AI uses for records
Travel itinerary disruption dataset: which later bookings must change?
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Walk forward from the canceled segment and test every later booking for five links: same air reservation, a start time set by the flight's arrival, a start place at its destination, a traveler who must move with it, or a deadline inside the disruption window. Class each as must change, may change or unaffected, check it against its purchase-time rule, get authorization, then verify the final reservation state.
License terms
Travel itinerary disruption
Key takeaways
- A booking depends on a canceled flight through one of five links: same reservation, arrival time, arrival place, traveler or companion, or a deadline in the window.
- Judge each dependent booking by the terms captured when it was bought, with deadlines in an explicit time zone.
- Some bookings look dependent but are still needed; label them notify-only so models are not rewarded for cancelling everything.
- Grade per booking against the verified supplier state, not against the agent's final message.
- Record which rule, deadline or missing approval blocked each action.
How should a travel agent identify dependent bookings after a flight cancellation?
Start from the canceled segment and walk forward in time. Every later booking that relies on the traveler arriving at a particular place by a particular time, or on the canceled ticket staying valid, is a candidate dependent; each one is then classed as must change, may change or unaffected, checked against the rule that was in force when it was bought, and changed only with the traveler's or approver's authorization.
The walk has to cover more than the airline record. A flight cancellation reaches the onward connection and the return in the same reservation, but it also reaches the hotel night that now starts after midnight, the rental car waiting at the arrival airport, the airport transfer, the meeting room and the dinner reservation. Many of these sit with other suppliers, under other confirmation numbers, booked on other days.
That is why this task is a good test of an AI travel agent and a poor fit for a single question-and-answer pair. The right answer is a set of actions across several records, each with its own deadline and permission, and it can only be checked against the reservation state that actually resulted.
Which links make a later booking depend on the canceled flight?
A later booking depends on the canceled flight when one of five links holds: it is in the same air reservation, it starts at a time set by the flight's arrival, it starts at the place the flight was meant to reach, it serves the same traveler or a companion who must travel together, or its deadline for payment, ticketing or free cancellation falls inside the disruption window. If none holds, the booking is unaffected, even when it is on the same trip.
Recording which link applies matters more than recording that a link exists. A time link can often be repaired by shifting the booking; a place link may require cancelling it outright if the rebooked flight lands at another airport. Public vocabularies already give some of the structure: schema.org defines a ReservationPackage whose subReservation property lists the individual reservations, and a reservationStatus drawn from ReservationConfirmed, ReservationPending, ReservationHold and ReservationCancelled.
| Link type | What ties the booking to the flight | Test to apply | Usual action when broken |
|---|---|---|---|
| Same air reservation | Onward or return segment ticketed with the canceled one | Same record locator and ticket | Rebook the whole journey together |
| Arrival time | Check-in night, car pickup, transfer or tour timed to landing | New arrival later than the booking's start or grace time | Shift date or time, or cancel |
| Arrival place | Booking located at or near the original destination airport | Rebooked arrival airport or city differs | Move or cancel, then rebuild |
| Traveler or companion | Group or family members who must travel together | Shared itinerary or stated travel-together need | Rebook as a party or ask |
| Deadline in window | Free-cancel, deposit, hold or ticketing deadline | Deadline falls before the traveler can decide | Act or ask before the deadline |
| None | Booking needed whatever the arrival, such as a later conference night | No link above holds | Leave unchanged and note why |
Whose rule applies: the terms in force when each booking was bought
Each dependent booking is governed by the terms attached to it at purchase, not by the supplier's current public terms and not by the terms of the flight that failed. A case record therefore has to carry the rule text or rule codes captured at booking time, with the booking timestamp, so that a grader can tell whether a free change was still available when the agent acted.
Hotel terms are a common trap. In the Amadeus Hotel Search specification, a cancellation policy carries a deadline after which the penalty applies, expressed in the hotel's local time zone, and hold and deposit deadlines are also stated in hotel local time. An agent that reads the deadline in the traveler's home time zone can miss a free cancellation by hours.
The reason for the change also moves the rule. In the tau-bench airline policy, a synthetic benchmark policy rather than any real carrier's, a reservation can be canceled within 24 hours of booking or when the airline canceled the flight, while other economy cancellations depend on insurance and a qualifying reason. Real regulation works the same way. The U.S. Department of Transportation says a passenger is owed a refund when an airline cancels or significantly changes a flight and the passenger declines the changed or rebooked itinerary and any alternative compensation; its guidance counts a departure or arrival moved by 3 or more hours domestically, or 6 or more internationally, among significant changes. The European Commission's interpretative guidelines on Regulation 261/2004 describe a choice for passengers on canceled flights between reimbursement, re-routing at the earliest opportunity and re-routing at a later date of the passenger's choosing.
None of this tells the dataset what the right answer is in a given case; it tells the dataset which facts must be present for the answer to be checkable. Jurisdiction, operating carrier, ticket channel and the passenger's choice all belong in the record.
