All built on a shared Lakehouse and delivered through one investigative workspace using Databricks Apps and Lakebase. The result is a fundamental shift in how AML works at TBC Bank: from reactive monitoring to contextual, AI-assisted investigation.
Talk By: Istvan Viz, Account Executive, Databricks ; Konstantine Khoperia, Chief Data & AI Officer, TBC Bank ;

Chapters

FAQs

How does TBC Bank use AI to fight money laundering on Databricks?

TBC Bank built an AI-powered investigation environment on the Databricks Data and AI platform that unifies watchlist searches, document comparison, and transaction analysis into one workspace. The system pulls from OFAC, EU, UN, and open-web adverse media sources to give investigators complete counterparty context.

What is the AML false positive problem and how does this platform address it?

Traditional AML systems generate false positive rates as high as 95%, burying analysts in irrelevant alerts and manual reviews. By consolidating data and applying AI models on Databricks, TBC Bank shifted from rule-based monitoring to contextual, AI-assisted investigation that surfaces only meaningful signals.

What are the four production AML use cases shown in this video?

The four production use cases are: a counterparty AML search engine across sanctions and watchlists, an AI engine that compares invoices, contracts, and customs documents, a transaction purpose scanner for dual-use and restricted goods detection, and Genie for natural-language analytics queries over financial data.

What is the role of Databricks Lakebase in the TBC Bank AML platform?

Databricks Lakebase is used to manage agent state and investigation context, providing the persistent memory layer for the AI investigation system. It enables the platform to maintain continuity across multi-step AML investigations within a single unified lakehouse workspace.

