How a $35B retail bank built its own agent on Wonderful
Bank Hapoalim is one of the largest banks in the Middle East, with over 250 branches serving a multilingual customer base across every major retail banking product. Banking at that scale means the integrations required to make customer-facing agents effective are significant, and that work continues long after the first deployment.
Bank Hapoalim didn't want to outsource their agent development indefinitely. They wanted internal teams to be able to design, build, iterate, and maintain agents without relying on a vendor for every deployment. Wonderful's model made that possible: forward-deployed engineers do the upfront work to get the first agents into production, while actively transferring knowledge and capability to in-house teams.
The first test: a savings campaign at scale
The bank had identified a meaningful opportunity for a new savings campaign: a segment of roughly 100,000 eligible customers (people without active savings accounts, loans, or mortgages) who could benefit from a promotional first deposit of about $150 and be incentivized to open up their first savings account. With a large audience and a defined promotional window, the bank needed a way for customers to get information about the new product, verify their identity, and confirm eligibility without routing every interaction through a human agent.
They assigned a dedicated data engineer from the bank to own the build, someone with no prior experience building agents.
Building it
Working within the Wonderful platform and leveraging the tools and integrations already built by Wonderful's FDEs, the data engineer coordinated with internal teams across identity, eligibility, and messaging systems and deployed a production-ready agent in 3 weeks.
The agent handles the full lifecycle of the savings campaign interaction: explaining the offer, walking customers through identity verification via ID and one-time passcode, checking eligibility against live bank systems, and confirming enrollment. It can also send SMS confirmations, forward calls to a human agent, and trigger satisfaction surveys, across both voice and chat, in two languages.
The architecture is straightforward: a RAG source containing campaign and product information, paired with 15 custom tools connecting the agent to backend bank systems. Those tools were originally built by Wonderful's FDEs for earlier deployments; the data engineer adapted them for this new context, reducing build complexity and time to production.
Impact
The campaign launched on schedule, targeting the full 100,000-customer segment with an expected 40,000 interactions a month.
Containment rate for the voice agent is 88%.
The agent handles voice and chat across two languages without degradation in either channel.
The metrics reflect a working agent, but what they don't capture is what the bank now has: an engineer who knows how to build the next one, and an internal practice that can handle agent development and management.
The organization is already identifying new workflows where agents can replace or augment manual processes, and they're not waiting for a vendor to start.