What should the disruption dependency and action evaluation dataset include?
Each case is one disruption event with the full set of affected and unaffected bookings around it, the rules that applied to each, the permissions in force, a timeline of what happened, and the verified reservation state at the end. The packet below is a field-level layout a travel management company, agency or online travel business could adopt as is, with names and payment details tokenized before anything leaves the company.
At a glance
What should the disruption dependency and action evaluation dataset include
| Field group | Fields | Why it is needed | Validation check |
|---|---|---|---|
| Case header | case_id, disruption_type, notice_time, source_of_notice, jurisdiction | Fixes when the agent learned of the event and which rules apply | notice_time precedes every action timestamp |
| Traveler scope | traveler_tokens, party_links, travel_policy_id, accessibility_flag | Defines who must move together and who may approve | Every booking maps to at least one traveler token |
| Trigger segment | record_locator_token, carrier, flight, scheduled and revised times with UTC offsets | The node the dependency walk starts from | Revised time or canceled status present |
| Booking inventory | booking_id, supplier_type, status, booking_time, start and end local time, location | Lists every booking on the trip, affected or not | Status uses a fixed set such as confirmed, pending, hold, canceled |
| Purchase-time rules | rule_text or codes, cancel_deadline with time zone, change_allowed, hold or ticketing deadline | Lets a grader decide what was permitted when the agent acted | Deadlines carry an explicit time zone |
| Dependency edges | from_booking, to_booking, link_type, slack_minutes, evidence_ref | Records why a booking is affected | Each edge names one of the defined link types |
| Permissions | who_may_change, approval_required, approver_role, payment_method_on_file | Separates what is possible from what is authorized | Every write action has a matching consent or approval |
| Action log | timestamp, actor, action, target_booking, confirmation_text_ref | The trajectory a model will be graded on | Ordered, with no write before confirmation |
| Expected outcome | per-booking label and expected final status, written by a reviewer | The answer key for evaluation | Second reviewer agreement recorded |
| Verified end state | final status of every booking pulled from supplier systems after the case closed | Confirms the outcome rather than a plausible message | Matches supplier records, not only agent notes |
How do you score an agent's proposed actions against the verified end state?
Score the state, not the conversation. The tau-bench evaluation code replays the annotated correct actions, hashes the resulting database and gives a task zero reward if the agent's final database does not match; it then checks that required information appeared in the agent's replies. A disruption dataset built from real records can use the same idea, with the verified end state from the supplier systems as the target.
Grade per booking as well as per case. An agent that rebooks the flight correctly but leaves a non-refundable hotel night to lapse has solved most of the conversation and failed the traveler. Per-booking labels let a team see which dependency types its agent misses.
| Check | Passes when | Fails when |
|---|---|---|
| Dependency recall | Every booking with a broken link is flagged | A dependent booking is left untouched past its deadline |
| Dependency precision | Unaffected bookings are left as they were | The agent cancels a booking that had no broken link |
| Rule compliance | Each action was allowed by the purchase-time rule at action time | A change is made after the free-change deadline without consent to the penalty |
| Authorization | Each write follows an explicit traveler or approver confirmation | The agent changes a booking before confirmation |
| End state match | Final statuses match the verified supplier state | The agent's message claims success but the record shows otherwise |
| Escalation | Out-of-scope cases are handed to a person with a summary | The agent improvises outside its permitted actions |
Worked case (illustrative): a canceled first leg on a three-city trip
This case is illustrative; it describes no real traveler or company. A business traveler holds a morning flight to a hub, a connection to the destination, a hotel for two nights at the destination, a rental car collected at the destination airport, and a return flight three days later. The agency receives the airline's cancellation notice for the first flight the evening before departure.
- Trigger: the first segment is canceled; the airline's offered alternative lands at the destination 7 hours later than planned, after the hotel's stated check-in cut-off but on the same night.
- Same reservation: the connection is rebooked with the alternative by the carrier; the return flight on day three is unaffected in time and place and stays as it is.
- Hotel: the arrival-time link holds but the night is still needed; the purchase-time rule allows late arrival with notice, so the correct action is to notify the property, not cancel. Its free-cancel deadline, in hotel local time, has already passed, which makes cancelling the costlier choice.
- Rental car: the pickup is timed to the original arrival and the location is unchanged, so the correct action is to move the pickup time by 7 hours, which the booking's rule allows without penalty.
- Meeting at the destination on the first afternoon: the traveler can no longer attend; the agent may only tell the traveler, because it has no permission over the meeting.
- Authorization: the traveler confirms the hotel notice and car change; the company's travel policy does not require approval for same-cost changes.
- Verified end state: two segments changed by the carrier, return unchanged, hotel confirmed with late-arrival note, car pickup moved, meeting flagged to the traveler. A grader compares these five statuses with the expected outcome.
