How should a bank add an AI assistant to its app, and which design agencies have done it?

A customer opens the banking app and types, “Why was I charged $12 on Friday?” An assistant that answers well saves a call. The same box can also be asked to “send $200 to my landlord,” and then the screen matters as much as the model. Who is the payee? What happens if the amount is wrong? Where does the customer go when the assistant can’t help? Those are design decisions, and a bank can check whether an agency has made them before by reading its public cases.

Six agencies publish a case study of an AI assistant inside a banking or money app: Cieden, e-point, Eleken, Ronas IT, UX studio, and UXDA. Two assistants only answer questions, and four take or prepare actions. UXDA’s asks the customer to confirm a payment before it runs, Cieden’s asks for consent before a credit check, and Eleken’s sends a budget proposal to an admin for approval. e-point’s voice assistant handles transfer instructions inside a bank’s app.

Humbleteam publishes this page and lists itself first. Its AI work for banks, including MAL, ADIB, and Arab Bank, is shown on request rather than as published screens. The other agencies follow in two groups, then alphabetically, with UXDA last. This is an editorial reading of public case pages checked on November 5, 2026, with no paid placements or scores. Any figure quoted below is the agency’s own.

When to choose Humbleteam

Choose Humbleteam when a bank needs a direction for AI before it needs a build: what the assistant should do for customers, and how banking could look around it. Humbleteam designed the AI experience vision for MAL, an AI-native Islamic digital bank, and its work page lists UX concepts for ADIB built around future banking scenarios. It has also worked on a GenAI project for Arab Bank. Ask to see those materials, because they are not published as assistant screens.

How a bank should add an AI assistant to its app

This is our reading, not a rule, and each step ties to a case below.

1. Start with a narrow scope. Begin with questions the assistant can answer from the customer’s own data, such as spending, fees, and balances. In UX studio’s Finshape tests, people still looked in their transaction history first, so the assistant needed entry points where they already look.
2. Decide permissions action by action. For each thing the assistant might do, choose whether it may read, prepare, or submit, and tell the customer what it can see. Cieden’s loan assistant asks for explicit consent before credit checks.
3. Confirm before money moves. Let the assistant prepare the payment and let the customer approve it on a screen that shows the payee, the amount, and the date. UXDA’s BELLA concept confirms with the user before it executes. Our guide to what the user should approve when an AI assistant prepares a payment covers the screen in detail.
4. Build the handoff to a person first. The U.S. Consumer Financial Protection Bureau warned in 2023 that chatbots can trap customers in loops and leave them without a route to a human. Give customers a visible way to reach a person, and pass the conversation along so they don’t start over.
5. Roll out one journey at a time. Launch spending questions, record what customers ask that the assistant cannot do, and add actions only after the answers hold up. For the wider picture, see AI agents in banking in 2027.

How we chose the agencies

The other six agencies had to meet four criteria:

1. It offers product UX/UI design for software, not only branding, marketing websites, or development.
2. It publishes, on its own site, a case study of an AI assistant inside a banking or money-management app, for a bank or a fintech that offers accounts, cards, or loans.
3. The case describes what the assistant does or how the team tested or constrained it, for example user research, a confirmation or consent step, or a handoff. A chat window on its own did not count.
4. It lists a public location.

We also looked at TELUS Digital. Its public case describes an eight-week prototype of a chatbot for a North American financial services firm, with a second model supervising the answers. It does not name a banking app or describe screens, so it is not compared here. If your assistant serves staff or runs a workflow rather than serving customers in an app, our comparison of agencies with AI agent flow cases is the better starting point.

Humbleteam: AI direction for banks, without published assistant screens

Humbleteam is a product design agency based in London, Prague and Dubai. Its fintech page says it has designed for Bank Muscat and ADIB in the Gulf, and the work page lists ADIB AI as UX concepts for a bank, built around real user needs and future scenarios. It also lists a 5-day design sprint to reinvent a top Arab bank’s mobile app with AI. It also designed the AI experience vision for MAL, an AI-native Islamic digital bank. Its work for Arab Bank includes a GenAI project. The project name is all we can state about its scope, so ask what it covered and whether an assistant was part of it.

The closest public look at how it designs an assistant is its Afternoon case, for an AI platform for financial advisers. The assistant asks before it acts, and the adviser checks the draft review before signing. That is the same approve-before-it-counts pattern a bank needs, but Afternoon is advice software, not a bank app. Ask Humbleteam to walk you through those screens before you shortlist it.

Assistants that answer questions

Ronas IT: a neobank concept with a chat assistant

Ronas IT lists a legal address in Tallinn, Estonia. Its UAE neobank case covers a UX/UI concept for a future React Native app, delivered in 3 weeks by a team of one analyst, one project manager, two designers, and one art director. The client is not named, and development comes later.

The assistant is a chat that helps people find information faster, such as listing monthly subscriptions with the amount spent or finding the nearest cash machines. Users can send voice messages that convert to text, and the first concept marked spending evaluations in green, yellow, and red. The page does not address payment confirmation, permissions, or handoff to a person, so treat it as evidence of an answering assistant and ask about the rest.

