How I replaced a broken search box with conversational AI to get mentees from confusion to a confirmed booking.
ADPList's mentor search was broken at the point of first-time activation. Mentees knew they needed help but couldn't translate a situation into a query that returned a relevant, available, trustworthy result.
I designed an AI-powered conversational search that captures mentee intent through natural language and maps it to mentor recommendations ranked by ADPList's platform signals.
Within the first quarter post-launch, activation rate increased 20% and 67% of users who completed the conversation booked a session.
ADPList - a global mentorship platform with 32,000 mentors across 140 countries.
Healthy Numbers, Broken Funnel
The funnel data looked deceptively fixable – 74% of users completed onboarding, but it took over 48 hours to make the first booking. Our assumption was that search quality was a data and filter problem.
What the research actually showed was a different problem – mentees were spending time shortlisting only to hit timezone mismatches, inactive profiles, and no-shows.
First-time users, especially career-switchers, couldn't formulate a query because they hadn't yet named what they needed. The search box punished them for it.
Mentors appeared in results but weren't actually active. Members booked, got no-shows, and stopped trusting ADPList as a reliable place to find mentorship.
Mentees searched for logos, not fit
The platform was optimising for filter depth while the real problem was mentee bias. Users searched by brand – Google, Apple, Meta – overlooking equally qualified mentors from other companies.
Hypothesis
"The retention ceiling is set by its activation floor. For us, that floor was discovery – and until a mentee could find a mentor who was relevant, available, and worth the risk of a first session, every retention intervention was solving the wrong problem."
Strategy
Search Became a Conversation
I made the call to lead with intent. By adding a conversational AI layer at the entry point, mentees could describe their situation in natural language and receive recommendations matched to their goals, availability, and fit — not just the brands they recognised.
The biggest drop-off risk in a conversational flow is the opening moment. I designed pre-written starters so the cognitive load of the first message was near-zero.
Quick View to Keep Momentum
Opening multiple full profiles to compare candidates was breaking the flow at the point closest to a booking decision. I added a profile quick view so mentees could evaluate a recommendation without leaving the conversation.
Ranking Is a Platform Decision, Not an AI Decision
I made the explicit call that the AI surfaces candidates but doesn't own final ranking. ADPList's signals – acceptance rates, attendance, activity recency – remain the ranking layer. This wasn't just bias prevention; it directly addressed mentors who appeared available but weren't reachable.
Trade-offs
Don't Wait for Certainty to Recommend
Engineering's position was that the AI needed multiple exchanges for all details, before it could recommend. I pushed back – deferring recommendations with multiple back and forth, would cost us completion rate. We must surface candidates after the first prompt and use subsequent turn to refine. 67% conversation-to-booking validated the sequencing.
Track the AI, Own the Ranking
The risk of AI-owned ranking was a black box – no visibility into why a mentor ranked above another, and no way to know if results were better or just different from filters. I kept ranking with ADPList's signals and pushed to track everything: queries, conversation turns, prompt selections, drop-off points. Engineering pushed back on the effort but I pushed back without that data, fast iteration was impossible.
Impact
The numbers below aren't feature adoption metrics. They're signals that the shift from volume to depth actually worked.
Post-Launch Iterations
Availability as a First-Class Signal
The second iteration extended the AI to include availability in search, so mentees could now express timing as part of their intent – "available this week", "open to sessions on Tuesdays" – and get recommendations that matched.
Bringing the Conversation Forward
Post-launch data showed mentees weren't giving the best prompts because the prompt suggestions were a click away. We moved them upfront, making intent easier to express from the first moment without any additional interaction.
Reflections
Fixing activation meant fixing discovery – but discovery wasn't broken because of bad search results. The effort required to find a reliable match was high enough that users discounted the outcome before they even met.
The AI layer made the two-sided nature of that problem visible: sometimes the bottleneck was mentee articulacy, sometimes mentor data quality, sometimes activity signals.
