Banks need predictive intelligence to see what happens next.

Banks are moving rapidly from experimenting with generative AI to deploying AI agents. The appeal is obvious. Large language models can understand complex information, summarize documents, converse naturally and reason across a growing range of tasks. Agents add another powerful capability: they can act.

But there is an important piece missing from much of the current conversation about agentic AI.

An agent can reason about what to do without knowing what is likely to happen next.

Consider a customer with a $500,000 CD maturing next week. An AI agent can identify the maturity, retrieve the customer’s relationship history, explain available products, create a personalized message and initiate an outreach campaign.

But should the bank contact this customer at all? How likely is the customer to move the money? Would an additional 15 or 20 basis points actually change that outcome? What is the probability that the customer will accept another product? And is retaining this deposit worth the additional cost?

“These are not primarily language or reasoning problems. They are prediction problems.”

That distinction will become increasingly important as banks move from using AI to improve productivity to using AI to make and execute decisions.

LLMs help banks understand. Predictive intelligence helps them anticipate. Agents help them act.

Banks will need all three.

From Understanding to Anticipating

For decades, financial institutions have used predictive models to answer some of their most important business questions: Who is likely to default? Which transaction is fraudulent? Which customer is likely to leave? Who is most likely to respond to an offer? What will happen to liquidity if rates change?

Generative AI does not make these questions less important. Agentic AI makes them more important.

Deposit Retention: Stop Paying to Keep Customers Who Were Never Leaving

Take deposit retention.

A bank might have thousands of CDs approaching maturity. An agent can identify those customers and automate personalized outreach. That is useful, but it does not necessarily make the underlying decision more intelligent.

Suppose two customers each have $500,000 CDs maturing next week.

The first has a very high probability of renewing without any incentive. The second has an 80% probability of moving the funds unless the bank intervenes.

Offering each customer an additional 20 basis points may retain both deposits. But the bank has unnecessarily increased the cost of retaining the first customer’s money.

Across a large deposit portfolio, those seemingly small decisions add up quickly.

“The more valuable question is not ‘Which deposits are maturing?’ It is ‘Which deposits are at risk, what action is most likely to retain them, and what is the economically optimal price to do so?’”

The agent can execute the action. The LLM can personalize the interaction. Predictive intelligence determines where intervention is likely to create value.

Collections: Some Accounts Need a Call. Others Need to Be Left Alone

The same principle applies to collections.

Imagine a bank has 100,000 delinquent accounts. An AI agent can contact borrowers, explain balances, negotiate within established parameters, arrange payment plans and document the interaction. That can dramatically reduce the cost of collections.

But using an intelligent agent to contact every delinquent customer is still, in some respects, automating yesterday’s process.

What if the bank knew that one customer had an 85% probability of paying within seven days without intervention, while another had only a 20% probability of paying but became significantly more likely to do so when offered a structured payment plan?

Those customers should not receive the same treatment.

The first may require no intervention. The second may warrant immediate engagement.

“Sometimes the most intelligent action an AI agent can take is no action. But knowing when not to act requires an understanding of the likely outcome.”

Credit: Faster Paperwork Doesn’t Answer the Only Question That Matters

Credit provides another example.

LLMs can read financial statements, tax returns and loan documents. They can summarize borrower information, identify missing documentation and help a lender prepare a credit memo. Agents can gather information, coordinate workflows and communicate with borrowers.

But the fundamental economic question has not changed: What is the probability that this borrower will repay?

An LLM may explain why a company’s revenues declined. An agent may collect additional documentation. Neither replaces the predictive models that estimate default probability, expected loss or future cash flow.

In fact, the more autonomous the lending process becomes, the more important these predictions become. The faster AI can act, the more important it is to understand the probable consequences before it does.

Reasoning Is Not the Same as Foresight

This exposes a weakness in how we sometimes talk about AI.

We increasingly use “reasoning” and “intelligence” almost interchangeably. But reasoning about a situation and predicting its outcome are fundamentally different capabilities.

An LLM might reason that a customer with a large maturing deposit deserves attention. A predictive model might determine that the customer has a 72% probability of leaving and that a particular intervention reduces that probability to 31%.

The agent can then determine how to execute the intervention within the bank’s policies and controls.

The real opportunity is therefore not LLMs or agents in isolation. It is the combination of capabilities:

• Understand
• Predict
• Decide
• Act
• Learn

The LLM understands the situation. Predictive intelligence estimates potential outcomes. Decision intelligence weighs those outcomes against business objectives and constraints. Agents execute the appropriate action. And the resulting behavior provides new data from which the system can learn.

This is much closer to how human decision-making works, and much closer to what banks will need if they want to move toward autonomous operations.

The Danger of Creating Another Generation of Silos

There is another implication for bank technology leaders.

Most established banks already have considerable predictive intelligence. Credit models, fraud models, propensity models, attrition models, pricing engines and forecasting models have been built over decades.

But much of this intelligence remains trapped within products, applications and organizational functions.

Deposit teams build deposit models. Marketing builds propensity models. Collections builds payment models. Risk maintains credit models. And now individual teams are beginning to build agents.

“Unless banks are careful, agentic AI could create yet another generation of technology silos.”

A better approach is to make intelligence reusable.

A customer attrition signal developed for deposit retention might also inform wealth management, relationship banking and marketing. Cash-flow predictions used for underwriting might support treasury management and early-warning systems. Customer propensity models might guide both campaigns and conversations initiated by agents.

This is why I believe the idea of an enterprise intelligence layer becomes more relevant, not less, in an agentic world.

Rather than embedding intelligence independently inside every agent or application, banks can create a governed layer through which predictive models, signals, enterprise context and decision logic can be discovered and reused.

Agents then become consumers of enterprise intelligence rather than independent islands.

The Real Goal: A Bank That Senses, Predicts and Acts

The longer-term opportunity is not simply to deploy more agents. It is to build a bank that can continuously sense what is happening, anticipate what is likely to happen and determine the appropriate action.

Imagine a commercial customer’s operating balances begin declining.

The bank detects the change. Predictive intelligence identifies a growing probability of deposit attrition. Other signals show increasing outgoing payments and falling average balances. The system determines that proactive intervention is warranted.

An agent provides the relationship manager with the relevant context, recommends the next-best conversation and, where appropriate, initiates the outreach.

No single technology produced that outcome.

The LLM helped understand the context. Predictive models anticipated the likely outcome. Decision intelligence determined the appropriate response. The agent executed the action.

Each did what it does best.

This is the distinction banks should keep in mind as they develop their enterprise AI strategies.

LLMs are extraordinarily powerful at understanding information. Agents will transform how work gets executed. But banking remains, at its core, a business of making decisions about the future: who will repay, who will leave, who will buy, where risk will emerge and which intervention will produce the best economic outcome.

The next competitive advantage in banking AI will therefore not come from having the most agents. It will come from giving those agents better intelligence.

“Autonomy without foresight is simply automation moving faster.”