Research framing, pattern definition, and the AEP examples.
Timeline
12 weeks, sponsored studio project.
Team
Seven designers.
Mentors
An Adobe experience designer, with a researcher and an AI engineer supporting.
Problem
Adobe added AI agents to its audience builder to save marketers effort. The agents answered every question and handed back a decision each time, so a single task turned into nine of them in a row.
Approach
I traced one real conversation to find where the decisions pile up, found three moments that recur in any agent, and wrote a rule for what the agent owes the person at each one.
Impact
Seven universal interaction patterns, built and tested inside Adobe Experience Platform. 2 minutes off the task.
our marketer
Needs
An audience that fits the brief.
Does
Describes the goal to the agent.
Gets
Six audiences, then two more questions.
Reacts
Picks, and hopes the first choice held.
Adobe gave the marketer an assistant to make this decision. I watched it hand back more decisions than it took away.
Twelve weeks to work out why that happens, and what an agent owes the person on the other side of it.
01The problem
Adobe Experience Platform is where marketers build audiences. They translate broad goals like "reach high-intent young shoppers" into precise, machine-readable segments drawn from behavioral events, demographics, predictive scores, and more.
To reduce that effort, Adobe added AI agents that summarize data, suggest audiences, and interpret intent. But the AI helps by offering more: more suggestions, more steps, more to weigh. Every suggestion becomes another decision, and now you're reading the data and second-guessing what the AI thought you meant, all while making back to back decisions.
Find a good audience to upsell luxury experiences.
Six audiences look relevant.
Business travelersWeekend travelersFrequent flyers+3 more
?Which one, and on what basis?
Weekend travelers.
Got it. Which channels should this run on?
EmailPaid socialPushIn-app
?Was that first pick even right?
Should I exclude the customers you set aside earlier?
?Which ones did I set aside?
By the ninth question the marketer has stopped weighing the answers and started guessing.
That is decision fatigue. Choices get worse as they accumulate, and at some point the work gets abandoned rather than finished.
02The moments
I went looking for the hardest decision in that conversation. There wasn't one.
Every question was answerable on its own. What wore people down was answering nine of them in a row, which a difficulty ranking will never show you. So I went back through the same conversation looking for where the pile gets bigger instead. Three moments kept showing up, and each one goes wrong differently.
01
Find a good audience to upsell luxury experiences.
Meaning
What do I even tell it?
You know what you want, but not how to phrase it so the thing understands, and nothing comes back confirming what it heard. So you over-explain, get something wrong, try again.
02
Should I exclude the customers you set aside earlier?
Orientation
Wait, what did I already decide?
You can hold about four things in your head at once. A multi-step conversation makes far more than four, and the agent is the only one keeping track of them.
03
Six audiences look relevant.
Trade-offs
There's too much to weigh.
Six options means fifteen comparisons, plus working out what order to compare them in. Nothing is ranked, and nothing tells you what any of them would cost you.
03What I made
One pattern for each of the three moments.
Sponsored by Adobe, mentored by one of their experience designers. The brief was patterns rather than features, and after a few weeks in AEP I understood why.
An agent doesn't answer the same way twice. Six audiences one time, one the next, and a layout built for six breaks. Adobe also runs a lot of agents, so a fix inside the audience builder would only ever help the audience builder.
A pattern
A rule about what the agent owes the person at a given moment. It works in any agent in any product, because it never assumes what the answer will look like.
A feature
A specific version of the pattern. Someone building the audience builder takes the rule and shapes it around their data, their layout, their constraints.
So the patterns don't replace features. They're guidelines for them: what a feature has to do, with the team who knows the product deciding how.
01
Confirm the request before acting on it
Pattern 01 · For meaning
The agent repeats your request in your own words before doing anything, and puts an edit control on each part it inferred. The usual alternative is a table of parsed fields, which nobody reads.
We want to launch a credit card signup targeting young existing customers. Feb to March, budget around $100K. Channels could be email or paid social.
Here's what I understood: launch a credit-card signup campaign for young existing customers, running February to March, with a budget of around $100K, using email and paid social.
Click one to edit
02
Show the decisions already made
Pattern 02 · For orientation
A normal breadcrumb tells you where you are in a hierarchy. This one tells you what you agreed to on the way here. Hover any step and you get what the options were and which one you took. Only decisions you actually committed to go in, which keeps it short.
Create an audience
You were askedReach a new segmentRe-engage lapsed customersUpsell to existing customers›Select creation method
You were askedReuse an existing audienceFind a similar audienceBuild through propensity model›Model trainingHover a step, then change the answer
Model trained on existing data. The most relevant audiences would be Business travelers and Weekend travelers.
03
Offer one option, not six
Pattern 03 · For trade-offs
Show the best-fitting option on its own rather than all six at once, with your position in the queue and a way out to the full list. Fifteen comparisons become a yes or a no. That exit matters. Without it, showing one option is just hiding the other five.
Best match for your goalOption 1 of 6
Business travelers
4.1M people · propensity 73 / 100
Travels for work and books premium services to make trips easier. The steadiest intent of the six and the closest fit to an upsell.
or view all 6
04Patterns in practiceHover an image
All three built into AEP's audience builder
Seven patterns in total. The three above are one per moment. The other four are the alternates, for when the first answer doesn't suit the situation. Check out the remaining patterns .
05Reflection
Three things I took from these twelve weeks.
I let the failed method be the finding.
I spent the first three weeks ranking the nine decisions by difficulty, looking for the worst one to fix. The ranking came back flat. Nothing was hard, and I had nothing to design. The flat result was the answer, and I only saw that once I stopped trying to make the ranking work.
I started asking what kind of gap I was looking at.
Capability, interface, or volume. A difficulty ranking can only see the first two, which is why it found nothing here. I ask which one a brief is actually describing before I pick a method for it.
I learned what handing the work over actually takes.
I left three things behind: a long document with the research and the reasoning, a short handoff covering only the seven patterns and how to build each one into AEP, and a set of videos recorded across the semester as findings landed. Nobody was going to read a hundred pages to find a rule, and nobody building the feature needed the literature. Adobe is where I learned to plan the delivery for each audience, not just the work.
One task, nine decisions, and not one of them hard on its own. The pile was the problem, not the agent.
Seven patterns came out of it, built into Adobe Experience Platform and 2 minutes faster to get through.
