Seeing the Unseen: Why Good AI Begins With Connected Data

Data: Noise or Intelligence?

So, how does an institution avoid these failures? How does it design and successfully implement a program that appropriately leverages current technological capabilities to focus resources on addressing the highest-risk threats, particularly in an environment that is currently more heavily focused on achieving efficiencies?

The answer is to use the current environment to create a reusable foundation that can be leveraged to solve both effectiveness and efficiency challenges.

The foundation begins with connecting fragmented data that resides in multiple applications and ensuring that those connections use a common language to interpret the data. In today's world, that interpretation must move beyond converting data into information; it must convert information into intelligence. For that intelligence to be reliable, the disparate pieces of data residing in multiple internal and external systems must be pulled together to create an accurate, holistic picture.

When creating that picture, unstructured data has historically posed particular problems, as has information stored in paper form. When created, this data was not designed to be analyzed by a machine; it was intended solely for human consumption. Given free reign to describe something, one never knows how it will be answered.

If Mary Smith is a conductor who decides to describe her occupation as “drug kingpin” on her CTR Form, it is up to the customer service representative to question her regarding her answer. Failure to do so would result in an investigation that would need to pull in all of the relevant information regarding the customer to determine whether “drug kingpin” was an honest answer, or an unfunny pharmacist joke.

In other words, the data must be used to create the context for the activity.

I will provide another quick, easy example that I think will resonate with most in the anti-financial crimes industry.

At some point, someone identified that sending money in round amounts could be indicative of illicit activity. For this example, let's take Mary Smith, who uses a payment network like Zelle or PayPal to transfer $20,000 to David Jones in July. Ding, ding, ding!

We have a round amount alert, which would result in an analyst determining whether the activity could potentially be suspicious. If the analyst cannot explain the activity, it would go to an investigator, who would then investigate the activity, possibly conducting an outreach to Mary to ask her who David Jones is and why she is sending him money.

All of this is resource intensive. It also is not a great customer experience. And it could be avoided if the data residing in the trust system was easily accessible to provide context for the activity in an automated way.

Contextualized data would have shown that Mary Smith had remarried, that her previous name was Jones, and that David Jones was actually her son who was enrolled in college. Suddenly, that round amount no longer looks suspicious because the use of the funds becomes quite clear: David is paying his tuition and buying books, deodorant, shampoo and beer (not necessarily in that order).

The transaction itself never changed. What changed was the context surrounding it. That context transformed what initially appeared to be suspicious activity into an explainable event without requiring additional investigation.

Figure 2: The same $20,000 payment from Mary Smith to David Jones, read two ways. Without context it generates review work, noise and a poor customer experience. With context it resolves into an explainable event.

Business Meaning: Will AI Agents Get It?

Now, let's assume that the bank in the example above decided to deploy an AI agent to work all of those annoying (almost completely useless) round amount alerts. Can it do so reliably?

When the AI Agent pulls in the information, will it pull in the information around the correct Mary Smith? Will it pull in all of Mary Smith’s information? If Mary Smith entered the retail bank through a branch 20 years ago, as Mary Jones, at a different address, and then entered the private bank at her new address, with her new name, after remarrying, will the AI Agent pull all of her accounts in? Will it know that she is also a signor on a commercial account in the name of her employer, and will that business’s accounts get pulled in? If Mary Smith moved $20,000 out of her employer’s account and into her personal, retail account a week before she moved it to her son, that might be material.

Deploying AI agents requires proper infrastructure. The systems and applications the AI agent gathers information from must have a consistent way of defining things. A "customer" in one system might have a different meaning as "customer" in another system. AI agents can navigate differences between systems, but they should not be left to guess which differences matter, otherwise they may encounter issues such as hallucinations.

Without a meaningful data infrastructure and appropriate controls to support its AI agents, an institution will simply miss risks and make wrong decisions faster.

