Designing Leumi’s AI banking agent

Overview

How might we reduce banker inquiries by making routine banking tasks fully completable within the chat?

Leumi AI is an AI agent embedded in Leumi’s app, Israel’s #1 ranked digital bank. It answers questions, surfaces account data, and completes full banking flows entirely within a chat interface. It adapts its response to the request type:

Problem

65% of all inquiries traced back to a small set of recurring needs.

Leumi Bank handles tens of thousands of digital banking interactions daily. Despite having a fully featured app, a significant portion of users were still reaching out to human bankers for routine tasks: transfers, balance checks, IBAN lookups, deposit management.

The existing digital flows were functional but fragmented. Users dropped off mid-process, missed key steps, or didn’t know the capability existed within the chat. Options that kept users inside the chat looked identical to those that sent them out of it, creating confusion at every decision point.

Opportunity

Tasks that should take seconds, completed inside the conversation.

Solution

Leumi AI: an agent that resolves, not just responds.

Leumi AI turns the bank into one conversation. Customers ask in their own words, and the agent understands what they need, pulls the answer from their own accounts and completes the task right there in the chat. No menus to dig through, no jumping between screens.

Every answer arrives as a designed component rather than a wall of text: a status tracker for a pending request, a consent form to review and sign, a breakdown of the month’s largest transactions. When something genuinely requires a banker, the handoff is immediate.

Core flows

Increasing a credit limit

The user asks once, confirms a few details, and signs the agreement, all within the chat.

Requesting a new loan

A loan request that is rejected after the agent checks eligibility, with a clear next step.

Checking cards linked to foreign currency

The user asks which cards are linked to their foreign currency account. The agent answers with an info card listing each card and its currencies, and links straight to currency exchange in the app.

Design decisions

Four principles behind every message.

1. Information and action look differentUsers always know whether they’re reading or doing.
2. Personal and conciseThe agent addresses users by name and keeps responses short.
3. Full flows, inside the chatUsers can complete actions end to end without leaving the conversation.
4. The user’s bubble is Leumi blueIf you’re with Leumi, you’re at the center.

Components

More component types.

Icon design

Why we chose this symbol

We mapped the icons of existing AI assistants on two axes: familiar to unique, and abstract to figurative. Most sit at the edges, either generic sparkles or cartoon characters. The Leumi AI orb sits in between: abstract enough to read as AI, soft and round enough to feel personal, and in Leumi’s own blue.

Impact

30% fewer banker inquiries in the first pilot.

Leumi AI launched as a pilot in a single branch, Pinkas in Tel Aviv. During the pilot, inquiries to bankers dropped by 30%.

Customers received the chat well. Early feedback pointed to first-version errors, which we’re fixing as we go, and the agent keeps evolving.

Reflection

• Designing behavior, not screens. An agent’s answers can’t be fully predicted, so a set of mockups was never going to be the spec. The real design work was the decision logic: when a reply should stay text, become an info card, link out to the app, run as a guided flow, or hand off to a banker. Every component only made sense once those rules were clear.
• Trust comes from what the agent admits. In a bank, a confident wrong answer costs more than a slow right one. Sensitive or unclear requests had to step back, ask a follow-up question or pass the conversation to a person with its full context. Designing those moments of uncertainty was harder, and more important, than designing the happy path.
• The model sets the material, not the experience. The same request can come back long, short or slightly off. Components had to hold up across that variability, and the copy rules, tone and structure of a reply became as much a part of the design system as the UI itself.