The Price of Maybe

If you've run servers, you've met the SYN flood. When a visitor says "hello," the server sets aside a little memory and waits for confirmation. Attackers send thousands of hellos and never confirm, crowding out real visitors until timers clear the junk. The fix, SYN cookies, reserves nothing until the visitor proves it will finish the conversation. A system that lets you say "maybe" for free tends to fill up with maybes.

Booking systems keep learning this. In March 2025 the U.S. Embassy in India cancelled about 2,000 visa appointments that bots had grabbed. Recreation.gov holds a campsite in your cart for 15 minutes, then releases it. California State Parks keeps every fee from no-shows and late cancellations and bans repeat no-shows for a year. Each rule admits that "maybe" was never free.

Those rules target scripts that grab too much and people who never show. What worries me is the polite assistant, which has a real person's intent and a script's stamina.

Assistants can already help you book restaurants and hotels. Their public docs don't describe booking three options and quietly cancelling two, but it's a short step, and it would feel like great service: grab the Tuesday dentist slot, keep looking, drop it if Monday opens up. That's a fair thing to want and a real problem for the dentist.

What kept this in check was effort. Booking three things and remembering to cancel two is a chore, so most people don't. The real design decision in any hold is how long it lasts, and booking systems have mostly left that to laziness, then built their forecasts on top.

A review of 105 studies found clinic no-show rates from 4% to 79%, averaging 23%, so some clinics overbook and bet on the empty chairs. Airlines model no-shows flight by flight and are very good at it: between April and June, about 3 in 10,000 passengers took a voucher to give up a seat, and just 0.15 in 10,000 were bumped against their will.

That math assumes people flake for scattered reasons (a sick kid here, a missed bus there) that average out into something forecastable. If a few assistant products all run the same book-now-cancel-later logic, the reasons could stop being scattered. Cancellations could land together, say in the hour a better slot opens across town. Then the clinic is no longer forecasting a crowd. It's forecasting a handful of enormous customers who change their minds at once.

The server fix doesn't carry over. A server can wait to reserve memory until you prove you're serious; a clinic can't, because 9:30 on Tuesday is the thing being held.

Somebody eats the cost. Say clinics respond with deposits and late-cancellation fees (Medicare already allows missed-appointment charges). A user whose assistant cancels an hour before the deadline pays nothing. A patient who can't front the cash never books, and one who loses track pays. The next patient never sees the slot that was held and dropped. Clinic staff absorb the overbooked days, which some studies link to overtime and longer sessions.

Clinics can refill dropped slots fast. At UCSF, a tool offering freed slots to waitlisted patients got those who took one seen a median 14 days sooner, though who took them varied by age and language. The damage comes from cancellations nobody catches in time.

If you run a booking system, check how long a hold lasts and whether anyone chose that on purpose. Then ask what your no-show model does when one person's "maybe" becomes five bookings that vanish in the same ten minutes, and how fast those slots find new takers. If nobody can answer, you're still counting on people being too lazy to cancel.