This is an important distinction. AI does not eliminate the need for organizing data around its context; it increases it. The more responsibility institutions ask AI to assume, the more important it becomes that the underlying data presents a complete and consistent picture of the people, entities and transactions being evaluated.

Figure 3: The same sequence of steps run on two different foundations. Without connected data, an identity error compounds through the process. With it, each step narrows toward an explainable, controlled decision.

Program Effectiveness: Two Birds With One Stone?

Once this infrastructure is in place, it should be reused to address the primary compliance concern: identifying patterns across the institution that would otherwise remain hidden. What if the bank never lacked the information, but it was missing the connections that nobody knew how to follow?

Institutions with more complex business models typically have multiple customer onboarding systems, as well as many systems that process transactions. They may also have customers and business partners who act as intermediaries between the ultimate payor and payee. Their more sophisticated customers may have a tremendous number of related business entities, each with multitudes of counterparties conducting what can seem like an infinite number of transactions. Many of those counterparties may themselves be entities, with beneficial owners and other related parties.

All of this creates challenges in confirming actor identity across data sources and following the transaction from beginning to end.

This is the complexity I mentioned earlier, and it is often associated with the highest-risk threats. With this, let's return to our Mary Smith example.

If, instead of being Mary's son, it turns out that David Jones is actually a member of the City Council of Jonestown, then the transaction between Mary and David starts to look different. If Jonestown is also in the process of accepting bids for a new multi-million-dollar media campaign, the picture changes further.

If it also turns out that Mary Smith is the Vice President of Business Development for Bad Guy Media Company, and that Bad Guy Media Company moved $25,000 to a related entity, Bad Guy Shell Company, at another bank on Tuesday of last week, it looks stranger still.

And if Bad Guy Shell Company moved $25,000 into Mary's personal account on Wednesday of last week, well, now it looks suspicious.

If Bad Guy Media Company also happens to be a commercial client of the bank seeking to increase its working capital loan to fund business expansion, including work associated with governmental entities, the bank suddenly has a much bigger problem on its hands.

The bank has constructive knowledge of all the data and information contained within its systems. The fact that those systems are not appropriately talking with one another, using a consistent language to connect dots between customers, is unlikely to matter in hindsight if it is later discovered that Bad Guy Media Company obtained its business by paying off public officials.

Figure 4: Above, the payment as it appears in isolation. Below, the same payment inside its network of employment, ownership, government and lending relationships, where a potential bribery pattern becomes visible.

Momentum: Another Glimpse Into The Crystal Ball?

While — as mentioned at the start — the current drivers for change are understandably focused on efficiency, it is important not to lose sight of the effectiveness requirements. The infrastructure required to support AI agents directly connects to the infrastructure needed to improve Anti-Money Laundering (AML) effectiveness.

Responsible institutions have an opportunity to use the current environment to create a reusable framework that supports both objectives. The same connected, contextual view of customers, counterparties and transactions that enables AI to operate responsibly also strengthens an institution's ability to identify complex patterns of activity that would otherwise remain hidden in oceans of data.

Technology will continue to evolve. AI models will become more capable. Regulatory expectations will almost certainly evolve alongside them.

What is much less likely to change is the responsibility placed on financial institutions to identify and report financial crime. Investigators will continue trying to distinguish legitimate activity from suspicious activity. Institutions will continue trying to connect fragmented pieces of information into an accurate understanding of what is actually taking place.

AML is a natural area in which to deploy artificial intelligence because the work involves large volumes of fragmented data, repeated investigative tasks, weak signals spread across many transactions, and patterns that are difficult for humans to find at scale. AI agents must operate on a shared understanding of data; they cannot be left to determine on their own what things mean or which differences matter.

In short, AI alone cannot replace the need for accurate, connected and contextual data.

The institutions that benefit most from AI will not necessarily be those that deploy it first. They will be those that use this moment to build a stronger foundation for both AI and more effective financial crime programs.