Supercharging Traditional Forensic Techniques with AI

The era of technology has changed for good. Service delivery has changed significantly, from Web 1.0, where we had static pages to Web 2.0, wherein interaction among users was hallmark. Now we are moving to the era of AI. The new buzzword in the AI domain is Agentic AI, which can make automatic decisions. Auditing as an industry has been sluggish due to change in technology. Barring global consulting firms, Indian firms are wary of investing in technology, but that has to change. Gone are the days of manual verification of inventory; IoT (Internet of Things) devices are changing the rules of the game. The evolution continues to this day, but it is no longer following a linear trajectory. It is dynamic, responsive, and disruptive. As technology advances from basic automation toward intelligent, autonomous agentic systems, the terrain of financial transactions is also going through a dramatic transformation.

When fraudsters swiftly adapt to these innovations, they exploit not just the vulnerabilities in digital payment gateways but also manipulate advanced platforms such as blockchain, cryptocurrency transactions, and smart contracts.

Global corporations have started integrating blockchain into their core business processes. As a consequence, it also introduces a unique complexity to audit trails and forensic investigations. Take, for instance, smart contracts that automate payments and contractual obligations in real-time. This makes auditors face entirely new challenges in establishing accountability and verifying the authenticity of transactions. We may be comfortable with our spreadsheets for now, assuming such scenarios remain limited in India presently, but they\'re rapidly gaining traction and will soon dominate the financial outlook.

The Two-Fold Problem

As criminals upgrade and evolve with newer strategies, organizations deploy increasingly sophisticated anti-fraud systems. Chartered Accountants and forensic practitioners, as a result, face a twofold adaptation challenge here.

1. First is to understand the deeply advanced digital infrastructures that their clients now operate on, and
2. The second and most important aspect is to embrace technology within their own forensic methods to match them.

A large-scale aircraft manufacturer, whose corruption investigation required analyzing nearly 60 million documents using AI-assisted methods. Similar investigative volumes and complexities will soon become the norm even within India\'s corporate fraud investigations, forcing the necessity to shift toward tech-enhanced forensic capabilities. Rather than viewing this seismic shift as a threat or a \'challenge\', forensic professionals can harness technology as their superpower, an enabler. Leveraging existing human expertise and augmenting it digitally using AI and machine learning tools can transform practitioners into agile, digitally adept detectives. It is precisely this integration of \'human judgment and technological precision\' that this article aims to explore deeply, offering a practical roadmap to bridge the digital demands of the financial fraud community.

Let\'s acknowledge the growing challenge of tech-driven fraud and the fact that organizations are rapidly adapting by integrating AI-driven solutions across risk management, internal controls, and advanced fraud detection systems. As organizations evolve, it becomes essential for forensic practitioners to do the same. But where does the challenge arise for them? Why are traditional investigation methods no longer sufficient?

This is where data plays a crucial role. Today, even simple forensic reviews generate vast amounts of data in multiple formats. The disputed sum may be as small as one or two crore rupees, yet an overwhelming volume of data is produced, including emails, encrypted communications, payment ledgers, bank statements, invoices, and other structured and unstructured documents. Manually sifting through thousands or even millions of such records-while adhering to the strict deadlines and accuracy demands of forensic engagements is not just inefficient but virtually impossible.

Traditional approaches were designed for a time when accounting was paper-based, transactional volumes were limited, and documents were systematically archived. While these methods were once a mark of expertise, they are now struggling to keep pace, creating vulnerabilities that tech-savvy fraudsters can exploit.

Ultimately, such old, traditional methods just need to evolve, simply because we must now triangulate data from various divergent sources in order to examine, analyze, and extract valuable insights, something called data blending. Yet, how exactly can forensic professionals practically implement data blending to identify and investigate fraud?

It means cruising through two separate but interconnected categories of data: structured and unstructured. Structured data is clearly organized numerical or transactional records. It can be your general ledger entries, financial statements, vendor invoices, payment records, and so on. In this realm, forensic analysis leverages statistical anomaly detection tools, such as Benford\'s Law, to identify irregularities.

Unstructured data, on the other hand, relates to qualitative information, like textual data. These are the unorganized data we discussed, like emails, chat logs, contracts, and even voice recordings form part of unstructured data. These do not neatly fit into tables or spreadsheets. Manually sifting through this information is labour-intensive and error-prone. With AI technologies, the traditional investigative approaches can now be augmented for both types of data, scaling the depth of analysis.

