Article | 2024
UX Meets LLM Chatbots
Advancements in large language models (LLMs) and unsupervised learning have redefined chat interface design. Here are key elements to create clear, agent-based human–AI interactions.
As product designers, we’re navigating uncharted waters—where prediction, not certainty, drives user interactions. This isn’t about "hoping" the system delivers accurate responses but about embracing a probabilistic model where perfection is elusive, and adaptability is essential. Designing for this paradigm can turn a bit unpredictable, occasionally frustrating and somehow weird.
Need finding, content strategy and early testing, are foundational to design effective LLM-powered chatbots. While these practices are UX pillars, their importance has soared with AI-driven conversational interfaces. Check out Raluca Budiu’s excellent research on bots, a great guide for getting started.
Content Strategy in LLM Chatbots
Content strategy for chatbots revolves around crafting meaningful, engaging, and contextually relevant interactions. Here are key elements to consider:
1. Welcome Message
"Hello! I’m a virtual assistant. Here’s how I can help..."
A friendly, concise introduction sets expectations for what the chat can do. It is important to get alignment on whether the bot should have a name; in any case, always clarify that users aren’t interacting with a human.
Define the bot’s scope and value clearly, which will be linked to the machine model in use but ultimately tailored to project goals. If feasible, think of contextualizing greetings based on user behavior or time of day (e.g., "Good evening! Need help wrapping up your day?"—assuming your bot supports dynamic usage).
2. Task-Oriented Prompts or Actionable Suggestions
Understanding the bot’s purpose and target users is critical – Surprise, surprise! (says the design thinking expert). Is this bot for troubleshooting? Product demos? Something else?
Establishing the bot priorities will be essential to design flows and prompts for tasks like:
- Common actions: "Track an order," "Find a product," "Get troubleshooting help."
- Seasonal or service alerts: “The system is currently experiencing an issue in your area due to a power outage,” "Ready to track your holiday orders? Click here!"
- Dynamic prompts based on user history or context: "You recently reviewed a pump. Need help pairing a sensor?"
Designing chatbot content is less about words, and more about building trust through meaningful, guided, and context-aware interactions.
3. Guided Flows for Key Tasks and Educational Tips
Use progressive steps to simplify complex tasks (e.g., account setup or troubleshooting). The algorithm’s role here depends on the desired flexibility:
- Decision-tree structures work for simpler, constrained flows.
- Context-sensitive, flexible bots require robust training but come at higher costs (e.g., pay-per-token).
Enable users to type keywords for instant answers. This is where feeding the bot with well-organized documents becomes incredibly useful, alongside the critical tasks of discriminating and tagging content to accelerate machine training and validation.
Additionally, consider including helpful process tips like: "Did you know I can send reminders when the [task] time is up?", “Did you know you can just talk instead of type?”
4. Fallback Responses and Error Handling
Designing for failure is imperative. The bot must acknowledge misunderstandings while offering alternatives. Avoid vague responses like, "Sorry, I don’t understand that."
Instead, consider:
- Clarifying questions: "Can you rephrase your question?"
- Suggestions: "Here’s what I can help with..."
- Searchable FAQs: "Not sure I understand, but these resources might help."
- Scope-driven replies: "I can’t do that, but I can assist with..."
The approach should demonstrate empathy, proactivity, and a focus on guiding users toward genuine solutions.
Other Key Content Strategy Elements
1. Navigation and Accessibility
- Allow seamless navigation without feeling like a traditional website. Sometimes just a hint is needed for the user to understand they can ask again: “Go back to the [topic] and explain [other option] with this [selection]”.
- Offer multilingual options and support for voice commands: "¿Prefieres español? Haz clic aquí."
2. Expectations, Feedback, and Haptics
- Include indicators (e.g., ellipsis to reflect the bot is typing/processing), visual, sound feedback, haptics, and progressive responses to mimic a more natural conversation flow.
- If needed, offer closing prompts (e.g., "Anything else I can help with?") to signal the end of an interaction.
3. Tone of Voice
- Sophisticated bots can change the tone of voice, remember to maintain consistency across interactions and channels or use a neutral tone.
- Depending on the objective and bot feasibility, rather than personalizing the tone of voice for each user, it could be better to adapt it to the context, prioritizing clarity and brand alignment. Think of an upbeat bot. Would it be acceptable to maintain a constantly overly cheerful tone, even when discussing health-related topics?
4. Transparency and Escalation Options
- Be upfront about the bot’s limitations and biases.
- Show where information comes from (e.g., paid content).
- Update accuracy details (how many options are displayed, when was the content updated).
- Ensure sensitive data is handled ethically and provide visibility of this information upfront.
- Provide clear human escalation paths, a phone number at least. Even if the bot cannot offer the answers, it could provide the basic information allowing users to ask straightforward questions, beneficial for both the user and the business.
5. Omnichannel Integration
- Redirect users to better-suited channels when appropriate, use the chatbot in addition and in collaboration with other channels.
- Allow advanced users to type quick queries for direct links to relevant info.
Testing LLM Chatbots: Wizard of Oz (WoZ) as a Technique
In addition to traditional prototyping, WoZ testing is invaluable. By simulating chatbot interactions through human operators, teams can:
App continuity is like the glue holding it all together. Imagine folding and unfolding your device like it's a high-stakes game of origami—your app should seamlessly transition between states without leaving users feeling like they've been thrown down a rabbit hole.
- Validate initial state and decision tree prompts.
- Test tone of voice and alignment with user goals.
- Evaluate fallback scenarios to ensure responses feel empathetic and actionable.
Tools like Voiceflow or Botmock are great for interactive mockups. For simpler setups, an online word processor works fine for mimicking scripted interactions. Use realistic use cases to identify pain points and refine response logic—ensuring the bot feels context-aware and helpful.
In summary, designing LLM-powered chatbots requires balancing innovation with practicality. From crafting meaningful content strategies to conducting thorough testing, the journey involves creativity, collaboration, and a sprinkle of patience. Like a good chat, it’s all about understanding and adapting to user needs—and sometimes learning to say an honest, "let’s try that again!"