Project Overview

Problem

Adobe Audience Builder helps users build a selected customer dataset, deciding who receives the campaign message. However, ❶ it is tedious to translate customer persona into data fields, and ❷ not all marketers have the data background to interpret the results of the prediction model.

Solution

Designed an AI tool that allows user to build datasets in natural language and provides model interpretation support in progressive disclosure and visualization.

Context

Imagine you're a marketing manager at Marriott

You're sending upgrade promo emails to guests who already booked a standard room. From that group, you want to narrow down to the ones more likely to pay, based on customer data.

This could take you hours. First❶, you need to translate "who booked a standard room" into concrete data fields → booking dates within a certain range, booked directly through the hotel website, opted into email, and so on. You're likely to get lost in different data field names. 🫨

Then you run a data model that's supposed to help identify customers who are more likely to pay for the upgrade. ❷Now you're staring at visualizations packed with terminology you don't fully understand. 🤯

Research

I interviewed five marketing teams from different industries and company sizes

I learned four things through interviews

❶ The user journey of a success workflow
❷ Edge cases to consider for AI agent responses
❸ Users' data skill level and scaffolds they need
❹ User mental model and top needs when running a marketing campaign

🎙️

[Company]-[Participant]-Marketing Team Interview & Cognitive Walkthrough

💡

Goal:

Understand how marketers currently build campaign audiences today, and where the process breaks down
Understand user mental model in building data sets and their top priority needs¹

Before you start

Confirm 30 min, recording consent

Send agenda ahead of time

Have follow-up prompts ready

Walkthrough questions

Could you give a brief overview about your role, your team and your background in data analysis?
Could you walk me through how you decide and build your audience dataset in your last campaign?

Possible follow-up questions

Input

What categories of campaigns do you usually work on?
What decisions are made by executives before you start working on audiences?
Would you have multiple KPIs for one campaign?

Output

How long does it take to onboard a new colleague on interpreting the model results?
How often and on what occasion do you reach out to the data team for support?
What are the biggest pain points in understanding model results?

Decision making

How do you decide one audience is better than another? What are the critical information you need?
Who makes the final decision? How's the timeline look like?

Wrap-up

What would a better workflow sound like to you?
Anything we didn't ask about that we should have?
Who else on your team should we talk to?

I mapped out the end-to-end workflow, and identified the key pain point to solve in each stage,

The journey is divided by the data model because it takes hours to run
In the input phase, the key user task is to set up all data constraints for the data model. The design focus of the AI agent is to make sure it covers all edge cases, not just the typical success use case.
In the output phase, the key user task is to interpret model results and make decisions. The design focus of the AI agent is to break down the complexity of the results.

Creating an AI agent from 0 -> 1 involves multiple big decisions. I sneak peeked the competitors to move faster.

I researched the entry and layout of competitive products, and decided on going with the second pattern because the workflow requires larger space for complex output and users want to track their progress.

Design Solution

Chat + templated canvas: talk, track and predict

With this design pattern of a chatbox and a canvas, I brainstormed with my two fellow designers and finally decided on a templated canvas where users can see what's going to happen upon landing on this agent page, which helps address a common fear of "jumping into a rabbit hole" among our time-strapped marketers.

Entry Point

The workflow starts with an overview of the campaign or a PRD file. Users can view the AI's interpretation of their input on the canvas and correct it if there's any gap.

Confirm generated data fields and run the model

From manually inputting those data fields to simply reviewing and revising them, users save much time in preparation for running the data model.

Review and compare model results

The in-conversation visualization card helps users understand model results and make decisions. Users can leverage the canvas to track their thinking process.

Finally, decide and publish

Users can archive any audiences they don't like to reduce cognitive workload. Once everything's locked, users hit the primary CTA on the top right to move forward.

Design Decisions

Input phase: use clarification questions for ambiguous input

I collaborated with our Adobe AI Engineer to determine what information is essential for the data model to generate reliable results and avoid a "rubbish in, rubbish out" experience.

Output phase: progressive disclosure of model results

Raw results from the data model are confusing, but they don't have to be. I categorized the numbers and translated them into a format that can be easily comprehended.

❶ Convert and label numbers to reduce noise
❷ Add context information
❸ Tuck information that isn't essential to decision-making deeper into an expandable card

Design Handoff

Design file organized by workflows, plus a Loom video

Like most designers, my own design file is, of course, a collage of all the iterations and discarded ideas. But I always tidy up the work that's ready to ship into a separate file for developers, for clarity. And I always find that attaching an async video first saves a lot of meeting time.

Contribute to the Adobe Spectrum design system

All UI is designed with Adobe Spectrum's tokens, including color, typography, spacing, and grid. For new UI elements I created for this project, I built them as reusable components and templates for scalability.

Take Away

Sixian Chen · 1st

Product Designer | AI UX, B2B, Internal Tools, Complex Systems | MS HCI/d at Indiana University

2d·🌐

⋯

This was my first time designing an AI agent, and it reshaped how I think about design. I spent most of my time designing around edge cases Traditional UX design starts with the primary flow and treats everything else as an edge case. Designing for AI reverses that. The model will misread intent, return a low confidence result, or simply get it wrong, and that isn't the exception, it's the baseline condition I had to design around. The real work wasn't polishing the ideal path. It was asking, again and again, what the interface should do when the model is uncertain, or wrong. Engineering constraints shaped as many decisions as user needs did This was also the project where I sat closest to engineering. User needs used to be the main constraint on my design decisions. Here, what the model could reliably do mattered just as much. I spent real time learning the model's capabilities so I could tell which features were worth building and which needed a fallback state, before a single screen got designed. Understanding the system became part of designing for it.