Empowering our design team to scale rapidly with Dust
As we navigate the rapid currents of a fintech scale-up, agility is not merely a perk — it’s a prerequisite. Our design team, expanding rapidly, reflects the escalating scope and intricacy of our projects. This environment demands to be in a solid enough position with enhanced efficiency to adeptly manage ever-growing challenges. We need to manage and access proliferating data for comprehensive research while integrating UX Writing as a new discipline to enhance user experience. As our responsibilities expand, the day still has only 24 hours.
To tackle these challenges and amplify our capabilities, we turned to Dust — a tool that automates time-consuming tasks. In this post, we’ll introduce you to this solution and explain how it helps us save time amidst our rapid growth.
Introducing Dust: tailored assistants for augmented operations
Founded in 2023 in France, Dust empowers organisations to create customized AI assistants by leveraging their comprehensive knowledge bases. Due to its French origin, Dust adheres to stringent EU data protection regulations, ensuring robust security and compliance. This foundation positions Dust not only as a tool for operational efficiency but also as a secure and trustworthy platform.
What’s the difference with ChatGPT?
Unlike general AI models like ChatGPT, which primarily respond to prompts based on a fixed dataset, Dust specialises in integrating a variety of data sources with different cutting-edge language models such as GPT-4, Gemini 1.5 pro, Claude and Mistral. This integration allows for dynamic and intelligent automation solutions that are tailored specifically to the needs and existing data of an organisation, offering a more customised and interactive experience.
Who can create AI assistants on Dust and how?
Dust democratises the creation of AI assistants, making it accessible for anyone at Pennylane, regardless of their technical skill level. This inclusivity ensures that all staff members can take advantage of AI tools to streamline their work:
The creation process is straightforward:
- Define the goal: Users start by setting a clear objective for the assistant.
- Customise the setup: They then customise the assistant’s capabilities, tailoring it to access and analyse the required data sources.
- Deploy and interact: Once set up, the assistant is ready to interact with, providing insights, automating tasks, or supporting decision-making processes.
Practical applications of Dust in Pennylane’s design team
We wanted to go beyond simply allowing anyone create assistants with Dust. So we developed several specific Dust assistants to make life easier for our designers. Here are the 3 main use cases we’re addressing:
Accessing user feedback
To streamline the search through our vast array of user feedback, we leveraged Dust to speed up the process. Our data sources are rich with valuable insights, but aggregating this feedback is no easy task. Our first goal was to enable designers to easily access and search through this data.
To make this happen, we created several assistants. Each linked to different user feedback sources like our ticketing service and various Notion databases. These assistants are given a basic overview of Pennylane and are tasked with acting as allies to the product team, helping to extract valuable insights from our data.
These assistants have been incredibly valuable, particularly when starting new projects. They enable designers to quickly gather feedback from various sources, which helps in defining and challenging their initial assumptions — an essential part of our user research process.
Enhancing UX Writing
The second key use case for our Dust assistants focuses on UX writing. These assistants are an excellent tool to enhance the quality of work a designer can achieve before requiring input from a UX writer.
We’ve set up a UX writing assistant that has access to our internal UX writing guidelines. Designers can use this assistant to help create written content within the app, such as modals or alerts. The process is straightforward: paste in the sentence you’re working on, and the assistant will suggest variations that align with our guidelines.
We’ve made it clear to designers that they shouldn’t just copy and paste the content generated by the assistant directly into the app. Instead, the assistant’s suggestions help them explore different options and refine their drafts.
We have seen several benefits to this UX writing assistant:
- It allows our UX writers to focus on more complex projects
- The assistant helps non-French speakers with their copy by providing accurate translations through our glossary
- It enables designers to interact with AI about their copy questions, alleviating any fear of judgment on their current work
Navigating and translating a technical product
Our final use case addresses the specialised vocabulary of accounting. This terminology is specific to the profession, and we can’t expect designers to be familiar with it, especially in multiple languages.
Traditional translation tools like Google Translate often fail to accurately translate this specialised vocabulary. To ensure we use the translations we’ve defined internally, we’ve created a glossary assistant connected to our internal glossary. This assistant helps designers accurately translate accounting-related language, maintaining precision in our communications and our product.
Evaluating Dust: benefits and limitations at Pennylane
Since Dust is based on large language models (LLMs), we still encounter some of the common shortcomings of this technology. Hallucinations do occur, but these are relatively easy to spot because we always ask our assistants to add their sources. The most frequent pain points we face are as follows:
- Exhaustiveness: Even with the largest LLM models available, Dust never seems to capture all the feedback on a given subject. We consistently find items that the bot misses.
- Prompt struggles: Both in crafting assistant prompts and the questions asked to the assistant, it’s challenging to understand what positively impacts the prompts. It often turns into a game of trial and error, which can be very time-consuming.
- Integration issues: Some of our data sources are difficult to use with a Dust assistant, particularly websites that require users to log in. This limitation prevents us from having one assistant that can summarise feedback from all sources.
Despite these challenges, we’ve seen the product improve over time. Some integrations have been added quickly, such as the Intercom integration, which has been a positive development.
Charting progress and iterating with Dust
While there is still room for improvement in our use of Dust, our enthusiasm remains high. As one of their first customers, we have the opportunity to help shape the development of a product that not only resonates with our values but also promises significant enhancements for our operational framework.
This collaboration fits perfectly with Pennylane’s forward-thinking ethos, we’re excited to see how Dust evolves to play a pivotal role in our processes. The journey of innovation and improvement continues, and we’re thrilled to keep working with a company that helps Pennylane stay AI-savvy.
A story by Gualtiero Mottola, Product Designer and Ludivine Kasteleyn,
UX Writer at Pennylane