Full transcript

[00:08] Let me ask a question. How is everyone? Exhausted a little bit. Waiting for the party. Good, good, good. Don't worry. Couple of hours left and you can have some drinks. Okay. But before that, we have to talk
[00:20] about a very important topic today, right? Which is going to be antimony loungering. Okay. But before we would get to the details, I'm bringing here a very important message. It's actually a
[00:33] love message from our lawyers. So please make sure that you make your purchasing decision according to the available technology, according to the available features. Okay? You can download this material. You don't need to read this.
[00:46] And one more ask, please please give us feedback. Okay? Give us a thumbs up, thumbs down if you like the session. If you don't like the session, if you have anything what you would change because this is actually helping us as
[01:00] bricksters to improve the next data and AI summit. Okay, really appreciate it. You can do this after the session. You don't need to do it now. Okay. This is my actually privilege to really
[01:12] walk you through about a very exciting story. Okay. how a bank from Eastern Europe, from Georgia, is actually utilizing latest technology, utilizing data and AI for actually stopping the
[01:24] money laundering. Okay. And we're going to spend the next, let's say, 40 minutes here together. And our goal is really to educate you, evangelize you, and a
[01:36] little bit inspire you about how a bank from another part of the world is utilizing data and AI. And actually this is not a story about
[01:48] a technology sake, right? This is really an important story, right? Because the stakes are real. The stakes are all about protecting the customer, protecting the consumers, protecting the brand, and really protecting everything
[02:00] which is around a very important um financial operation. And uh let's face it, fighting against financial crime is no longer a compliance story. is no long
[02:12] no longer a compliance process. This is really a data and AI challenge. Okay. Uh I would like to bring you a little bit of context around what's happening in Europe. You can see the
[02:25] numbers but uh what is actually important to mention that it's all about the numbers right? It's all about you know how other banks are actually trying to protect against um against you know
[02:39] financial crime because they are managing sensitive data. They are actually managing personal data and hence banks are really investing in in such topic and what's the scale of the problem. It's very simple. This is not
[02:52] just a risk issue anymore. It's all about customer experience now. It's all about growth. It really touches reputation. Right? As you can see the numbers one of my favorite is like 50% of
[03:05] customer abondment when onboarding takes more than five minutes just imagine I mean you are also consumers right if you need to take more than five minutes for onboarding you might find another bank
[03:18] right or 50% of bank headcount dedicated for KYC right KYC activities 90 days of average time to open a new corporate account this is a little bit tricky I
[03:30] would say and $114 billion is the annual EMIA financial crime compliance spent. This is a huge number as you can see and yeah 95% false positive rate in AML
[03:43] alerts. This is also something which is which is uh which is very important to me mention and the message is clear. Message is clear. This is not a back office problem anymore. This is actually nowadays a frontline challenge. A
[03:55] frontline challenge which should be really handled. Just think about it. If onboarding is slow, right, revenue will suffer. If controls are weak, there's going to be fines, right? There's going to be reputation damage. There's going
[04:08] to be uh lots of other issues. And if you know, if false positives are too high, your most talented people are chasing a noise instead of a risk. So money laundering is really a compliance
[04:21] problem. It's not a compliance problem anymore. It's a trust and a growth problem. And in Europe there is a regulatory tsunami. Hope it's not a problem that I
[04:34] call like this but uh we started like from 2025. Yeah. Dora applied for operational resilience or there's a buffing updated AML guidance or there is anti-money laundering authority based
[04:48] out in Frankfurt which is operational since uh since July of 2025. Which means that all of those facts are showing us that the pressure the pressure is rising around the banks. The
[05:01] regulation is actually accelerating. There are unified regulations around around the world. Let's put it this way. There are mandatory AML protocols and quick info sharings are required from
[05:14] banks. And there is also a push from regulators to use or to utilize perpetual QYC. Perpetual KYC meaning you need to actually check your customers
[05:27] constantly not just while you are onboarding them. What does it mean in a practice for you? It means that the old model of periodic review and the fragmented tooling is just simply running out of the road and the manual
[05:41] processes are actually slowing down bunks down. And this is this is exactly why data and AR are becoming central around the banks
[05:54] all over the world. And this actual deck shows plain out across fraud detection risks or or personalization or utilizing alternative data data is really crucial nowadays and anti-moneyaundering is at
[06:08] the center of the trend. It requires more intelligence, faster actions which are actually creating value, business value at the end of the day. And there is actually a shift there is a shift
[06:20] from can we use AI around this topic to the way to how to use AI, how to use AI to accelerate all of the processable processes in a governed and auditable and operational way.
[06:34] And that brings us actually to the architecture, right? Banks are following the lakehouse architecture. Lakehouse pattern which actually giving them flexibility of managing structured and unstructured data defining path
[06:48] defining patterns right defining patterns with machine learning and AI capabilities. Lakehouse is bringing together the entity extraction, the entity resolution and also entity contextualization
