The fun part of an AI booking assistant for travel is the planning. Ask it where to go, have it surface options, picture the trip. That is the demo everyone runs, and it is genuinely useful. It is also the easy half.
The hard half starts after the booking exists, when the customer needs to change a date, cancel a leg, or get a refund for a flight the airline moved. That is where the volume lives, where the rules get strict, and where an assistant either earns its place or creates a mess that lands on a human anyway. This piece is about that second half: what AI can safely automate in changes, cancellations, and refunds, and where it has to stop.
Planning is the easy half
The travel industry's AI conversation has been dominated by the planning use case, and the big platforms have leaned into it. When OpenAI opened ChatGPT to third-party apps in October 2025, Expedia and Booking.com were among the first partners, letting travelers search and compare trips inside the chat. Notably, the actual booking still finalizes on the travel brand's own site. The conversation plans; the transaction lands somewhere accountable.
That split is telling. Planning is forgiving because a wrong suggestion costs nothing, the traveler just ignores it. Booking and post-booking are unforgiving because money and contracts are involved, and a wrong action has real consequences. The interesting work in travel AI is moving from the forgiving half to the unforgiving one.
For customer service teams, the planning use case is mostly someone else's product. The requests that actually hit your queue are about bookings that already exist: change my flight, cancel my hotel, where is my refund. Those are the ones worth automating, because they are high-volume, repetitive, and currently eat agent time that could go elsewhere.
Changes and cancellations: structured, so automatable
A booking change is a well-defined transaction underneath the stress. The customer wants to move a date or cancel a segment. The system needs to find the booking, check the fare rules or cancellation policy, calculate any fee or fare difference, confirm with the customer, and execute the change in the reservation system. Every step has a defined answer.
That structure is what makes changes a good automation target. The assistant works from rules the airline or hotel already wrote, applying them faster and more consistently than a queue of agents reading the same policy documents. Expedia rolled out an AI service agent built to handle booking changes, cancellations, and support in a single interaction, collapsing what used to bounce across search, service, and checkout.
The customer benefit is immediate resolution at the moment of need, often outside business hours, without a hold queue. A change that took a fifteen-minute phone call becomes a two-minute conversation that the customer drives. That speed is most of the value, because the typical change request is simple, just gated behind a wait.
The line to draw is the same one that holds across travel automation. Standard changes within clear policy: automate. Anything that needs an exception, a goodwill gesture, a complex multi-party itinerary, or a judgment call about an unusual situation: route to a human. A system that hands off when it is unsure, the way our Agent 5 model does, keeps the assistant on the changes it can complete correctly and escalates the ones it should not touch.
Refunds: where compliance enters the room
Refunds are where an AI booking assistant gets genuinely sensitive, because money is leaving the company and regulation often governs the timing and the amount.
In the United States, this got sharper in 2024. The Department of Transportation's automatic refund rule, in effect since October 2024, requires airlines to automatically issue cash refunds when a flight is cancelled or significantly changed and the passenger declines the alternative or a travel credit, with refunds for credit-card purchases due within seven business days. That is a hard requirement with a clock on it, and it changes what "good" looks like for refund handling.
An AI assistant can do real work inside that frame. It can tell a passenger whether their situation qualifies, explain the rule in plain language, confirm the choice between rebooking and a refund, and kick off the refund process. Done well, it makes a compliant outcome faster and clearer than a human reading from a script. The assistant becomes the fastest path to the refund the law already requires.
What the assistant should not do is improvise on eligibility or amount. A refund that depends on interpreting an edge case, a contested charge, a fare class with murky rules, or a goodwill exception belongs with a person who can be accountable for the decision. A confidently wrong refund answer can cost more than an unhappy customer; it can become a regulatory or financial problem. Conservative behavior on anything uncertain is the safe default, and it is the difference between an assistant that helps and one that creates liability.
What to automate, and what to route to a human
Post-booking actions, sorted by whether the rule is clear-cut or needs human judgment. Refund-timing basis: U.S. DOT automatic refund rule, effective Oct 28, 2024.
- Standard date change within fare rules
- Cancellation within stated policy
- Fare-difference calculation + confirm
- DOT-qualifying refund: cancelled or significantly changed flight, traveler declines the alternative (credit-card refund due within 7 business days)
- Goodwill gesture or fee waiver
- Complex multi-party / multi-leg itinerary
- Contested charge or disputed fare class
- Edge-case refund eligibility or murky fare-class rules
The trust problem is the real ceiling
The technical capability to automate changes and refunds is mostly here. The limiting factor is trust, and it cuts two ways.
Customers are wary of letting an AI touch their booking, because the downside of an error is high. A mistaken cancellation or a refund sent to the wrong card is the kind of mistake that turns a customer into an ex-customer. The assistant earns trust by being transparent about what it is doing, confirming before it acts, and making the human path obvious. An assistant that always shows its work and asks before executing a change feels safe in a way that an opaque one never does.
Companies are wary too, and they should be. Handing financial actions to automation without guardrails is how you get a story about a bot that refunded the wrong thousand customers. The guardrails are the product: confirmation steps, clear escalation rules, scoped permissions on what the AI can execute, and a conservative posture on anything it is unsure about. Get those right and the automation is safe to trust. Skip them and one bad week erases the savings.
This is why the metrics around booking automation deserve care. A high automation rate on changes and refunds looks great until you find it includes actions customers did not fully intend, or refunds that should have been reviewed. The honest measure pairs automation rate with error rate and customer trust signals, the gotchas that come with the support metrics AI changes. Speed without accuracy is not a win in a domain where the errors are expensive.
Where this is heading
The trajectory is toward assistants that do more of the post-booking work autonomously, monitoring for disruptions, surfacing options, and executing approved changes. The sharpest version of this is already live in the air, where airline chatbots handle disruptions, rebooking, and baggage at the scale a storm produces. That future is real, and the building blocks (conversational interfaces, system integrations, structured policies) are in place. If you want the foundations, our guide to conversational AI covers how these systems actually work.
The pace will be set by trust more than by capability. The companies that move fastest are the ones that automate the structured, low-risk changes first, prove the assistant is accurate and transparent, and only then expand into the sensitive territory of refunds and exceptions. Earning the right to handle a customer's money is a gradual thing, and the assistants that respect that will end up handling far more of it.
The booking assistant that wins will not be the one that books the most trips. It will be the one customers trust to fix a trip when something goes wrong, because that is the moment that decides whether they come back. Booking changes are one stage of a longer arc, set in context by our overview of AI in hospitality across the guest journey. Planning is where the demos shine. Changes, cancellations, and refunds are where the loyalty is actually made or lost, and that is exactly why they are worth automating carefully.