The AI agent running 90% of Bank Hapoalim's foreign transfers

Bank Hapoalim, one of Israel's largest banks, once required a banker to manually gather customer, transaction, and compliance information for every incoming foreign transfer before making a decision. Project Sentinel changed that, handling more than 90% of transfers automatically and bringing in a banker when human judgment is needed.

Every transfer started with a banker

Incoming foreign currency transfers were one of Bank Hapoalim's most operationally complex processes. A bank that has led Israeli financial services for over a century, Bank Hapoalim wasn't content to solve this with more headcount. Every transfer required a banker to pull information from KYC records, CRM, foreign exchange systems, customer history, pending transactions, and risk and compliance signals, then determine whether to approve, reject, or escalate it within a single business day.

As volumes grew, Bank Hapoalim faced a simple reality: It couldn't keep adding people to keep pace.

What if an AI agent could make the first decision?

Bank Hapoalim built Project Sentinel, an AI agent designed to evaluate incoming transfers much like an experienced banker would.

Every 30 minutes, Sentinel brings together a 360-degree view of the customer, evaluates the transfer against operational and risk rules, and recommends whether to approve, reject, or escalate. It can also label the transaction for regulatory reporting and send execution instructions to RPA systems.

Today, Project Sentinel handles more than 90% of incoming foreign transfers automatically, while handling time has fallen from approximately six minutes to two. For the first time, the bank wasn't simply using AI to support an operation. It was putting an agent directly into the operation itself.

Ensuring confidence in agentic decisions

In banking, knowing when not to automate is just as important. Every Sentinel recommendation includes a confidence score, supporting evidence, reasoning, and an audit trail. High-confidence transfers can move automatically. Lower-confidence cases go to a banker. If an API returns incomplete or inconsistent data, confidence falls and the transaction moves to human review.

And when that happens, the banker doesn't start from scratch. Sentinel provides the evidence and explains why the case needs attention. When the AI is confident, it acts. When it isn't, a person takes over.

Bank Hapoalim didn't turn that autonomy on overnight. Initially, bankers confirmed or overrode Sentinel's recommendations. As the team validated its performance, the process progressed toward straight-through automation.

First, recommend. Then, validate. Then, automate.

Built on Dataiku, Sentinel coordinates LLMs, internal data, business rules, operational systems, and human review in one governed workflow. Now the bank sees Sentinel as more than a foreign-transfer solution. It's a tested blueprint for introducing AI agents into other regulated banking operations. The agent handles the repeatable decisions. People handle the ones that require judgment.