AI data market
Travel disruption agent data market: 12,680 dialogues, zero rebookings
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
The travel disruption agent data market turns on one scarce record: a case that keeps the rule in force, the alternatives offered, the traveler's choice and the verified ticket or refund outcome. Booking conversations are plentiful; exception cases are not. Google's Schema-Guided Dialogue dataset holds 12,680 travel dialogues and no intent to change, cancel or refund a booking. Score segments on rules, outcomes, rights and preparation before backing supply.
Agent run
Travel disruption agent data
Key takeaways
- The tradable unit is a disruption case linking rule version, alternatives offered, traveler choice and verified outcome, not an itinerary or a flight-status feed.
- In Schema-Guided Dialogue, 94.0% of 10,931 travel booking calls succeed, and none of its 46 intents changes, cancels or refunds a booking.
- The rules that define a correct outcome moved: DOT's 2024 automatic refund rule, EU rail's 100-minute self-rerouting right and a June 2026 EU air agreement with a three-hour rerouting deadline.
- Carrier commitments form a second policy layer that DOT notes airlines can withdraw at any time, so every case needs the commitment version in force.
- On this article's scorecard, US involuntary airline changes score 6 of 8 and hotel relocations 3 of 8; measured buyer utility is unpublished for every segment.
- Traveler data, carrier and client contracts and scattered outcome records limit supply more than disruption volume does.
What exactly is scarce in travel disruption data?
The scarce record is a disruption case that ties four things together: the rule in force when the trip broke, the alternatives the traveler was offered, the choice the traveler made, and the ticket, refund or voucher record showing what actually happened. Schedules, flight status and booking histories are abundant by comparison. They tell a model that a flight was cancelled; they do not tell it what the traveler was owed, what was offered or whether the outcome matched.
That narrows the investor question. The issue is not whether travel generates data, but which disruption segments produce cases a buyer can grade against a rule and an end state, and who holds the right to license them. An AI developer building a rebooking or refund agent needs exactly those cases; everything else is context.
This article tests the thesis four ways: an original count of a public dialogue dataset, the rule changes that decide a correct outcome, a six-segment scorecard with its evidence gaps, and the risks that would break it. How to judge whether a single change was authorized is covered in the live guide to travel rebooking authorization data; this page asks where investable supply sits.
12,680 public travel dialogues, and not one rebooking
Google's Schema-Guided Dialogue dataset, which its README describes as more than 20,000 annotated conversations between a human and a virtual assistant across 20 domains, contains 12,680 dialogues that touch a travel service and no intent for changing, cancelling, rebooking or refunding a booking.
Of its 22,825 dialogues across the train, dev and test splits, 12,680 (55.6%) involve flights, hotels, buses, trains, rental cars, ride sharing or attractions. Those dialogues make 10,931 booking or purchase calls: 10,270 (94.0%) end with a success notice and 661 (6.0%) with a failure notice. Flight reservations account for 1,594 of the calls and 94 of the failures. A failure here is a booking that did not go through, voiced as a one-line apology such as "I'm sorry I couldn't book the flight"; no fare rule, entitlement or alternative is recorded. Of the 571 dialogues with a failed booking call, 464 record a later successful booking in the same conversation.
Among all 46 intents in the dataset's schemas, none matches change, cancel, modify, update, refund, rebook, exchange or reschedule. That is not a flaw: the README lists intent prediction, slot filling and dialogue state tracking among its uses, and it serves them. It does show where the public record is thick, the happy path of a new booking, and where it is silent, the after-sale exception where a disruption agent earns its keep.
Counting rules, as the script applies them: a travel service is any service whose name starts with Flights_, Hotels_, Buses_, Trains_, RentalCars_, RideSharing_ or Travel_; a booking call is a system frame whose service call names an intent marked transactional in that split's schema; its outcome is the NOTIFY_SUCCESS or NOTIFY_FAILURE act in the same frame. Source: the dstc8-schema-guided-dialogue repository on GitHub at commit e852981 (August 2023).
Rulebooks that define a correct rebooking keep changing
In each major jurisdiction, the rule that decides what a disrupted traveler is owed has changed recently or is changing, so a case is gradable only if it carries the rule version in force on its date.
In the United States, DOT's April 2024 final rule requires airlines to refund passengers automatically when a flight is cancelled or significantly changed and the passenger declines the alternatives offered. A significant change includes a domestic departure moved 3 or more hours earlier or an arrival 3 or more hours later (6 hours for international flights), a different origin or destination airport, more connections, or a downgrade in class. Refunds are due within seven business days for card purchases and 20 calendar days for other payment methods.
In the European Union, Regulation 261/2004 gives passengers on a cancelled flight a choice between reimbursement, re-routing at the earliest opportunity and re-routing at a later date (Article 8), and proof of extraordinary circumstances releases the carrier from paying compensation (Article 5(3)). On June 15, 2026, the Council announced an agreement with Parliament on revised air passenger rules: a passenger who chooses the earliest rerouting must be offered an alternative within three hours, which may use other airports, airlines or transport modes at the airline's expense, and may arrange their own if the airline misses that deadline.
