Atlas equips AI agents with an open-source skill to make flight reservations

Currently this issue is only partially addressed by players

Atlas publishes as open source a Skill allowing AI agents to conduct an end-to-end flight booking, from search to payment.

AI agents are struggling to get past the crucial booking stage, so they need the tools to go further. Atlas, a specialist in the distribution and marketing of low-cost airline content, is releasing Atlas Flight Booking, an open-source skill for these AI agents to enable them to complete this process. 

With this additional skill, agents will be able to search for flights and fares in real time, select additional services, and issue a booking and process payment once user confirmation is received.

Although this skill is open-source, it requires an account on ATRIP, Atlas's technology platform, which this add-on, installed on a compatible agent, can connect to. 

The skill is installed with a single command on a compatible agent. Currently, this version does not support credit cards and does not cover refunds, cancellations, or any after-sales operations.

With this contribution, Atlas addresses a need already partially met by players such as Amadeus (with its Enterprise API Portal), but emphasises the measures implemented to protect against the risks associated with fully automated booking. 

"We released our flight booking skill for AI agents as open source, not because we wanted agents to book flights completely autonomously, but precisely because that's not what we wanted," said Mary Li, CEO and founder of Atlas.

While there's a need to streamline the transition from search to payment, Atlas is mindful that travelers remain wary of entrusting bookings to a robot, as confirmed by an Amadeus study in early August. 

"Real-world transactions involve four points where a human must say yes authorisation, a price change, an alternative seat, and payment," she added. 

"We've integrated these checkpoints into the process itself, not just the documentation. If you're developing agents capable of processing transactions, the challenge isn't usually the API call; it's knowing when not to make it."