Augmenting Structured Data

When we talk about structured data, one highly effective technique to illustrate is Benford\'s Law, a statistical principle widely discussed in financial fraud contexts. Benford\'s Law has been a trusted method to detect anomalies in transactional data in accounts payable or general ledger entries by checking if the occurrence of first-digit frequencies conforms to preset patterns. That is, numbers beginning with digits like 1 and 2 occur significantly more often than those starting with 8 or 9. Forensic professionals traditionally use this concept (\"first two-digit test\"), computing their frequency distribution and comparing these to the expected Benford probabilities. Financial records, including ledger balances, expense reports, invoice amounts, and sales numbers, all follow this structure. So, if a dataset deviates from Benford\'s predicted distribution, it can be a sign of data tampering or fraudulent manipulation. It works well when dealing with medium-sized datasets of 2,500-5,000 records. But in contemporary financial settings, where businesses execute millions of transactions every day across several departments and business divisions, it suffers. Here, there is a scalability issue. Since fraud doesn\'t occur uniformly and it often happens within specific vendors, regions, or payment methods, this traditional Benford analysis of the entire company\'s data might overlook irregularities in one department or business unit.

To overcome this, a concept of Benford Subset Divergence Analysis (BSDA), a method that uses AI to analyze financial data dynamically, can quickly examine thousands of smaller subsets, such as:

• Individual vendors or groups of suppliers
• Specific geographic regions
• Different payment methods (NEFT, RTGS, cheque, etc.)
• Particular general ledger (GL) accounts or document categories

Rather than running Benford\'s Law once across an entire dataset, the BSDA approach can analyze different subsets across a multi-billion-dollar transactional dataset in minutes, an efficiency inconceivable through manual analysis alone.

Traditional method

Consider the case of a multinational corporation that processes millions of journal entries annually. Forensic professionals suspect potential fraud, such as fictitious entries or rounding anomalies, but manually analyzing the data is impractical due to its volume and time constraints. In the traditional scenario, the practitioner prepares the data by extracting, say, one million journal entries from the general ledger and applies Benford\'s Law to analyze the first two digits of each transaction amount. Then, using spreadsheets, they calculate the frequency of each digit combination (10 to 99) and compare it to the expected distribution.

They identify deviations where actual frequencies exceed or fall below the expected range, such as

• Spikes at 20 and 21: Transactions starting with these digits occur more frequently than expected.
• Valleys at 90-99: Transactions in this range are underrepresented, suggesting potential rounding or threshold manipulation.

Here, the limitations that they may face would be

• Deviations spread across the dataset, making it hard to pinpoint specific fraud risks.
• Manually filtering data by vendor, department, or geography to identify anomalies is time-consuming.
• Small but significant frauds within subsets (e.g., a single department or vendor) may be \"washed out\" in the aggregate data.

AI Method

Now, if the same dataset is loaded into an AI-powered fraud detection tool and configured to perform the BSDA, it may find the specific vendor showing the deviations in transactions starting with 20 and 21. Let\'s call him Vendor X. The tool can pinpoint hidden/specific deviations as follows.

• For the spike at 20: 12% of Vendor X\'s transactions start with 20, compared to the expected 4.5%.
• For the spike at 21: 10% of transactions start with 21, compared to the expected 4.1%.

The AI tool can also identify \"more-like-this\" transactions (popularly called predictive analytics) across other vendors and GL accounts, which may reveal any fraud ring involving multiple entities. Additionally, transactions under specific GL may also reveal any unusual patterns, suggesting potential data manipulation.

Based on this, the investigation can be targeted by better focusing on Vendor X and related subsets, uncovering -

• Fictitious entries by creating fake invoices below the threshold of Rs. 20000.
• Rounding anomalies of Rs. 20000 and Rs. 21000
• Duplicate transactions as double payments.

Here, this AI-augmented approach of Benford\'s Law could handle complex datasets across multiple dimensions and also predict similar patterns across the dataset.

Some use cases of BSDA are as follows.

• Fake Beneficiaries: Fraudulent recipients exploiting government subsidy programs
• Invoice Splitting: Divided invoices inflating infrastructure project costs
• False GST Refunds: Shell companies claiming illegitimate GST refunds.
• Election Funding: Manipulated funds compromising electral transparency
• MSME Anomalies: Irregularities in MSME loans indicating potential fraud.
• Stock Schemes: Manipulative stock practices harming market fairness.