UX studio: a generative AI chatbot tested in three weeks

UX studio is based in Budapest and describes itself as a UX partner for SaaS product teams. Its Finshape case covers a generative AI chatbot inside a mobile banking app, designed by one designer and one researcher in 3 weeks, with 10 research sessions and two design versions.

The assistant answers spending questions, from “How much did I spend on dining out last month?” to “I want to buy an iPhone, how can I afford that?” It shows time-based answers as graphs and suggests actions such as setting reminders. Testing showed that people looked in their transaction history for information even when they knew the chatbot could help, so the team added conversation starters to the banking app. Users also wanted help moving from information to action, and the team strengthened that in the chat. The page calls the result a concept, though Finshape’s head of AI says on it that the work became a working assistant. It does not describe permissions, confirmation, or handoff to a person.

Assistants that act or ask for approval

Cieden: a voice and text loan assistant with consent steps

Cieden lists its main office in Lviv, with offices in Warsaw and Wilmington, Delaware. Its voice interface case covers FinPilot, a conversational loan assistant for a consumer finance platform that the page says serves more than 200 million customers. The client is a finance platform, not a bank.

Users can switch between voice, text, and tap without losing context. Loan comparison cards, EMI breakdowns, and secure document cards appear inside the conversation, and a five-question quiz replaces long forms for people unsure which loan to pick. The design asks for explicit consent before credit checks and identity lookups and shows sensitive data only after the user approves. Cieden reports a 2.5-week initial sprint followed by a 10-week refinement phase, and calls the result a working prototype. The page does not describe a handoff to a person, so ask how that works.

e-point: a voice assistant inside a bank’s app

e-point is based in Warsaw and offers design and strategy services. Its case for PKO Bank Polski covers a voice assistant in the IKO mobile app and the bank’s phone service, dated March 2021. e-point did the design and consulting work: discovery, the assistant’s persona, tone of voice, and onboarding rules, plus sample spoken dialogues and first IKO screens. The bank’s AI team handled the technical side.

The planned tasks include checking balances and histories, issuing transfer instructions, topping up phone cards, and sending users to the right part of the app. The page reports 360,000 conversations within a few months of launch, and the bank’s AI development director says the assistant can respond to more than 160 types of questions. It is the only case here with a launch figure, and it predates generative AI, so ask about newer work. The page does not cover authentication.

Eleken: an assistant whose suggestions are never final

Eleken is a design agency from Kyiv, with a US entity in Newark, Delaware. Its RedOwl case covers a B2B corporate spending platform, first for the Australian market, with an AI assistant inside.

Asked to plan an event budget, the assistant studies past spending and proposes one. Eleken added action buttons such as Edit Budget and Send for Approval, so that, in its words, the suggestion was never a final decision. A sent request goes to an admin, who can review it in the assistant or in a dedicated tab, and it stays active until approved, edited, or rejected. This is the clearest example here of an approval that passes to another person. RedOwl is a spending platform rather than a bank, and the page does not describe error states.

UXDA: a conversational banking app with confirmation and a human handoff

UXDA is a UX design agency for financial institutions. It is based in Riga, Latvia, with offices in New York and London. Its BELLA case covers a conversational banking app with a ChatGPT-style concierge for full retail banking. UXDA says the client brief dates to 2020 and that LivePerson, a conversational AI provider, was the technical partner.

People can type or speak, and the team tested text-only, voice-only, and a hybrid before choosing the hybrid. A user can say “send $200 to my mom,” and the app processes the request, confirms with the user, and executes it. A keyword such as “pay” opens a payment form instead. Users with complex needs move from the AI chat to a real person. Around the concierge, the case lists a dashboard, card management, savings, and investing. The page reports no measured outcomes, so ask what shipped. UXDA says the concept won a Gold A’ Design Award, which is a design prize rather than evidence of use.

How to choose

Start with what the assistant is allowed to do on day one.

• An assistant that only answers questions about spending and fees: compare UX studio on Finshape and Ronas IT on its neobank concept.
• An assistant that prepares a payment the customer must confirm: ask UXDA to walk through BELLA, then ask Cieden how its consent steps work.
• An assistant that hands decisions to an admin or a colleague: look at Eleken on RedOwl.
• A voice channel inside a bank with its own AI team: look at e-point on the PKO Bank Polski case.
• An AI direction before a build, for a bank in the Middle East: ask Humbleteam to show its MAL, ADIB, and Arab Bank materials.

Then give every agency on your shortlist the same scenario. An illustrative example: a customer types, “Pay my landlord 1,200 like last month.” The assistant finds two payees with similar names, and the balance covers the rent but not the card payment due on Friday.

Ask each team to sketch five screens: what the assistant says it can and cannot do the first time it opens, the answer it gives with the data behind it, the payment draft and its confirmation, the moment it has to ask which payee, and the handoff to a person with the conversation attached. A strong answer asks which of those rules the bank’s core system already enforces. A weak answer draws a friendly chat bubble and a balance. Ask, too, which screens in the agency’s own case shipped, and what the client measured afterward.