Trustworthy and Assistive AI Chatbot

A design concept for a more productive internal business support agent

The Problem

AI Chatbots are known to suffer from hallucination where text generated can be factually wrong, irrelevant or even nonsensicial.

With the scope of an internal business IT/HR support chatbot, how might we design a chatbot that users feel they can trust?

Process

We applied a double diamond approach to this project, seeking to understand the issue and defining problem in the research phase followed by developing and delivering a solution in the design phase.

Research Phase

The research phase followed the process of:

• Qualitative data gathering and analysis
• Quantitative data gathering and analysis
• Data triangulation and defining themes/design suggestions

Design Phase

The design phase followed the process of:

• Sketching
• Wireframing
• Usability testing
• Iteration of design based on feedback followed by another round usability testing for mid and high-fidelity
• Final design

Research

Resarch Questions

1. What are users' expectations and current perceptions of chatbot as a solution at workplace?
2. What design aspects of AI chatbots would be most likely to build and maintain users' trust over time?
3. What accessibility considerations should be included in the design of AI chatbots?

User Interview and Contextual Inquiry

We had conducted 12 sessions of semi-structured interviews coupled with contextual inquiry with participants who are in industry, some from our sponsor company. This was done to understand the expectations and preceptions users have of AI chatbots. The contextual inquiry consisted of providing participants with several workplace scenarios which they will try to solve the issue using chatgpt(instructed to act as a workplace chatbot)/workplace chatbot. This allowed us to probe at perspectives that may not have been mentioned in the interview.

Survey

We released a survery which gathered 42 responses to validate findings from the user interview and contextual inquiry through quantitative results. The survey consisted of a likert scale to capture possible nuances to their preference, preference testing and a follow up open-ended questions to understand the reasoning of their preferences.

The results such as preference for a certain style of answer had better informed our designs decisions before proceeding to the design phase. The special case of a 43%(concise answer) and 46%(detailed answer) divide of answer detail level had led us to a new idea of giving the users the ability to switch between the two when they felt necessary.

Research Findings

The qualitative and quantitative data had resulted in 3 main themes with several sub-themes.

A challenge of trustworthiness in AI Chatbot

Lack of credibility

Lack of transparency

Data privacy concern

A need for a more engaging conversational experience

Lack of human-likeness

Lack of contextual understanding

Lack of flexibility

A need for chatbot's assistance to boost productivity at work

Task execution

Boosting work productivity

Design Suggestion

We decided to tackle the themes above through prompt engineering coupled with UX design. Prompt engineering will be able to improve the chatbots's response while UX design will be able to improve on the experience and visual aspects whilst using the chatbot.

Prompt Engineering

Image below shows an example of a chatbot response after being prompt engineered by us which we implemented the design suggestions from above. Response was generated using Langchain, OpenAI API drawing information from openly available data from Nvidia's website.

Sketches/Wireframes

The initial design was inspired by Chatgpt as it was what most users are expected to be familiar with. Sketches by teammmates were collated and certain design of features were decided on to proceed to the wireframing stage.

Sketches

Wireframe

Usability Testing

3 rounds of usability testing were conducted - once with the wireframe, once with the mid-fidelity design and once with the high-fidelity design, all with new participants as we wanted to also see if useres will be able to intuitively use the system.

Round 1

Testing with the wireframe showed that users preferred to interact with the bot through typing rather than using pre-generated templates when using its productivity features, finding typing to be more efficient. A participant also mentioned for more integration of productivity features which led us to include meeting calendar, to-do list and notification as part of the features.

Round 2

Testing with the mid-fidelity prototype showed that users preferred real-time editing of task support as well as branded UI to show a better link of task and software. Another finding was that the purpose of the bot was unclear which led us to add informational buttons and a help page.

Round 3

The main purpose of the final round was to identify any bugs as well as to help finalise the design with validation that users were satisfied with the prototype.

High Fidelity Design

Landing Page

An introductory message allowed for users to understand what the bot can help with as well as guide them on how to interact/use it.

Request Processing

Integration of a split-screen to see information updated in real time as well as the UI of the third-party software.

Further work

We would have liked to conducted more user testing of the final design to be able to integrate quantitative data and analysis to better develop and evaluate the prototype. During these user testing, having a diverse group of users will also be beneficial to uncover any accessibility pain points in the design.