WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=Iwe_RY-fYgI

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/Iwe_RY-fYgI.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:04.040 --> 00:00:05.920
Hello everyone.

s2
00:00:05.920 --> 00:00:13.120
Thank you to the AI engineering world fair team for providing this wonderful opportunity to share my research.

s3
00:00:13.120 --> 00:00:18.680
It is truly an honor to be speaking alongside so many talented researchers and practitioners.

s4
00:00:18.680 --> 00:00:20.520
My name is Varsha Shah.

s5
00:00:20.520 --> 00:00:26.000
I am an enterprise technical architect working at Tata Consultancy Services working for Microsoft.

s6
00:00:26.000 --> 00:00:34.520
I'm focused on artificial intelligence enterprise compliance, finance governance, and intelligent automation.

s7
00:00:34.520 --> 00:00:43.440
Today, I would like to share my research on AI-driven multi-document correlation for enterprise financial compliance and fraud detection.

s8
00:00:43.440 --> 00:00:56.360
As organizations continue to digitalize their operations today, they generate a numerous amount of data for the financial system across payroll, tax, procurement, transaction system.

s9
00:00:56.360 --> 00:01:05.080
Ironically, while we have more data than ever before, compliance teams continue to struggle with hidden fraud patterns, regulatory risk.

s10
00:01:05.080 --> 00:01:10.640
The reason is the most existing solution analyze the documents independently.

s11
00:01:10.640 --> 00:01:18.040
While many of the most critical risk only become visible when the information is connected across the multiple systems.

s12
00:01:18.040 --> 00:01:38.080
In this presentation, I'll introduce you to a framework that combines the graph-based entity correlation, probabilistic risk modeling, and cross-jurisdictional normalization to uncover these hidden relationships and transform enterprise compliance from reactive process into proactive intelligence capability.

s13
00:01:38.080 --> 00:01:51.720
The framework was evaluated using approximately 3 million financial records across four jurisdictional demonstrating both the strong detection performance and meaningful operations improvement.

s14
00:01:51.720 --> 00:01:59.640
With that context, let's begin by looking at the compliance gap that organizations continue to face today.

s15
00:02:01.160 --> 00:02:05.840
Let's begin by understanding the compliance gap that many organizations continue to face.

s16
00:02:05.840 --> 00:02:10.560
Enterprise compliance has become significantly more complex over the last decade.

s17
00:02:10.560 --> 00:02:20.720
Organizations now operate across multiple countries, regulatory frameworks, and financial system, each with own reporting standards and compliance requirements.

s18
00:02:20.720 --> 00:02:25.680
At the same time, the volume of the enterprise has grown exponentially.

s19
00:02:25.680 --> 00:02:38.800
Payroll records, tax filings, procurement transactions, and financial documents are generated every day, making manual reviews both time-consuming and increasingly impractical.

s20
00:02:38.800 --> 00:02:42.800
Adding to this challenge, fraud has been evolved.

s21
00:02:42.800 --> 00:02:48.560
Modern fraud rarely appears as an obvious error within a single document.

s22
00:02:48.560 --> 00:02:54.360
Instead, it exploits subtle inconsistency across multiple system.

s23
00:02:54.360 --> 00:02:59.120
Patterns that often remain invisible when the records are reviewed independently.

s24
00:02:59.120 --> 00:03:02.680
This creates a fundamental limitation.

s25
00:03:02.680 --> 00:03:14.320
Traditional rule-based and document-level NLP system are designed to validate individual records, but they are not built to understand relationship across the documents.

s26
00:03:14.320 --> 00:03:17.519
And that's the gap this research aims to address.

s27
00:03:17.519 --> 00:03:24.720
Moving beyond this isolated document analysis to uncover the hidden risk through the cross-document correlation.

s28
00:03:24.720 --> 00:03:35.519
To better understand this limitation, let's look at why traditional document level analysis often fail to detect the most sophisticated fraud patterns.

s29
00:03:36.959 --> 00:03:44.680
To understand why traditional approaches struggle, let's consider how most compliance system operates nowadays.