Labyrinth of Unstructured Data

While AI-driven BSDA is a powerful tool to address numerical anomalies within structured datasets, forensic investigations today extend beyond numbers alone. Investigators often face massive volumes of unstructured information like emails, chat logs, and digital invoices, which makes manual review both impractical and time-consuming. It is only natural that an AI model capable of using both structured and unstructured information is the answer. Multimodal AI, as they are referred to, integrates and analyses data from various sources and formats, such as text, images, and audio. Now you may think of it as an LLM like GPT. But GPT is founded on Generative AI wherein it is used to create new data or content and can process all these file formats. Multimodal AI, on the other hand, also can integrate these multiple types of data. Think of it like a cook and a star chef. Gen AI is a cook who can give you a recipe based on your available ingredients and preferences. But Multimodal AI is a star chef who can understand a picture of your fridge contents and suggest a recipe while listening to your instructions and adjusting cooking times based on heat sensors or even the season!

In the financial fraud world, multimodal AI is capable of giving a holistic view of a transaction, simultaneously reviewing structured invoices/payments, unstructured vendor emails, WhatsApp conversations, recorded phone calls, CCTV footage indicating vendor and employee meetings, and so on. What used to be achieved through keyword searches using forensic toolkits, rudimentary speech-to-text tools, or cross-referencing emails to voice chats is now achieved using multimodal. Global financial services companies are already employing such models, stepping up their fraud detection game.

The core components of the model are given below.

Components of Multimodal AI

Advanced Natural Language Processing (NLP) for text analysis

• It shows the capability of analysing emails and chats beyond keywords, detecting sentiment, intent, and coded language.
• For example, \"medicine delivery\" in an email could be flagged as a drug trafficking euphemism based on context.

Audio processing with speaker recognition

• AI transcribes voice chats in real-time and also enhances the audio quality by removing background noise and identifying speakers using vocal biometrics, cross-referencing against databases or prior samples.

Graph-based correlation

• AI constructs relationship networks, linking entities like people, devices, and events across emails, voice chats, and metadata.
• A call mentioning \"meet at 5\" could be tied to an email with a location, visualised instantly.

But how is this technology relevant to practitioners? In a hypothetical scenario, imagine investigating a Rs. 500 crore corruption scandal involving politicians, bureaucrats, and shell companies. We are presented with terabytes of emails, WhatsApp chats, and intercepted voice calls. But in this case, these are all fed into a multimodal system.

• The NLP flags all the coded bribe terms such as \"facilitation fee,\" \"expediting payment,\" \"consultancy charges,\" or \"special handling charges.\"
• The audio AI identifies a politician\'s voice in a call
• Graph analytics links the call to an email scheduling a payoff.
• Integrates any subtle linguistic cues and emotional tone that might indicate bribery negotiations or suspicious dealings.
• Establish critical timelines or suspicious interactions, secret meetings, previously unnoticed between persons.

Within hours, the model produces a clear timeline of events, maps interactions among parties involved and shell companies, and provides investigators with actionable \"evidence clusters.\"

Potential Use Cases: Blending Structured and Unstructured Data

Is the scenario far-fetched?

While this might sound like science fiction, law enforcement agencies across the world are already employing such models. Global regulatory agencies use multimodal AI to analyze complex financial crimes and terrorism cases. Multimodal AI is already being explored extensively beyond finance. Ecological research, for instance, employs models like the TaxaBind framework, integrating text-based taxonomy data, geographic coordinates, and satellite imagery, clearly showcasing multimodal AI\'s effectiveness in synthesizing diverse data streams into insights.

Recent developments in the Indian scenario illustrate tax authorities considering a review of various digital sources, including emails, social media activities, online investments, and financial transactions to detect tax evasion, which may be a potential use case of such integrated models. Such scenarios, if they materialize, considering the data protection laws, would require authorities to integrate vast volumes of structured and unstructured data effectively, though challenging with traditional models. In these cases, emerging approaches, such as multimodal AI, might become a necessary tool to enable investigators to analyze data seamlessly across various formats.

Importantly, technology transcends borders, and data knows no jurisdictions. It is becoming impossible to analyze these developments through a country-specific lens, given their interconnectedness. As organizations across the world adopt such cutting-edge digital models for risk and fraud management, forensic practitioners can no longer afford to lag. Practitioners need to actively understand and incorporate such novel methods to stay competitive and indispensable. Instead of being overwhelmed by technological advancement, we are in a unique position to welcome it and use our knowledge and expertise superpowered with AI to detect new-age financial crimes.

Ultimately, the future of forensic investigations lies in a harmonious marriage of human judgment and technological accuracy. It is the technology itself that is transforming practitioners from mere number-crunchers into highly capable digital detectives, preparing us to tackle financial frauds head-on in this data-centric world.