[07:02] which are always required. Just think about when you need a real-time analysis around the transaction data around the swift swift messages or when it comes to resolution you want to also validate
[07:14] entities with the knowledge knowledge graph with AI or when it comes to entity contextualization you have to follow the money you have to understand where the money is coming from and where the money is supposed to
[07:27] go right you have to identify those trends in any money laundering AI should not replace judgment Right. AI should really support uh your investigators, your in your guys who are
[07:39] actually working on investigation. It should really accelerate judgment and reaching the line of defense with AI with AI connecting the dots around graph and finding patterns with machine
[07:53] learning. And actually if you would think like this is really theoretical the headlines say otherwise. All of the news are around like banks
[08:07] from bank for from Europe like there is a Swiss bank who was forced into liquidation over money laundering allegations. So those those um headlines are really proof like how enforment
[08:20] trends are actually appeared in in in in Europe right. It's it's actually showing that regulator regulators are prioritizing sanction compliance or beneficial ownership transparency and
[08:32] AIdriven transaction monitoring again following the money right also this slide is actually showing you that major banks they have huge amount of fines right and they're actually really having
[08:47] an issue around reputation because of all of these trends and also Um I just had a chat with uh one of the bank in Europe. They're really suffering with um
[09:00] uh with with false positives right so they are investing huge amount of time for the people who are actually following the noise instead of really real frauds and that's why datadriven risk
[09:12] assessments are are necessary. Banks are nowadays they are not judged what they catch. They are actually get judged how strong are their systems before
[09:24] something goes wrong. It is actually my privilege to to introduce you TBC bank. Okay. TBC bank is Georgia's largest bank based out in
[09:38] TBC and fast growing business in Usbakistan. Okay. They are actually combining the traditional banking so like current accounts, saving accounts and cards with other let's say ecosystem
[09:50] products like e-commerce, mobility and other e other business services and I'm really really enjoying working with with them because I feel that they are one of the most innovative
[10:02] techdriven bank in the other part of the world inevenia I would say they are actually utilizing the latest technology uh from data and AI. They're utilizing uh uh data and AI not just for
[10:15] anti-money laundering also uh to gain competitive advantage. Right? This is actually uh something which which which is super awesome. I would say they have re rebuilt their core
[10:27] anti-moneyaundering capabilities with low latency and with transactional access. Fun fact, they give actually workspace with AI assisted analytics and natural language exploration for all of their
[10:40] investigators. So they are really a great proof that they are really accelerating the way how they can protect you know their business against all of the financial crimes
[10:52] and I would like to really welcome on stage Constantine who actually traveled over you know from Tibilis right uh by the way anyone have been had been traveling to Tibilis anyone knows
[11:05] Georgia great food great wines >> great >> you liked it okay the Rest of the people, please, please visit Georgia one day. You will enjoy their hospitality big time. Okay, welcome
[11:18] Constantine. >> Thank you. Thank you. Good luck. >> Thank you for the introduction. Happy to be here. Actually wanted to
[11:30] start it with the words like super excited, but it seems that super excited is owned by data bricks. So you need a special permission to say these words. So happy to be here and um let me start
[11:44] with um uh why AML had to change in the first place.
[11:56] So uh there was a time when investigations were relatively simple. You could work with a short transaction purpose field, do a manual lookup in a single source, rely on limited data
[12:10] sources, or run rule-based uh alerts on transactions. In many cases, investigations could have been completed in a couple of minutes, but
[12:23] that world was gone. Today, investigators are dealing with thousands of words across documents, invoices, contracts, uh, customs documents,
[12:36] multiple sanctions and watch list sources, complex counterparty networks, uh behavioral history, and the need for cross referencing analysis. What used to
[12:48] take minutes now can take hours and sometimes even days. This is why uh AML is no longer a checkbox compliance exercise but rather
[13:03] it became uh real time data analysis challenge uh which requires speed breath and precision precision which manual
[13:17] processes just simply cannot handle. So the key point is that uh complexity did not increased a little it increased
[13:29] structurally. All the anal was designed for simpler cases, slower signals and narrower context. But when complexity explodes, the real
[13:43] bottlenecks become very clear. So what was the real bottleneck? It was not a shortage of rules. It was the fact that investigators
[13:56] had to work across multiple systems with no unified view switching between screens just to piece together a single case. At the same time, invoices, contracts, lending
[14:10] bills, custom documents were being uh reviewed by hand which made the process slow, inconsistent and errorprone. So critical challenge remained siloed
[14:23] across uh multiple sanction lists, internal watch lists and external sources with no real cross referencing layer. That's why slide calls out for
[14:36] scattered tools, manual processes, manual document chats and disconnected uh sources. The real issue was fragmented intelligence uh context scattered across
[14:51] different tools, documents and data sources that AML officers had to stitch together under time pressure for every single investigation. And when
[15:03] context is fragmented, three things happened. Investigations slow down. Uh quality becomes inconsistent and the strongest analysts spend their
[15:15] time assembling evidences instead of applying judgment. So the problem was never too little data. The problem was too little