s30
00:03:44.680 --> 00:03:47.760
They evaluate each document independently.

s31
00:03:47.760 --> 00:03:51.800
A payroll register is validated against the payroll rules.

s32
00:03:51.800 --> 00:03:55.360
The vendor invoices are checked against the procurement policies.

s33
00:03:55.360 --> 00:03:59.080
A tax filing is reviewed under the tax regulations.

s34
00:03:59.080 --> 00:04:05.160
If each document passes its individual validation, the transaction is generally considered compliant.

s35
00:04:05.160 --> 00:04:12.520
The challenge is that many sophisticated fraud patterns doesn't really appear within a single document.

s36
00:04:12.520 --> 00:04:16.239
They emerges only when the multiple documents are analyzed together.

s37
00:04:16.239 --> 00:04:19.760
For example, a payroll record may appear accurate to us.

s38
00:04:19.760 --> 00:04:23.360
A vendor invoice may seem legitimate to us.

s39
00:04:23.360 --> 00:04:27.280
A tax filing may be correctly submitted.

s40
00:04:27.280 --> 00:04:34.480
But, when these records are connected, they're revealing the inconsistencies and indicate the fraud and compliance risk.

s41
00:04:34.480 --> 00:04:37.000
The information already exists.

s42
00:04:37.000 --> 00:04:41.480
What missing is the ability to understand the relationship between these documents.

s43
00:04:41.480 --> 00:04:53.120
That is why the research shifts the focus from document level validation to cross document intelligence, enabling the organizations to detect the risk that would otherwise remain hidden.

s44
00:04:53.120 --> 00:05:01.120
So, if the problem is understanding relationships rather than individual documents, what kind of architecture can solve this?

s45
00:05:01.120 --> 00:05:04.960
Let me introduce you to the framework.

s46
00:05:05.240 --> 00:05:09.640
Now that we have established the problem, let's look at the proposed framework.

s47
00:05:09.640 --> 00:05:22.800
Rather than relying on a single model or algorithm, the solution is built on three complementary components that work together to transform the raw enterprise data into the actionable compliance intelligence.

s48
00:05:22.800 --> 00:05:26.200
The first component is the entity correlation engine.

s49
00:05:26.200 --> 00:05:39.680
It the purpose of this is to connect the related information across the payroll, tax, procurement, financial systems, creating a unified view of enterprise activity rather than isolated records.

s50
00:05:39.680 --> 00:05:53.440
Once these relationships are established, the second component that is adaptive probabilistic risk model which is which evaluates the connected data to determine which pattern represent meaningful compliance.

s51
00:05:53.440 --> 00:06:02.960
Instead of generating alerts based on single rule, it it prioritizes the cases using the multiple risk signals here.

s52
00:06:02.960 --> 00:06:16.360
The cross-jurisdictional normalization layer provides the regulatory context by standardizing the currency, tax structure, reporting standards, and compliance rules across the different jurisdictions.

s53
00:06:16.360 --> 00:06:24.400
This ensures that the risk are evaluated consistently regardless of where the transaction originated.

s54
00:06:24.640 --> 00:06:27.680
So, individually, each component provides a value.

s55
00:06:27.680 --> 00:06:35.960
Together, they enable the framework that move beyond the document validation and towards the enterprise-wide compliance intelligence.

s56
00:06:35.960 --> 00:06:42.120
In the next few slides, I'll briefly explain you how each of these components contribute to the process.

s57
00:06:42.120 --> 00:06:51.400
So, let's begin with the foundation of the framework, the entity correlation engine, which makes the cross-document intelligence possible.

s58
00:06:52.760 --> 00:06:56.640
So, the first component is graph-based entity correlation engine.

s59
00:06:56.640 --> 00:07:03.520
The role is to connect the entities across the payroll, tax, procurement, financial system into unified network.

s60
00:07:03.520 --> 00:07:14.720
Rather than analyzing the documents independently, it it identifies the relationship between the employee, vendors, account, transaction, regulatory files.

