After launch, Booth Selection saved organizers significant time managing exhibitor booth registration. Setup ran across three steps: upload floormap, customize booth tiers, and place pins on the map.
The first two worked well — but for events with hundreds of booths, placing 50–200+ pins one at a time was the most time-consuming step, and the one organizers were most likely to abandon.
"AI could detect and pin the booths on the map, so organizers wouldn't have to."
Yes, but what could the detection realistically handle, and where would it break?
Incorporating AI to detect and pin booths on the map was a decision made from day one. My next step was to find out what it could actually do with real-world maps.
Engineering ran the initial detection tests against real event booth maps pulled from the platform. Once they had results in hand, I set up working sessions with them to walk through the findings together. Alongside that, I explored different prompts in Gemini to see how detection results changed with different framing, and used Claude to synthesize the working sessions and Slack discussions into a shared view of what we were seeing.
In the 65 maps we pulled from actual events, three types emerged based on how much detail they carried:
SOURCE
INSIGHTS
01
Detection quality varied wildly by map type
Color-tier maps worked well, but size-based grouping often needed correction. Numbered-only maps were reliable at high resolution and fragile at low.
02
Unpredictability of organizers' map input
Organizers sometimes include booths they aren't selling, or number non-booth items like stages and info tables. Detection is accurate, but relevance isn't something we can decide for them.
The core design challenge was making the result review feel effortless. My goal was to build enough trust that organizers didn't need to understand the system. And enough clarity that reviewing AI output didn't feel like more work.
I anchored the design around three principles:
01
Progressive disclosure based on user decisions
Asking users to review all detection output at once would be overwhelming; showing too little would make it hard to make decisions. The design should focus each screen on one decision that moves setup forward.
02
Human-in-the-loop
Even when detection is confident, users should have the final say and the ability to adjust results at key steps.
03
Design for the edges, not just the happy path
Real event maps are unpredictable. Major edge cases surfaced in the audit need a designed fallback that let users continue setup, not an error state that stopped them.
With the principles set, the next question was how they'd hold up against real event maps.
Low confidence booth detection due to blurry maps
Colors ≠ tiers
Non-booth items with numbers
Booths already claimed before setup
Size detection isn't always accurate
The edge cases weren't evenly distributed. Each map type gave detection a different level of information, and that changed what organizers needed to do next:
Color/size maps → Detection suggests groupings; organizers bulk-assign.
Numbers-only maps → No grouping; organizers assign one by one.
No detection → Manual setup fallback.
After walking through a few flow options with the PM, we landed on three branches for the three detection outcomes. All of them converge on the same review-and-manage dashboard.
USER FLOWS THAT ADAPTS TO WHAT THE DETECTION RETURNS
With the flow set, I built a matrix mapping every step against every case, putting user story, design requirements, edge cases, and constraints in one place. I brought it to the PM and we worked through the edge cases together, deciding which had to ship at launch and which could wait.
DESIGN SPECS
To introduce organizers to smart detection and prepare them for what was coming, I made two changes to set expectations:
01
Guidance on what details speed up setup
Asking organizers to select what their floormap looks like gave them a chance to learn what works with smart detection.
02
Real-time detection updates
After upload, a progress state showed the system working on the booth detection, so organizers weren't left wondering whether anything was happening.
AI detection is fast, but it isn't always right. Real maps in our sample were sometimes blurry, contained non-booth numbers (like table numbers or room labels), or included booths the AI missed entirely. That created a real design question: how much editing should we let organizers do at the review step?
FLOW EXPLORATION
Reviewing detection results is a decision moment, so the layout had to make that clear. I explored three ways to surface it and tested with the internal team. In the end, we pickedthe modal split view that gave the review its own space without competing with the dashboard for real estate.
LAYOUT EXPLORATION
With the 'review detection results in modal' behavior locked in, I moved on to how organizers should move through it. For maps with booth color/size, the flow needed to walk organizers through detected booth groups. For numbers-only maps, it needed to walk them through individual booth-tier instead. Across both cases, I designed around the same two criteria: scannability and speed to decision.
Maps with booth color/size
Maps with booth numbers only
MAPS WITH BOOTH COLOR/SIZE
Explorations
❌ All booth groups on one screen
Gave organizers a complete overview, but the screen got overwhelming for maps with many groups.
❌ Group-by-group wizard
Focuses each decision to one group at a time, but hard to go back and forth if a group needs editing.
Final design
✅ Collapsible list
Scannable at a glance, focused when expanded, and organizers can jump into any group without walking a sequence.
Maps with booth numbers only
01
Combining steps 2 and 3 into a single dashboard. All paths (smart booth detection or manual setup) land here.
02
Tiers, pins, and descriptions all editable in one view.
Maps with numbers and tier info
01
Combining steps 2 and 3 into a single dashboard. All paths (smart booth detectio or manual setup) land here.
02
Tiers, pins, and descriptions all editable in one view.
With detection and assignment designed case-by-case, the risk was ending up with what felt like four different products under one feature. I audited each path for overlap and consolidated the pieces that could work the same way regardless of how organizers got there.
Before
01
Setup split across three steps: upload, customize tiers, place pins.
02
Editing a pin color or tier description meant navigating back to an earlier step.
After
01
Combining steps 2 and 3 into a single dashboard. All paths (smart booth detectio or manual setup) land here.
02
Tiers, pins, and descriptions all editable in one view.
Trade-offs
The dashboard is denser than any single step in the old flow, but it gives organizers a holistic view of their tiers and booths in one place. Every action lives where organizers are already looking, without jumping between steps to edit different info.
