Navigation AI Assistant
Artificial intelligence often flips the way that users interact with computers. Traditional interactions usually function like steering a computer toward a certain goal or building an idea from blocks. However, with AI, the user describes the goal (or final form) and the AI system attempts to build it from its model and training.
This fundamental shift in thinking was pivotal to this feature.
Tools
Collaboration
Highlighted Challenge
We knew users could easily become overwhelmed as a complicated platform added features. Many functions could be lost or buried in a large system.
AI interactions pose unique challenges: they are new, their reasoning can be unpredictable, and occasionally their results are unreliable.
Hypothesis: Could we implement AI in a helpful way that could build trust and benefit users?Solution
I designed our AI experience around the following principals derived from user research:
- Keep AI optional for the user
- AI should be focused and narrow in scope
- The experience should guide the user by suggestion
- Users verify AI responses before taking action
As mentioned above, I surveyed concerns internally and externally around application complexity, user overload, and specific issues important to how people respond to AI in applications. This research resulted in the principles explained in my solution below.
The Goal: Building User Trust
Keep AI features optional: When testing this tool it was clear AI was helpful for discovering unknown actions, but also more cumbersome than traditional navigation for functions users already understood. Therefore, our assistant was designed and trained to aid the user in navigation and feature discovery, but not to replace traditional navigation. In order to build trust, it is important for users to maintain the ability to choose.
AI should be focused: The AI assistant was given a very limited scope (initially), only being trained how to direct users to application features. This allowed us to minimize unexpected results. If asked questions outside the scope of its knowledge, the AI was trained to give friendly responses directing the user back to its limited capabilities.
The experience guides the user: Because AI interactions are still new, some users may be uncertain where to start, or what they can ask. I designed our AI experience to include friendly prompt suggestions, gently guiding the user toward things they could try, and demonstrating different ways they could ask questions.
Users verify AI responses: Once the AI assistant responded, the user was given the choice to confirm and follow the suggestion, try something else, or exit. For instance, if asked, "How can I update Alison's profile image?" the system would respond, "You can update Alison's profile image by editing her user page. Would you like to go there now?"
I added this validation step to keep choice in the hands of our users. It allowed them to verify results, experiment, and choose how the AI could help, rather than having it automatically make hidden decisions for them.