s61
00:07:14.720 --> 00:07:24.720
And these connections reveal the hidden pattern and structural anomalies that are often invisible to the traditional document level analysis we were doing.

s62
00:07:24.720 --> 00:07:30.360
In simple terms, this component answers one fundamental question, what is connected?

s63
00:07:30.360 --> 00:07:36.080
It provides the relation relational fundamental for the rest of the framework.

s64
00:07:36.080 --> 00:07:45.800
Once those relationships are established, the next step is determining which one actually represents the meaningful risk.

s65
00:07:48.760 --> 00:07:55.720
Once relationships have been established, the next step is to determine which one actually represents meaningful risk.

s66
00:07:55.720 --> 00:07:58.840
That's the role of the adaptive probabilistic risk model.

s67
00:07:58.840 --> 00:08:15.440
Instead of relying on static rules, it combines the multiple risk indicators such as anomaly anomaly strength or the source reliability historic patterns has happened to calculate confidence best risk score.

s68
00:08:15.440 --> 00:08:22.640
This helps prioritize the cases that require immediate attention while reducing the unnecessary investigations.

s69
00:08:22.640 --> 00:08:32.479
The important advantage is its ability to learn from the audit outcomes, allowing the models to continuously improve its accuracy over the time.

s70
00:08:32.479 --> 00:08:40.560
Simply put, this component answers the question that what is most likely to be genuine compliance risk.

s71
00:08:40.560 --> 00:08:52.000
Once we have identified the high-risk cases, the final step is to ensure those risk are interpreted consistently across the different regulatory environments.

s72
00:08:52.720 --> 00:08:56.680
The final component is cross-jurisdictional normalization layer.

s73
00:08:56.680 --> 00:09:04.400
Global organizations operate across different countries, currencies, tax structure, and reporting standards.

s74
00:09:04.400 --> 00:09:10.520
Without normalization, the same transaction can be interpreted differently depending on the jurisdiction.

s75
00:09:10.520 --> 00:09:22.920
This layer harmonizes the financial data by standardizing the currencies, tax rules, the reporting periods, and the classification schemes as well.

s76
00:09:22.920 --> 00:09:29.280
As a result, the risk are evaluated consistently regardless of where the where it has originated.

s77
00:09:29.280 --> 00:09:38.240
In short, this component answers to the question, "How should the risk be interpreted within the appropriate regulatory context?"

s78
00:09:38.240 --> 00:09:47.200
With all these components working together, the next step was the evaluated how the framework performs under real-world enterprise conditions.

s79
00:09:48.000 --> 00:09:56.920
So, before discussing the result, let's um uh it's it's very important to uh briefly understand the evaluation of the environment.

s80
00:09:56.920 --> 00:10:07.920
The framework was evaluated using the approximately 3 million of the financial records collected over 5 years of period over the four different regulatory jurisdictions.

s81
00:10:07.920 --> 00:10:24.360
The objective was to assess how the framework perform under the real realistic uh enterprise condition and where uh large volume of interconnected financial data and varying regulatory uh requirement introduce the significant complexity here.

s82
00:10:24.360 --> 00:10:28.560
Uh with that foundation established, let's look at how the framework perform.

s83
00:10:28.560 --> 00:10:34.320
So, how effective was the framework in detecting hidden compliance risk?

s84
00:10:34.320 --> 00:10:37.400
Let's look at the results.

s85
00:10:37.760 --> 00:10:41.920
So, here now uh look at the detection performances.

s86
00:10:41.920 --> 00:10:51.680
The framework achieved approximately 91% of precision, meaning the vast majority of the flat cases were confirmed as a genuine anomalies here.

s87
00:10:51.680 --> 00:11:07.240
At um it also achieved 87% um recalls, demonstrating its ability to identify most true fraud cases while minimal uh minimizing the misdetections here.

s88
00:11:07.240 --> 00:11:17.480
So, together, these results produce a F1 score that is 0.89, indicating a strong balance between the precision and recall here.

