Don't botch the bot: Designing interactions for AI
Summary
It seems like every company is adding a conversational AI chatbot to their website lately, but how do you actually go about making these experiences valuable and intuitive? Savannah Carlin will present a case study on a conversational AI chatbot—Marqeta Docs AI—that she designed for a developer documentation site in the fintech industry. She will share her insights, mistakes, and perspectives on how to use AI in a meaningful, seamless way, especially for companies like Marqeta that operate in highly regulated industries with strict compliance standards. The talk will use specific examples and visuals to show what makes conversational AI interactions uniquely challenging and the design patterns that can address those challenges. These include managing user expectations, handling errors or misunderstandings within the conversation, and ensuring that users can quickly judge the quality of a bot’s response. You’ll gain a deeper understanding of the intricacies involved in designing interactions for AI, along with practical advice you can apply in your own design processes. Take-aways What to consider before you add AI to your product to ensure it will be valuable, usable, and safe for its intended workflows The interactions that are unique to conversational AI experiences and the design patterns that work for them Common challenges in designing conversational AI experiences and how to overcome them
Key Insights
-
•
Defining a clear and specific primary use case is crucial before starting any generative AI chatbot project.
-
•
High-quality, thoroughly reviewed training data is foundational to delivering accurate and useful AI outputs.
-
•
Initial state messaging must clearly frame what the chatbot can and cannot help with to reduce irrelevant or off-topic queries.
-
•
Loading indicators for AI text responses should be subtle, with progress reflected by text appearing rather than distracting animations.
-
•
Supporting efficient scrolling and prompt review is vital since users frequently check and refine their inputs against often long answers.
-
•
Error states in AI chatbots shift from traditional fixed errors to helping users write better prompts to get more relevant results.
-
•
Transparency about accuracy, AI limitations, and source citations builds user trust, especially in regulated domains like FinTech.
-
•
Providing users with prompt engineering guidance via documentation significantly improves the quality of chatbot interactions.
-
•
Accessibility considerations, like keyboard navigation and screen reader compatibility, must be integrated from the start, especially given the large text outputs.
-
•
Chatbots can reduce customer support friction and encourage users to ask questions they might not have otherwise, enhancing user engagement with the product.
Notable Quotes
"If you have any doubts about the quality of the training data, do not proceed."
"You want to assist people in framing the interaction and setting their expectations correctly so they know how to be successful."
"The biggest difference with error states in AI bots is helping people write prompts effectively, not just recovering from simple failures."
"Loading text itself is a loading indicator; the letters appearing show progress better than jumpy animations."
"People often forget what they wrote and then want to check their prompt again before refining it."
"Every output should have at least three source links, almost like citations in a research paper."
"We had to be very careful about accuracy because we're in FinTech and compliance is critical."
"People started asking questions to the bot that they wouldn’t have taken the time to email about."
"It’s really important to be clear and transparent about what the tool is good at and what it’s not good at."
"Accessibility testing included making sure everyone could navigate it using a keyboard alone."
Or choose a question:
More Videos
"Doing work in one-week cycles fits the rhythms of companies working with agile and Scrum frameworks."
Victor M. GonzalezPracticing Learners and Learning Practitioners
March 10, 2021
"Toxic stress is experiencing too much too fast for too long without healing."
Alla WeinbergHealing Toxic Stress
September 23, 2024
"Maybe I too had joined another tech company that really wasn’t walking the walk when it came to diversity."
Libby MaurerTreating Diversity & Inclusion in Hiring as a Design Problem
December 5, 2019
"When design ops comes in hot, it means honest, orchestrated, and timely."
Briana Thomas Christina RodriguezWhen Design Ops Comes in H.O.T. : A Tale of a Transformed Design Org
September 30, 2021
"Every decision in a design system is like a puzzle piece; together they form a bigger picture that evolves over time."
Mitchell BernsteinOrganizing Chaos: How IBM is Defining Design Systems with Sketch for an Ever-Changing AI Landscape
September 29, 2021
"The best sources for tool info are other researchers in communities like LinkedIn and Slack, not really any independent compendiums."
Jon Temple Kalee Dankner Bruce Falk Lauren GalanterPanel: Stacks, Security, and Stakeholders: The Hidden Work of UXR Tool Procurement
March 12, 2026
"Auto captions on YouTube struggle with accents and technical terms, so manual or human-reviewed captioning is essential for accuracy."
Sheri Byrne-HaberThe Importance of Accessible Design Systems
January 8, 2024
"We organize our work around our life, instead of our life around our work."
Ana FerreiraDesigning Distributed: Leading Doist’s Fully Remote Design Team in Six Countries
January 8, 2024
"Doing at least two benchmarking rounds lets you see if optimizations have measurable impacts over time."
Daniela Magaña Flores Ariane Rahn Jeff Ephraim BanderAhead of Competition: Learn What UX Benchmarking Can Do for Your Business Today
March 10, 2022