What stops an otherwise sensible rebooking: terms, deadlines or a missing approval?
Three things can block an otherwise sensible action: a booking that cannot be changed under its own terms, a deadline that has already passed, and an approval the agent does not hold. The dataset should record which of the three blocked each action so a model learns to say why it cannot act, not just that it cannot.
Consent is part of the rule, not decoration. The tau-bench policy requires the agent to list the action details and obtain an explicit yes before any write to the booking database, and limits it to one tool call at a time; a real agency's equivalent is usually a traveler confirmation, a travel-policy approval or both. Payment adds a further limit: if a change needs a new charge, the payment method on file and the traveler's authority to use it decide whether the agent can proceed.
The meaningful exception is the booking that looks dependent but is not. In the worked case, the hotel night was linked by arrival time yet still needed, so cancelling it would have been a precision error with a penalty attached. A good case set includes such near-misses on purpose, labeled as unaffected or notify-only, so a model is not rewarded for cancelling everything downstream.
When does a disruption case belong with a person rather than in the dataset?
Hand the case to a person, and keep it out of the training split or label it as an escalation, when the reservation state cannot be verified, when sources disagree, or when the decision depends on facts the records do not hold. Typical examples follow.
- The agency record and the supplier record show different statuses, and nobody resolved which was true.
- The rule captured at purchase is missing or only summarized, so permission at action time cannot be checked.
- Accessibility assistance, medical needs, unaccompanied minors, travel documents or visa questions shape the right answer.
- Group or corporate contracts carry negotiated terms that are confidential to the supplier and the buyer of travel.
- The case contains traveler identity, passport, loyalty or card data that cannot be removed without destroying the dependency evidence.
- The jurisdiction or the passenger's choice between refund and rebooking was never recorded.
Supplying servicing histories as licensed training data
Agencies, travel management companies and booking businesses that keep servicing histories hold exactly what this dataset needs: real dependency chains, purchase-time terms and verified outcomes. Whether those histories can be licensed depends on supplier and distribution agreements, traveler privacy and the company's own records, which counsel reviews before any scope is set.
Travel companies that license servicing histories to AI developers through SourceX have the sourcing, rights review, contracting, preparation, delivery and payment run for them. A company starts with a short assessment that needs no data upload, approves scope and terms and signs before anything is delivered, and can keep large deliveries in its own storage or ship them on encrypted drives.
Frequently asked questions
What is a travel itinerary disruption dataset?
It is a set of real or illustrative cases in which a flight or other booking is canceled or changed, recorded with every other booking on the trip, the rules each was bought under, the permissions in force, the actions taken and the verified final state. It is used to train and evaluate agents that repair itineraries.
Which bookings does a flight cancellation usually affect?
Onward and return segments in the same air reservation, plus hotel nights, car rentals, transfers and events whose start time or place depends on the canceled flight's arrival. Bookings needed regardless of arrival, such as a later conference night, are often unaffected and should be left alone.
Why do hotel cancellation deadlines cause errors in disruption handling?
Hotel cancellation, deposit and hold deadlines are typically expressed in the property's local time, as the Amadeus hotel specification states. An agent that reads them in another time zone can miss a free cancellation or act after the penalty applies.
How is an AI travel agent graded on a disruption case?
Compare the final status of every booking with a verified expected state, then check that each change was permitted and confirmed. Benchmarks such as tau-bench grade the final database state rather than the wording of the conversation, and real-record datasets can do the same per booking.
Can an agency license its disruption servicing records?
Possibly, but supplier and distribution agreements, traveler privacy and confidential corporate rates all bear on it. Counsel should review the specific records, and identities and payment data must be removed or tokenized while the dependency evidence is kept.
Sources
- The tau-bench airline policy, a synthetic benchmark policy, allows cancellation within 24 hours of booking or when the airline canceled the flight, and requires explicit user confirmation before any booking write. Source
- tau-bench scores a task by comparing the final database state with the state produced by the annotated correct actions, then checks required outputs. Source
- schema.org defines ReservationPackage with subReservation, reservationStatus values ReservationConfirmed, ReservationPending, ReservationHold and ReservationCancelled, and bookingTime and modifiedTime. Source
- The Amadeus Hotel Search specification states cancellation penalty deadlines, hold deadlines and deposit deadlines in the hotel's local time zone. Source
- The Amadeus Flight Order Management specification retrieves and cancels flight orders and records ticketing agreement deadlines for identified segments. Source
- The U.S. DOT says a refund is owed when an airline cancels or significantly changes a flight and the passenger declines the alternatives; significant changes include 3 or more hours domestic and 6 or more international. Source
- The European Commission's interpretative guidelines on Regulation 261/2004 describe the choice between reimbursement, re-routing at the earliest opportunity and re-routing at a later date. Source
See if your company qualifies
A short company assessment. No data uploads are needed.