Overview
Adobe released its AI agents in the Adobe Experience Platform (AEP), which is a B2B software help marketers unify customer data and support campaign launch. I worked on one of the AI agent that helps marketers build campaign audiences.
Challenge
Users experienced decision fatigue when interacting with the audience agent because the conversation often went very long and included too much data terminology.
Solution
Goal-orientated Conversation
At the start of the conversation, AI agent asks user what the project is about and what audience they're going to build together. Once receiving user's input, AI agent translates the goal in canvas view to let users know their needs are well translated, and help users focus on this goal throughout the conversation.
To assure it's a goal-orientated journey, the AI agent would first fill this "audience description" card before it moves on to other cards. This also makes sure that the data propensity model at the back-end get the basic input to provide insights.
Canvas with Clear Steps
Before, the canvas goes with the conversation in real-time. Every time the user or the agent brings up one new task, there appears a new item card in the canvas, like a growing tree. This adds to the feeling of uncertainty as users don't know when it will end.
I converted the canvas to a pre-built template with clear steps to follow, based on the input the AI model needs to recommend an audience. Users are very clear what they are required to do and where the end point is in this work flow.
In this template, users first describe the project they're working on, and the AI agent will transcribe their input to item cards on the canvas. Sometimes users don't have or forget to mention about what the KPI is or what destinations they want, AI agent will take the initiative to ask for more clarify, which helps to build trust.
Recommendation Card
With sufficient input from users, AI agent is able to generate recommendations through data training model. Users will compare the options, make iterations and make decisions based on the model results.
Although the propensity score is the most objective and accurate metric, most marketers can be very confused about what the score indicates: likelihood to purchase, or model fit score?
The new recommendation card reflects the design principle that in human-AI interaction, users should be given full autonomy in decision-making. Instead of presenting a "propensity score", I chose to present two most important metrics in plain words to facilitate decisions.
Through user research, I also learnt that this simplified recommendation card can be insufficient for (a) expert users with data background, (b) final check before launch. So I designed this card to be expandable for better user segmentation.
Research
Given the constraints of an industry-sponsored project, direct access to enterprise users was limited. I addressed this challenge by exploring the problem from two complementary directions, as shown in the figure.
I created this "fatigue map" to highlight where users feel exhausted during the journey, which is a synthesis of research insights.
I also created a signature use case that reflects marketers' mental models, which emphasizes the iterative nature of their workflow.
Design Iteration
Design principle — those are strategies derived from research and sketches, which support finer-grained design decisions in later phases.
I dropped this branch feature because it made users explicitly manage decision paths, which conflicted with the proactive principle.
