571 AI Designs, 210 Pieces: A Two-Day Software Conference

The brief

At a B2B software company's two-day customer conference, the activation was designed to give every attendee a takeaway they made themselves: original artwork, created with AI and produced on site, instead of a preset giveaway.

It had to work under real conference conditions. It ran from about 8 a.m. each day, kept the line moving so guests didn't walk away, and stayed brand safe without anyone hand-checking every image.

What we built

Three guest stations fed on-site production. Guests picked a product, a canvas tote, an aluminum keychain or a jigsaw puzzle, then designed the artwork themselves with AI.

Guest experience flow:

1. Guest picks a product at the station
2. Types a prompt describing the design they want
3. The AI generates the design in about 15 seconds, shown on screen
4. Guests can regenerate. On average each guest looked at 2.3 designs before choosing one
5. The chosen design goes straight into the production queue
6. The guest gets a text when the piece is ready and collects it

Every piece was sublimated, which bonds the ink into the polyester coating, so the print is permanent and doesn't crack, peel or wash out.

The tech

• Image generation: Google's Gemini image model
• Brand safety: two layers. Our own rules screen prompts for real people, violence and other companies' brands before generation, and the model provider's own safety filters run on top. 12 of 571 generations (2.1%) were blocked, with no staff intervention needed.
• Production: sublimation on polyester canvas totes, aluminum keychains and puzzles
• Data capture: every prompt, design and order was logged, and guests could opt in to text updates

The numbers

247 guests designed something, and 207 of them (84%) ordered it. Most activations count people who walk past. This one counts people who made a design and committed to it.

571 designs, about 15 seconds each. Average generation time was 14.7 seconds (median 14.4, 90th percentile 16.6), well inside the 30-second limit where guests start to drift. Guests regenerated freely, an average of 2.3 designs per guest, because each one came back fast.

210 finished pieces in roughly 16 hours over two days:

Totes outsold the other two products combined. Given a choice, guests picked the item they'd actually use.

66% opted in to text updates. 139 of 210 guests asked to be texted when their piece was ready, so they kept moving around the floor instead of waiting at the booth.

Peak demand hit 31 orders in a single hour on the morning of day two.

What we'd do differently

• Plan production capacity for the peak hour, not the average. The median order was ready in 15 minutes, but the slowest 10% waited 54 minutes or more when the day-two morning rush stacked up. More production capacity during that window would have flattened it.
• Close the loop on the 40 guests who designed but didn't order. 16% of guests generated designs without placing an order. A "text me my design" option would turn that interest into a contact and a piece of content.
• Scan every pickup. 83 pickups were logged against 210 finished pieces. Most likely many guests collected without a scan, which leaves the post-event data less complete than it should be.
• Watch the failure rate. 19 generations (3.3%) failed at the model, and guests had to try again. A second fallback model would make that invisible to guests.

The four commitments: did we hit them?

Credits

• Client: B2B software company (name withheld)
• Production: Sunny Side Ink
• Generative activation platform: LIVEPRINTING.AI
• Stations: 3 guest stations
• Activation duration: about 16 hours over 2 days

Planning a conference activation? See how an AI activation works, check what live event printing costs, or contact us with your dates.