ChatGPT Said the Visa Would Be Accepted. It Wasn't.

From the Travel Tech Podcast newsletter for episode 33, ChatGPT Said the Visa Would Be Accepted. It Wasn't..

Rob Lee's company shares an office wall in Chicago with the Chinese consulate, and every day he watches hundreds of people file in to pay $150 to $170 for a visa. That single detail tells you more about the visa industry than most trend pieces do. I put this conversation in front of you because Rob has spent almost 20 years solving a problem that AI keeps claiming to have already solved, and he's got the receipts on why it hasn't.

He started running passport applications between a U.S. passport agency and consulates in Chicago, timing same day submissions to beat competitors drowning in FedEx volume. That edge grew into a business serving corporate travelers, cargo airlines, and Fortune 500 companies, and this year it turned into an AI built database that Rob had priced out at roughly $780,000 back in 2018 and finally built for a fraction of that cost. Apple Podcasts, also streaming on Spotify and YouTube.

The tension running through the whole conversation: AI has made it radically cheaper to build the infrastructure behind visa processing, but Rob still puts a human between every AI generated result and the application a customer submits.

That Gap Between a Database and a Document

That tension has a structural cause. Visa requirements don't live in one place. They live across hundreds of individual consulate and government websites, each updated on its own schedule, and the industry's long standing source of truth (IATA's Timatic database) reportedly employed around 15 full time staff just to track daily changes during the pandemic. Rob's team built a scraper and database system to replicate that at a fraction of the headcount, but it still routes every result through a human before it reaches a customer.

The AI pressure shows up on the other side of that gap, at the application itself. Rob says customers regularly tell his team that ChatGPT confirmed a form would be accepted, only to have it rejected, because general models are trained on whatever visa requirement page ranks well rather than whatever page is current. India's eVisa payment system alone can reportedly reject a card up to 15 times before it clears its fraud checks, the kind of granular, frequently changing friction that a general AI model has no reliable way to track. It works well for scraping and structuring data in bulk; it doesn't yet work for the single, consequential answer a traveler needs 48 hours before a flight.

That's the split Rob Lee, CEO and co-founder of Swift Passport and Visa Services, walks through in this episode: starting the company in 2008 serving private aviation clients, the three to four week Brazil visa era, the industry wide shift to eVisas and ETAs, and why his own AI built platform still keeps a human in the loop on every application.

Key Takeaways

• Same day submission was the original moat: Swift beat larger competitors not on price but on processing small batches the same day they arrived, while high volume competitors sat on applications for one to two days.
• eVisas didn't kill the business, they relocated the friction: Paper bottlenecks gave way to payment fraud checks, file format rejections, and photo sizing requirements that still require expert handling.
• AI answers on visa requirements are unreliable by design: Models surface whatever content is well optimised for search, not whatever reflects the current version of a form or policy.
• The $780,000 database became buildable at a fraction of the cost: A project Rob shelved in 2018 for pricing reasons became feasible this year using AI assisted scraping and development, with mandatory human quality control on the output.
• Visa delays are often diplomatic, not technical: Processing times can function as reciprocity leverage between countries, meaning faster technology won't necessarily produce a faster visa.

Alex Brooker, Host of the Travel Tech Podcast, Founder of Airside Labs