Vrbo AI Penalty Box

Vrbo hosts were canceling at five times the rate of competitors, and Vrbo collected only about 15% of the cancellation fees it was owed. Fees were billed by manual invoice, emailed three to four weeks after the cancellation — leaving roughly $27.5M uncollected every year.

I led the end-to-end design of Vrbo's first AI-automated fee collection system: deter cancellations at the moment of decision, collect fees automatically from host payouts, and build it to scale globally. Concept to launch took about two to three months, with a PM, a data analyst, three engineers, a UX researcher and a content designer.

The low collection rate wasn't a payments failure. It was a behavioral one. Hosts felt no consequence at the moment they decided to cancel, and by the time an invoice arrived weeks later, the cancellation felt like ancient history.

Airbnb and Booking.com make consequences visible at the exact moment of decision. Vrbo needed the same — but with two very different kinds of hosts (individual owners and professional property managers), global payments and legacy systems, “just show the fee” wasn't enough. The consequence had to feel real, transparent and human, without burning the host relationships the marketplace runs on.

Every screen the host sees has an automated decision running underneath it. I designed the host experience and the AI engine as one pipeline, in five steps:

• Trigger — the host starts to cancel, and the engine pulls the booking context
• Deter — an empathy sheet shows who's affected: a family of five, seven months of planning, arriving in two weeks
• Validate — the host picks a reason and messages the guest; the engine checks both against their cancellation history
• Enforce — the fee, a seven-day listing suspension and blocked dates are spelled out in plain language before the host confirms
• Determine — the engine sets the final fee and account standing, and the fee comes out of the next payout automatically

AI tools sped up every phase of my own process, not just the product:

• Claude and ChatGPT for problem framing and synthesis — research on cancellation reasons backed the finding that hosts game the reason picker
• Cursor and Figma Make for exploration — generating layout variants faster than I could sketch them
• Figma MCP into Cursor for prototyping on the live Vrbo design system, so research sessions tested working, high-fidelity flows
• Claude and ChatGPT for content — drafting and stress-testing penalty language with Content Design ahead of Legal review

I partnered with UX Research on seven moderated sessions with active hosts — owners and property managers who had canceled in the past six months. Three findings reshaped the flow:

• Hosts picked whichever reason they thought minimized the penalty, so a required message to the guest made honesty and empathy part of the flow
• The success screen held attention better than the terms screen, so the waiver and next steps moved there
• Auto-blocking the canceled dates read as helpful, not punitive, so it was framed as protecting the host from a second cancellation

The ideal system was fully automated fee collection across every market, currency and payment method at launch. Legacy payments couldn't support that on day one.

So I sequenced by lever. The deterrent didn't need global payments to work, so the disclosure, deterrent and fee flow shipped first in the U.S. and drove most of the drop in cancellations on its own. Automated collection for the global rollout followed and drove the recovery.

In the first three to four months after launch, the system recovered $63M in fees, raised collection to 77% and cut host cancellations by 13.5%.

Design was the connective tissue. The same experience had to satisfy Legal (in-context policy language that reduced “I didn't know” disputes), Payments and Operations (automated fee deduction instead of four-week invoices), Finance (clear payout records), and International Growth (currency formats and policy language that adapt by market).

I'd bring Legal in on day one. We lost a week of creative momentum waiting for policy alignment we could have shaped earlier. Earlier constraints mean earlier freedom.

Next, I'd solve for the other side: the guest who gets canceled on. Rebooking options, visible compensation and a clear timeline, before they have to ask.

Behavior change at scale isn't a UX problem. It's a systems problem where UX is the lever — a better cancellation screen alone would have moved nothing.