[15:28] connected data. And that's exactly why we didn't wanted to add more rules. We wanted to build an intelligence layer.
[15:44] And let me continue with the principle that shaped everything we built. The answer as I mentioned was not to add more rules. It was uh to an already overstretched system. It was uh the
[15:56] strategic decision that we decided to build an intelligence layer, a unified platform that puts together context in front of investigators instantly powered by data bricks. The
[16:10] intelligence layer is made up of five parts. First is uh the lakehouse as the governed and unified data layer. The second is lake base as the search and
[16:24] retrieval layer. Third one is datab bricks apps as an operational interface for the ML officers. Fourth
[16:38] single uh investigation window. Uh and just to keep in mind this multibot is again the product of datab bricks apps.
[16:50] So each of these components plays a distinct role but together they form one coherent AML intelligence workspace that replaces fragmented tooling with a
[17:03] governed and AI powered investigation environment. The key point is this. We didn't try to automate around fragmentation. We removed fragmentation first and then
[17:16] proceeded. And once you start from that design principle, the architecture becomes much easier to understand.
[17:30] So what you're looking at on the slide uh is the operating backbone of uh the intelligence uh layer. We bring together multiple sources like
[17:42] databases, web services, uh files and uh real time sources. Those flows come through components like event hub for the real time or near realtime data
[17:56] sources and uh Azure data factory for batch processing which later goes into the lakehouse where we manage the pro information across bronze, silver and uh
[18:10] gold layers for the further processing. From there we support uh historical analysis, machine learning checks and AI agents and then expose that intelligence
[18:25] through Genie search and unified uh unified interface built with lakebase and datab bricks apps. So why does this matter? Because ML is no longer a one
[18:40] single workflow. It's a chain of workflows, ingestion, enrichment, retrieval, analysis, investigations and uh outcomes decisioning.
[18:54] So if uh those steps live in a separate silos uh you lose time at every handoff. What we wanted was one governed architecture where data layer, AI layer
[19:08] and the investigation experience all reinforce and support each other. If your architecture is fragmented, your investigators are fragmented as well.
[19:20] And uh now let me take you further to uh the concrete uh to make let me make it this concrete for you through actual use cases. And the use case number one is
[19:34] counterparty AML search engine. Uh what gets searched in one one query? Um, we search internal watch lists and flag entity registers, sanctions lists
[19:49] including OFAC, EU, UN, um, and internal uh, local lists, do risk intelligence database, open web and uh, adverse media
[20:02] sources and uh, historical transactions and counterparty behavior. This means that investigator no longer has to jump from one system to another uh to answer a very basic
[20:15] question. Who is this counterparty and what do we already know about them? The key outcome is moving from multi-system uh manual lookup to one governed search with a dramatically faster counterparty
[20:30] risk assessment and consistent auditable results every time. Technically this process is uh powered by lakebase for unified search apps as a UI layer
[20:47] and govern data in uh lakehouse. This is one of the moments where a very simple uh user experience uh hides a lot of engineering discipline behind. The
[20:59] experience looks like one search bar. The value uh is the answer comes back governed, contextualized and ready for action.
[21:11] Search is not just a retrieval. It is uh in IML searches risk compression. The second use case takes us from uh search to document intelligence part. Uh this
[21:23] is where the value of AI becomes really very visible. uh invoices, contracts, custom documents um uh are ingested and parsed with by AI engine. The system
[21:38] then compares fields, values, parties, dates, uh products, volumes, numbers, etc. uh across all documents simultaneously. So uh inconsistencies, discrepancies and
[21:54] um suspicious patterns are surfaced instantly with evidence trails and structured findings are delivered uh directly to the AML officer through the multibot interface. What used to require
[22:07] a specialist or analyst uh spending hours on document review can now be completed in a fraction of time with greater consistency and full uh audit
[22:20] trail maintained uh in the platform. Uh the matter this matters and uh this matters because uh document heavy uh investigations are exactly where manual
[22:33] effort becomes expensive, inconsistent and difficult to scale. Human expertise is still essential. However, um now humans are focused more on judgment not
[22:45] on um mechanical comparisons. So they start stop being integrators and that's the very good thing for them. The goal is not to read uh fewer documents. The
[22:59] goal is to extract more truth uh from them faster. The third use case addresses uh the hardest inputs in AML which is messy free text. And uh you
[23:14] agree that payment descriptions are really messy free text. They are often ambiguous full of abbreviations or deliberately obuscated and manual review
[23:27] in scale is simply impossible. So we built a transaction purpose scanner. It applies AI language models to read and classify payment descriptions at a high
[23:39] volumes, mapping free text against uh curated goods uh curated goods classification taxonomy and sanction uh filed product codes. Every transaction
[23:52] uh is analyzed in context uh not just by keywords uh not just by keywords matched. Uh this helps detect sophisticated obuscations
[24:04] uh that simple uh rule-based angels uh will definitely miss. The results are uh surfaced again immediately in the animal officers workflow flagging high-risk uh
[24:17] transactions for the prioritize review. And because it runs on a datab bricks platform, lakehouse platform, it can uh operate at the full transaction uh volume with the governed and auditable
