Why Your AI Models Need to Talk to Each Other (And Maybe Take Yoga Together)

Jerry Duan, SVP, Director, Credit Risk Models, United Community Bank

Jerry Duan, SVP, Director, Credit Risk Models, United Community Bank

Banks have become impressive at collecting data, such as account histories, loan performance, credit scores, and transactional behavior. But knowing a lot isn’t the same as knowing everything that matters. More often than not, the most valuable insights are the ones that slip between the cracks: the relationships, the context, the “who’s connected to whom and how.”

Take this scenario:

Customer A owns a company that supplies raw materials to Business B, a manufacturing client with a sizable credit line. Customer A also happens to be the guarantor on a loan for Customer C, who just applied for a working capital facility. That same customer has a personal account that’s shown some suspicious activity recently. Each of these data points exists in your systems—but in different places, and in isolation, they don’t look like much. It’s only when you connect them that the picture comes into focus.

Problem #1: Fragmented Customer Knowledge

In most banks, customer information is spread out like mismatched socks in different drawers. One department sees personal banking activity, another sees business account transactions, and a third watches the commercial loan portfolio. No one sees the whole wardrobe.

This disjointed view leaves out critical context. A business owner may appear low risk from a personal banking perspective, but if their business is losing its biggest buyer—who also happens to be another bank customer— that risk isn’t captured. Or maybe the business they own is placing massive orders with another client, driving up both companies’ exposure. Without a unified view, banks are flying blind to these real-world interconnections.

Problem #2: AI Models in Silos

Now layer on the AI. Banks have built separate models for different problems: credit risk, fraud detection, customer churn, and loan prepayment. Each one is trained on a specific dataset for a specific goal, and it operates like a brilliant hermit—smart but isolated.

Here’s the issue: the world isn’t divided like that. A customer who looks likely to prepay a loan might also be a churn risk. Fraud patterns often mirror high credit risk. And if two businesses are tightly connected, distress in one can quickly bleed into the other.

“By layering graph intelligence on top of your existing models, you get smarter risk scoring, more accurate fraud detection, and better customer engagement”

When models don’t share data or context, they miss these cross-signals. The fraud model flags nothing because it doesn’t see the credit exposure. The churn model fails to notice that a key supplier went bankrupt last week. It’s like having five detectives working the same case without talking to each other.

The Solution: Think in Graphs, Not Tables

Enter the graph. A graph-based approach treats data not as isolated rows but as a network of relationships. In a graph, every customer, account, business, and transaction is a node—and the connections between them (owns, supplies, guarantees, co-signs, etc.) are edges.

This means you can map out complex structures:

• Customer A owns Business X

• Business X is a primary supplier to Business Y

• Business Y is a major borrower with the bank

• Customer A is also a guarantor on another loan

Suddenly, you’re not looking at isolated records—you’re seeing a web of dependencies. A late payment from Business X might predict liquidity problems for Business Y. A default in one area can be understood as a network signal, not just a single event.

What Graph AI Can Do for You

By layering graph intelligence on top of your existing models, you get smarter risk scoring, more accurate fraud detection, and better customer engagement. Here’s how:

• Integrated Risk Awareness: Instead of a credit model only seeing one entity’s finances, it sees that entity’s entire ecosystem—partners, buyers, subsidiaries. A credit downgrade in a related business now becomes a warning light, not a post-mortem.

• Smarter Fraud Detection: Graphs help detect patterns that rule-based systems miss. If five seemingly unrelated businesses suddenly start sending funds to a new company— and they’re all indirectly linked through a common owner— that’s a red flag only a graph can wave.

• Customer-Centric View: Graphs allow banks to treat customers as people and businesses in context, not as isolated accounts. That means better personalization, more relevant offers, and fewer awkward “we didn’t know you already had that product” moments.

• Model Collaboration—and Consolidation: Here’s the real game-changer. Instead of maintaining multiple isolated models, banks can train one integrated, multi-purpose model using graph-enhanced data. By feeding in all datasets— credit, fraud, churn, and prepayment—and connecting them through relationships, this unified model can outperform each specialized model on its own. It doesn’t just predict outcomes more accurately—it understands why they’re happening and where they might spread.

A Real Competitive Edge

In a world where banking relationships are more complex than ever, the winners won’t be the banks with the most data—they’ll be the banks that understand how that data is connected. Graph-based AI gives you that edge. It turns your data from a set of spreadsheets into a living map of your customers’ world.

So yes, maybe your AI models don’t actually need to take yoga together—but they do need to collaborate, share intelligence, and learn from each other’s strengths. With graph-powered integration, they finally can.

Fraud

AI