s89
00:11:17.480 --> 00:11:29.400
More importantly, these results are achieved across four jurisdictions and large enterprise scale data, demonstrating that the framework performs consistently in the complex real-world uh environments.

s90
00:11:29.400 --> 00:11:32.840
The next question is um equally important.

s91
00:11:32.840 --> 00:11:38.640
How do these improvements translate into operational value for the compliance teams?

s92
00:11:38.640 --> 00:11:41.680
Detection accuracy is one part of the story.

s93
00:11:41.680 --> 00:11:49.360
The real value um comes from improving the efficiency of the compliance operations.

s94
00:11:50.040 --> 00:11:54.960
Beyond detection accuracy, the framework delivers measurable operational benefits.

s95
00:11:54.960 --> 00:12:08.760
One of the most significant outcome was a 76% uh reduction in false positive, meaning um investigators spend uh far less than reviewing the cases that they uh they're ultimately

s96
00:12:08.760 --> 00:12:10.280
legitimate.

s97
00:12:10.280 --> 00:12:22.840
So, in addition, the framework reduces the manual audit efforts by approximately 40% allowing compliance teams to be focused on the only high-risk cases instead of uh routing the document reviews.

s98
00:12:22.840 --> 00:12:32.720
Ultimately, uh this translates into faster investigations, improved uh resource utilization, and uh greater confidence in compliance decisions.

s99
00:12:32.720 --> 00:12:42.760
In other words, the uh the framework doesn't just detect fraud more effectively, it helps compliance teams work more effectively.

s100
00:12:42.760 --> 00:12:53.800
So, these operational improvements become even more meaningful when we compare uh the uh the framework with the traditional rule-based approaches.

s101
00:12:58.040 --> 00:13:03.080
This comparison highlights the advantage of moving beyond the traditional rule-based system.

s102
00:13:03.080 --> 00:13:15.440
Uh conventional approaches are effective at validating the individual records, but they often struggle to detect the complex fraud patterns uh that spans multiple uh documents and systems.

s103
00:13:15.440 --> 00:13:29.760
So, by incorporating entity uh correlation, probabilistic risk modeling, and regulatory normalization, the uh proposed framework consistently um performs the baseline across the key performance metrics.

s104
00:13:29.760 --> 00:13:31.440
The key takeaway is simple.

s105
00:13:31.440 --> 00:13:44.000
Connecting data across documents produces better detection, fewer false positive, and more actionable, uh, compliance intelligence than analyzing the documents is, um, in in isolations.

s106
00:13:44.000 --> 00:13:52.320
One of the reason the framework continues to improve over the time is that, uh, doesn't rely on the static rules anymore.

s107
00:13:52.320 --> 00:14:00.320
Instead, it continuously learn from the completed audit and investigate, uh, and the investigator feedbacks here.

s108
00:14:00.600 --> 00:14:05.800
So, this framework is designed to continuously improve through the feedback.

s109
00:14:05.800 --> 00:14:10.200
Every, uh, completed audit provides the valuable information.

s110
00:14:10.200 --> 00:14:14.400
Uh, confirm the fraud case, strengthen the future detection pattern.

s111
00:14:14.400 --> 00:14:19.880
While false positive helps refining the risk scoring and reduce the unnecessary alerts.

s112
00:14:19.880 --> 00:14:27.480
This create a continuous learning cycle where the system becomes more accurate with each audit and investigation.

s113
00:14:27.480 --> 00:14:36.360
As fraud patterns evolve and business, uh, environments changes, the framework adopts rather than relying on manual rule updates.

s114
00:14:36.360 --> 00:14:41.960
In simple terms, every completed audit helps make the next audit more efficient.

s115
00:14:41.960 --> 00:14:48.440
This continue, uh, continue learning, uh, capability enables a broader shift here.

s116
00:14:48.440 --> 00:14:57.600
From reactive to compliance issue after they occur to identifying the risk before they become audit, uh, findings here.

s117
00:14:57.600 --> 00:15:00.560
It's more of a proactive.