[24:33] outputs. So what does this scanner detects? It detects dual used goods, restricted goods and prohibited goods. And this is important. And this is
[24:45] important because old models assumes that risk is obvious in the text fields while in reality risk is often hidden behind vague wording, coded language or
[24:57] incomplete descriptions. And basically context is what turns text into signals. And bad actors hide behind ambiguity. AI
[25:09] helps us read through it. And once you have those signals, the next question is how analysts explore them uh at speed. And here comes the fourth use case. And
[25:24] the first use case is uh Genie for instance AML analytics. Genie can instantly surface entities with patterns of high-risk transaction descriptions, enabling behavior
[25:37] profiling without SQL or data engineering support. It can cross reference transactions um purpose scanner output and with geography to
[25:49] identify jurisdiction specific trade control uh trade control exposure in real time and it can perform uh trend analysis on counterparty transaction
[26:02] behav behavior changes uh or shifts that may indicate emerging risk or typology changes. This deck also shows um the kinds of questions uh that investigators
[26:16] can ask directly like show counterparties um with related suspicious payment purposes, find transactions mentioning restricted goods, compare behavior over time and
[26:29] many many other uh the philosophy behind this uh very simple. ML analysts should not spend their time filing to file a ticket to data
[26:41] engineering team every time they have analytical uh questions to answer. Genie enables natural language quering directly against governed AML data
[26:53] turning questions into immediate auditable uh insights. And this is not just a productivity improvement. This is uh a speed of thinking improvement. The
[27:06] best AML uh for me basically the best AML um case or insight is one that you ask for in the moment you need it. And now let me uh show how this capabilities
[27:21] work together in the real world.
[27:34] The slide is uh really operating model in one picture. So transformation is not four separate tools bolted together. It is one investigation cockpit, a unified experience served through multibot that
[27:47] gives every AML officer the full power of the platform through a single intuitive interface. The workflow moves from alerts received to counterparty
[27:59] search to document review to transaction context to decision and audit. Um along the way, the scanner uh applies AI language models to classify payment
[28:11] description. Cross source lookup happens instantly uh across sanctions, watch lists, dojun, and web intelligence. The AI document engine compares uh
[28:25] invoices, contracts, lending bills, etc. Uh customs and uh customs automatically and the purpose scanner plus genie provides uh behavioral and analytical context on
[28:39] demand. The officer then makes a governed documented decision with a full evidence trail. And why does this matter? Before the architecture, the same investigations required switching
[28:52] between five or more screens, more tools, waiting for a data team to answer and manually assembling context. The datab bricks based solution uh collapses
[29:05] that into one govern workflow that is faster, more consistent and fully auditable. For me, this is the heart of the transformation. It is not AI as a
[29:18] demo rather it is AI integrated into the processes that uh changes how works get done. Uh and uh one workspace again means less
[29:33] wheelchair compliance uh and more informed decisions and uh when you design the workflow end to end the impact or business impact shows up very quickly and the business impact is
[29:46] clear. Searches are completed uh in a fraction of time through unified search engine. AI document comparison replaces hours of manual works of the AML
[30:00] analysts. Officers arrives at decisions with richer, more complete context which reduces the risk of missed signals. Frontline AML officers also gain
[30:13] self-service uh capabilities through Genie and um data rigs apps reducing the backlog of data engineering team for routine investigation. um queries and
[30:26] the slide summarizes the impact as reduced manual lookup time and the faster review, low technical dependency, better investigation, context and flexibility. And this last part really
[30:39] matters because in AML um AML uh risks uh are always changing and uh companies should change together with
[30:51] this. So flexibility is not something that is nice to have. It becomes the integral part of the framework. So
[31:03] what did we learn from uh building all these use cases? Uh we close with uh six lessons learned that shaped the build uh and that apply to any financial
[31:16] institution considering a similar transformation. The first start from investigator workflow. If you start from technology alone, you will you're risking
[31:29] optimizing components instead of outcomes. Second, unify uh data before automating decisions. Automation on top of fragmented uh data is only scales
[31:44] inconsistency. Third, apps turn analytics into operations. Insights has value only when they are delivered to the front line in a usable
[31:57] form. Force Genie is powerful with the governed data only. Natural language is useful uh only when uh the answers are
[32:09] trusted. And fifs think in terms of intelligence. Um, not only rules. Yes, rules still matter, but uh they are not
[32:21] enough on their own. And six platform thinking bits point solution. The future belongs to the institutions that can connect data, AI, governance and
[32:33] workflow into one system. In AML, uh, point tools solve tasks, but platforms solve complexity. And let me close on that thought.
[32:45] Uh but before I would like to introduce our team from TBC and also our data bricks friends which made this happen and which are standing behind this
[32:58] success and um if uh there is um uh one message I hope stays with you is that uh is following ML modernization is not
[33:11] about replacing people with AI. I uh rather it's u uh giving people better intelligence systems to make faster, stronger and more accountable decisions.
[33:25] And so thank you very much or as we would say in Georgia the Madlova and Madlova is thank you. Thank you.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.