EU rail runs on yet another rulebook. Under Regulation (EU) 2021/782, a passenger not offered a usable alternative within 100 minutes of the scheduled departure may book other rail, coach or bus services and reclaim the necessary, appropriate and reasonable cost, and the force majeure clause excuses delay compensation only, not re-routing or assistance. The Commission's 2026 evaluation found the self-rerouting right is not consistently applied, and that the main railway undertakings in around half of member states have given up the force majeure exemption.
For a dataset this means three things. Every case carries its jurisdiction, rule version and event date; cases that straddle a change are tagged, not pooled; and the label is the outcome the rule required, checked against the record, not whatever the human agent happened to do.
Carrier commitments add a second policy layer that can change overnight
Beyond regulation, US airlines publish voluntary commitments for disruptions they cause, and those commitments differ by carrier and can be withdrawn, so a case also needs the commitment in force on its date. DOT's December 2024 advance notice of proposed rulemaking on airline passenger rights reported that six of the ten largest carriers had committed to fee-free rebooking on a partner or other airline for controllable cancellations, and that nine of the ten had committed to hotel accommodation for controllable overnight disruptions. The same notice observed that airlines can remove such commitments from their customer service plans at any time.
This layer is where many agent exceptions live. Whether a meal voucher, a hotel night or a seat on another airline is owed depends on whether the cause was controllable and on which carrier's plan applied that day. An archive that kept only the final itinerary has lost the reason, and with it the label.
For an investor the layer cuts both ways. The variation makes real cases valuable, because a written benchmark policy cannot reproduce one carrier's commitments on one date. It also means every case needs version metadata that operational systems rarely store, which is preparation cost a supplier has to absorb or a buyer has to pay for.
Travel disruption scorecard: six segments and their evidence gaps
Scored 0 to 2 on four criteria, US involuntary airline changes rate highest at 6 of 8 and hotel relocations lowest at 3 of 8, and no segment has published evidence of measured buyer utility. Rule checkability asks whether a published rule or written policy decides the correct outcome. Outcome verifiability asks whether a system of record shows what happened. Rights clarity asks whether the holder can plausibly license the case. Preparation scores high when linking and de-identification are light.
The scores are editorial judgments made for this article from the sources it cites. A 2 means the criterion is documented by default, a 1 that it is partial or varies by firm, and a 0 that it is undocumented or blocked. Treat them as hypotheses to test in diligence, not as measurements; the per-criterion breakdown is in the methods section.
The six segments, with the gap that most limits each:
| Segment | Rule that decides the outcome | Outcome evidence | Main evidence gap | Score of 8 |
|---|---|---|---|---|
| US involuntary airline changes | DOT 2024 refund rule: significant-change thresholds, automatic refund if alternatives are declined | Reissued ticket or refund record with timestamps | Offers made and declined are rarely stored as fields | 6 |
| EU and EEA flight disruptions | Regulation 261/2004, with revised rules agreed in June 2026 | Rerouting record and compensation claim decision | Cases that straddle the rule change | 5 |
| Corporate travel disruptions | Client travel policy and approval chain | Reissue linked to an approval trail | Client contracts may limit reuse of records | 5 |
| US controllable disruptions | Carrier service commitments, which vary and can be withdrawn | Voucher, hotel and partner rebooking records | Which commitment version applied on the date | 4 |
| EU rail disruptions | Regulation 2021/782: 100-minute self-rerouting, force majeure for compensation only | Reimbursement claims, often filed later | Operators apply the exemption differently | 4 |
| Hotel relocations of walked guests | Property policy only; no external rule cited here | Folio and relocation booking | Nothing outside the property to grade against | 3 |
Why a busier network does not mean more licensable cases
Because the binding limits are rights, linkage and rule versions, not volume. A carrier or agency that handles a great many disruptions can still hold few cases that a buyer can grade and the holder can license.
Three filters cut supply. Personal data comes first: a disruption case is built around a named traveler, payment details and sometimes accessibility or family information, all of which must come out, while the de-identified case keeps its timeline and rule links. Contracts come second: agencies work under carrier, distribution and corporate client agreements that may restrict what servicing records can be used for, a question each holder's counsel has to answer. Linkage comes third: the alternatives offered, the traveler's choice and the refund often sit in different systems, and under DOT's rule the refund itself can land days after the conversation ends.
The dialogue count shows what fills the gap when real cases are missing: conversations about making new bookings. That substitutes for some agent skills and not for exception handling, which is why the scarcity argument rests on disruption records specifically and not on travel data as a whole.
The buyer pilot that would prove demand
The first strong evidence would be a rights-screened pilot in which a buyer grades an agent on held-out disruption cases from one segment and reports utility against its current data; none is public for any segment. Until one exists, the opportunity is a hypothesis about where demand should be, not proof that it is there.