s118
00:15:02.320 --> 00:15:08.839
Beyond improving detection accuracy, this framework changes how organization approach compliances.

s119
00:15:08.839 --> 00:15:11.839
Traditionally, the compliance has been reactive.

s120
00:15:11.839 --> 00:15:17.839
Issue were identified after audits, investigation, or regulatory reviews have happened.

s121
00:15:17.839 --> 00:15:25.480
This framework enables more proactive approach through continuous, uh, monitoring, cross document correlation, adaptive learning.

s122
00:15:25.480 --> 00:15:32.000
Instead of asking what went wrong, organizations can begin asking what is likely to go wrong next.

s123
00:15:32.000 --> 00:15:45.840
That shifts from reactive validation to predictive governance, allowing the compliance teams to identify the risk earlier, prioritize the investigation more effectively, and make better informed decisions.

s124
00:15:45.840 --> 00:15:53.560
Ultimately, the compliance becomes an ongoing intelligence functions rather than a periodic review process.

s125
00:15:53.920 --> 00:16:00.640
While the framework is uh research driven, it has been designed with the enterprise deployment in mind.

s126
00:16:00.640 --> 00:16:04.520
Successful implementation depends on uh four key consideration.

s127
00:16:04.520 --> 00:16:13.000
First one is seamless integration with existing enterprise systems such as ERP, payroll, procurement, tax uh tax platforms.

s128
00:16:13.000 --> 00:16:21.800
Second, the jurisdictional specific uh configuration to ensure the compliance with local regulations and reporting standards.

s129
00:16:21.800 --> 00:16:33.360
Third one is um alignment with the audit framework enabling the investigator to focus on prioritized risk code cases instead of manual document reviews.

s130
00:16:33.360 --> 00:16:35.080
And finally, scalability.

s131
00:16:35.080 --> 00:16:45.920
The evaluation demonstrate that the framework can effectively process millions of financial records um making it suitable for the large enterprise environments.

s132
00:16:45.920 --> 00:16:56.080
Uh so, together these uh consi- these considerations help bridge the gap between the research and the practical enterprise adoption.

s133
00:16:57.839 --> 00:17:02.160
As I conclude, I would like to leave you with four key takeaways.

s134
00:17:02.160 --> 00:17:10.680
First is many of today's most significant compliance and fraud risk exist between the documents, not within them.

s135
00:17:10.680 --> 00:17:18.760
Detecting these risk require more beyond document level analysis and cross document intelligence.

s136
00:17:19.000 --> 00:17:31.080
Second is combining the graph based entity correlation, probabilistic risk modeling, and cross jurisdictional normalization creates a more comprehensive and scalable approach to enterprise compliance.

s137
00:17:31.080 --> 00:17:49.280
Third is the result demonstrated that this particular framework is not only improves the detection accuracy, but also reduces the false positive and significantly lowers the manual audit efforts, delivering the measurable operational value.

s138
00:17:49.280 --> 00:18:03.520
Finally, by continuous learning from the audit outcomes, and the framework enables the organizations to move beyond reactive compliance toward predictive intelligence driven risk management here.

s139
00:18:03.520 --> 00:18:19.400
I believe this represents an important step toward the future of enterprise financial govern- governance, where AI is not only helping the organizations to detect risk, but also anticipates and

s140
00:18:19.400 --> 00:18:23.120
prevent this risk going forward.

s141
00:18:25.080 --> 00:18:34.760
So, thank you once again to the AI Engineering World Fest team for this wonderful opportunity, and thank you all for your time and attention.

s142
00:18:34.760 --> 00:18:43.720
Please feel free to reach out to me with your questions and your enthusiasm on this email below, or connect me on the LinkedIn.

s143
00:18:43.720 --> 00:18:45.400
I'm happy to help you.

s144
00:18:45.400 --> 00:18:57.240
If you're building with such kind of amazing systems for compliances using the AI wanted to transform them, and thank you all again.

s145
00:18:57.240 --> 00:19:00.040
See you next time.