From manual curation to self-service: an AI advisor that improved learning adoption and retention
MY ROLE ON THE PROJECT
I was the sole product designer on this project, leading the experience from discovery through launch alongside Product and Engineering. I owned the UX strategy, the recommendation flows, the behavioral analysis after launch and the iteration that followed. My objective was to help learners build their own development paths and reduce how much the whole system leaned on someone manually creating and assigning a program for them.
Learners had plenty of content but not enough guidance, and manual program curation could not keep up
The organization had already invested heavily in structured learning programs and content libraries.
Learners could self enroll in content, and administrators could create and assign curated learning programs, but neither side was scaling well.
Portfolio Experts and administrators spent a lot of time curating, assigning and maintaining programs, and that operational burden kept growing as the catalog did.
At the same time, learners struggled to know where to start, how what they were learning connected to their career growth, and which skills actually mattered for their goals.
Adapting tools we already had to give learners more personal guidance.
The obvious answer was to give the platform real personalization: a recommendation engine that understood each learner and surfaced the right next step for them.
The problem was that our existing learning recommenders ran on old rule based logic with no genuine personalization behind them, so delivering that experience through them would have meant rebuilding the capability almost from scratch.
I did not need to go there, because the pieces already existed. We had a recommendation engine built for the Sales team, and a front end already built for the platform tutor.
Working with engineering, I adapted both into a personal career advisor instead of rebuilding a learning recommender from the ground up, since that gave learners a subject matter expert they could turn to at the moment they needed direction. Reusing logic and interface patterns that were already proven, kept the experience consistent with the other AI touchpoints on the platform, and let us reach learners much sooner. It was the more practical path and, given what we already had, the more sensible one.
BEFORE:
AFTER:
I then placed the advisor where a learner was most likely to be thinking about their next step.
To increase the likelihood of meaningful engagement, I decided to place the Advisor on two key touchpoints:
The Dashboard, the platform's most visited page and the primary entry point for many learners.
Course completion moment, including when learners revisited completed course pages, where they were naturally evaluating their next learning step.
These touchpoints represented the strongest opportunities to connect learners with personalized guidance and content discovery.
Engagement was lower than expected, and the real reason was timing.
Engagement with the feature was lower than expected overall, and my first worry was recommendation quality.
Looking at data we found out that around 80 percent of users who opened/saw the advisor never sent it a message. It was tempting to assume the suggestions were just not good enough, but when I went back through the behavioral analytics and screen replays across both touchpoints, the picture was more specific than that.
The Dashboard placement was driving most of the real engagement, while the course completion placement had a high dismissal rate and very little meaningful interaction.
THE DATA REVEALED THAT
I had misunderstood what learners came to do at course completion.
USERS COMPLETING COURSES WERE TYPICALLY:
Retrieving certificates
Revisiting completed material to refresh learnings
They were not in a moment of planning what to learn next, so the advisor was arriving in the right place at the wrong time.
This mattered, because once someone did engage with the advisor meaningfully, 76 percent of them went on to create a learning plan. That number told me the experience itself was working. What needed to change was the moment I introduced it.
I suggested to collapse the advisor by default at course completion so it stayed available without interrupting learners who came for something else.
I had two options at that touchpoint. I could remove the advisor from the course completion page entirely, or I could keep it there but stop it from imposing itself.
I recommended the second, changing its default state from expanded to collapsed, because removing it would have cut off learners who did want guidance in that moment, while an expanded panel was interrupting everyone who had come to the page for another reason.
A responsible AI decision that shaped the layout.
An advisor that shapes someone's learning path is influencing their development, and I did not want it forcing an interaction on a learner who had simply come to download a certificate.
Leaving the person in control of when to engage with the AI, rather than pushing it in front of them, is the same principle that frameworks like the EU AI Act apply to systems that touch education and skill development, and it felt like the right way to design regardless of the regulation.
The intent was to keep personalized guidance one click away without spending the learner's trust to get their attention.
The change also made the data more "honest". Overall engagement went down, but that was because the numbers were no longer inflated by people who never intended to use the advisor in the first place.
When learners chose to engage, most went on to create a learning plan. This reinforced what the earlier data suggested: the core experience was working, and the bigger issue was when we introduced it.
Nearly 3x higher retention
Retention on personalized AI-generated pathways: 70% VS 24% on direct content library enrollment
The advisor turned learning discovery into a self service experience, lifting adoption and retention while cutting manual curation
This work turned learning discovery from something that depended on manual curation into a self service experience that could actually scale.
63% of learners adopted AI generated learning programs, and pathways generated this way saw 70% retention, compared with 24% for direct enrollment into the content library.
What mattered most to me was that this addressed a business problem and a user problem at once: it reduced the manual burden of building and assigning programs, while helping learners find content that was relevant to them and stay engaged for longer.
Lesson learned: with AI features, timing and context matter more than the model itself
This project changed how I think about what makes an AI feature work. The model itself was rarely the constraint.
What decided whether people used it well was whether I understood the moment they were in, and whether the experience respected that moment instead of treating every visit as an invitation to interact.
That is the question I now ask first with any AI or agentic feature: not only whether the underlying system is capable, but whether it shows up when the person wants to actually use it, and whether they stay in control of what happens next.
