DBS Bank chatbot
At DBS Bank, I collaborated closely with the SME Chatbot team, which did not have a dedicated UX designer. Alongside my work on the design system, I supported this team by providing design guidance wherever needed.
This project had unique dynamics. The chatbot was already useful for quickly answering customer questions, but the team struggled to demonstrate real value internally.
In a large organization with distributed feature ownership, we also faced the challenge of overlapping products: how do you justify two tools that serve the same function while encouraging collaboration rather than competition?
Defining the vision
The first step was alignment. Terminology like “chatbot” and “virtual assistant” was often used interchangeably, internally and across the industry. We needed a clear definition as to what we are working towards:
| Parameter | Current | Future |
|---|---|---|
| Communication channels | Single | Multiple channels including emails and SMS service |
| Adaptability | Non-adaptive | They can learn as artificial learning and Machine Learning are integrated |
| Linguistic acumin | Limited | More advance and accurate understanding of queries |
| Response chain | Little to no memory | More conversational and has the ability to recall details |
| Understanding context | Fails to understand queries when specific keywords are not present | Understanding of context and intent |
Virtual Assistant's feel like a real assistant, communicating across channels, understands complex queries, context, and intent, and most importantly, is proactive. Chatbot's on the other hand are primarily reactive, answering user queries on demand.
Roadmap planning
We realized we couldn’t rely on standard feature by feature implementation. For example, the team was working on a money transfer status feature because the current interface didn’t clearly display status or explain its meaning.
I questioned the rationale behind this request. Should the chatbot act as patchwork for gaps in the product, or genuinely add value? These type of feature requests also prompted me to make a simple framework to categorize features and understand if it is valuable or not.
By establishing this framework, we could also set business expectations around the lifespan and impact of features implemented in the assistant. The transfer status feature? A Category 01: important now, but likely to be superseded by a simple dashboard update.
We need to prioritize features that give the chatbot long-lasting value, creating functionality that can’t be easily replicated elsewhere in the product.
Proactive assistant
Small businesses tend to review their financials at predictable times throughout the year. The virtual assistant should learn these patterns and prepare reports before they’re requested, anticipating user needs rather than reacting to them.
Chatbot vs. virtual assistant ability range
Another opportunity lies in local insights. The bank already offers market data and news, the assistant could digest this information and proactively deliver what’s most relevant to each business.
A dashboard that adapts to user behavior
We recently launched a dashboard with customizable widgets. One of the proposals for the virtual chatbot was to manage this dashboard for users. The experience a user may need on a mobile may be different than on a desktop. Suggesting to hide or show different widgets based the platform and behavior of the user may create a more seamless approach.
Of course, we don't want elements suddenly disappearing, so this is one of those cases where the user confirms with the virtual assistant that it's something they'd like to do.
This enabled users to rearrange widgets to their needs, but also customize the view of the widget, from having shortcuts, or key data, plus a list of information. These widgets made implementation of features easier to manage and created a micro-system of consistency.
Understanding language, context, and lingo
Designing a virtual assistant means making sure it understands how people actually speak across different contexts, languages, and regions. For example, a user might type “Exchange 500 SGD to USD.” Do they want to view the exchange rate, or complete the exchange? Interpreting intent depends on understanding phrasing, slang, and tone.
As Bank of America shared challenges they faced in their 2018 chatbot release:
Bank officials have been surprised by the language people are using to ask their questions, using slang like “dough” instead of “cash.” Attempts are being made to adapt the product. “We sometimes ask ourselves, why didn’t we think of that?”
It requires enormous amounts of data to feed even a simple chatbot. To make it more proactive is tougher still.
In Singapore, with its mix of four official languages and many dialects, it’s especially important to bridge the gap between how people speak and how the assistant responds.
Simplifying the language
The language used in chatbots feels robotic, and long-winded. What I mean by this is many confirmations on actions. Another goal of ours was to simplify the output, if the chatbot has an 85% confidence on the ask, it should output a proper response rather than confirming what the customer just stated.
Our dear content team made a great video on this very topic. You may also read more about it in the article published on Medium by our lead UX content writer Liva.
Customer service approach
Focusing on building a proactive virtual assistant doesn’t mean it can’t handle traditional customer inquiries. After all, it is a low hanging fruit that can be addressed swiftly. For instance, the transfer status feature was requested due to the high volume of calls the bank receives on this topic—costing the bank significant resources.
In the future enabling the assistant to authenticate users and access account information, we could answer questions directly about specific transactions, reducing call volume and improving efficiency.
Finding our implementation strategy
Where do we start? Many of these features are long-term and complex, so we prioritized creating awareness and early value for the assistant. Wherever Joy appears in the product, it must also be able to answer questions in the chatbot.
This method aligns with research from Bentley University (2017):
People create first impressions with chatbots just as they would when meeting people. ...people reach into their own experiences to create lasting impressions.
To ensure success, the assistant must feel valuable from day one. We begin with static experiences, onboarding steps and form assistance before gradually addressing open-ended questions.
Looking ahead, a dedicated application could enable voice interactions, significantly increasing accessibility and engagement.