A credible design draws cases from one segment and one jurisdiction over a period that includes a rule or commitment change, tags each case with rule version, alternatives offered, choice and verified end state, and holds out the later period. The buyer pre-registers its metric, for example the share of cases in which the agent's proposed outcome matches the outcome the rule required. This article reports no pilot results; the measures below are what one should report:
- Case yield: the share of disrupted bookings that produce a complete case with all four links.
- Label agreement: how often two reviewers assign the same required outcome to the same case.
- Utility: the buyer's pre-registered metric on held-out cases, against its current data.
- Preparation cost: reviewer and engineering hours per accepted case.
- Repeat demand: whether the buyer asks for more cases from the same segment, and for which period.
Four developments that would shrink the opportunity
The thesis fails if exception cases become cheap to simulate, hard to license or captured by the firms that already hold them. These four developments would show it, and each can be watched:
- Synthetic cases catch up. Public dialogue data already covers the booking path; if simulated disruption cases built on written policies matched real ones on a buyer's held-out test, the premium for real cases would fall.
- Holders keep their cases. Carriers and large agencies that run their own disruption tools may train in-house instead of licensing, leaving a thinner set of independent suppliers.
- Rules settle down. Version-tagged archives are worth most while rules churn; a long stable period once the revised EU air rules apply would reduce the edge of dated cases.
- Privacy limits tighten. Disruption cases center on identified travelers; if de-identification that preserves timelines proved unworkable for a segment, that segment's supply would shrink.
Questions that separate a disruption archive from a booking log
Ask any supplier, SourceX included, to show the four links on a sample rather than describe them. A supplier that can only produce final itineraries is selling booking logs.
SourceX is one place to put these questions to work. It takes records a company chooses to license and handles the steps between that choice and a paying AI developer: checking fit, reviewing rights, writing the contract, preparing and delivering the data and settling payment. For buyers, it sources datasets on request instead of holding inventory. None of its supply, deal or commercial figures is published, and company metrics go to investors privately. Its current raise, under Rule 506(c), is open only to accredited investors; materials can be requested through sourcex.si/investors, where any of the questions below can be put to it.
- What share of your disrupted bookings carry the rule or commitment version in force on the event date?
- For a sample of cases, can you show the alternatives offered and the traveler's recorded choice, not only the final itinerary?
- Which system holds the outcome record, and how is it linked to the conversation and the original booking?
- Which carrier, distribution or client agreements govern reuse of servicing records, and has counsel reviewed them?
- How are traveler identities and payment details removed while timelines and rule links survive?
Scoring rules, sources and blind spots
The scorecard and the dialogue count were built on 2026-10-27. The count comes from a short Python script, kept with this draft for reviewers, run on a git clone of google-research-datasets/dstc8-schema-guided-dialogue at commit e852981 (August 2023) with the counting rules stated above and no sampling. Rule facts come from DOT's April 2024 refund rule announcement and December 2024 advance notice, EUR-Lex texts and summaries for Regulations 261/2004 and 2021/782, the Commission's 2026 rail evaluation and the Council's June 15, 2026 press release.
Per-criterion scores in the order rule, outcome, rights, preparation: US involuntary airline changes 2, 2, 1, 1; EU and EEA flight disruptions 2, 1, 1, 1; corporate travel disruptions 1, 2, 1, 1; US controllable disruptions 1, 1, 1, 1; EU rail disruptions 1, 1, 1, 1; hotel relocations 0, 1, 1, 1.
Not measured: case yields, label agreement, preparation cost, prices or buyer budgets for any segment, at SourceX or anywhere else; the share of real disruption records that carry rule versions; and how the revised EU air rules will be applied once in force. No market size, growth figure or company ranking is used or implied.
Frequently asked questions
Is travel disruption data valuable for AI agents?
The valuable part is narrow: cases that link the rule in force, the alternatives offered, the traveler's choice and the verified ticket or refund outcome. Those let a buyer grade a rebooking or refund agent against a required outcome. Flight status and booking histories alone describe what broke, not what the traveler was owed.
Does public data cover flight rebooking for AI training?
Not in the public dialogue set counted here. Google's Schema-Guided Dialogue dataset has 12,680 travel dialogues and 10,931 booking calls, but none of its 46 intents changes, cancels or refunds a booking. Its failures are bookings that did not go through, with no rule or alternative recorded.
Why do rule versions matter in a travel disruption dataset?
Because the correct outcome depends on the rule in force on the event date, and those rules moved: DOT's automatic refund rule in 2024, EU rail rules under Regulation 2021/782, and a June 2026 EU agreement setting a three-hour rerouting deadline. A case labeled under the wrong version teaches the wrong answer.
Which travel disruption segment is the strongest data opportunity?
On this article's editorial scorecard, US involuntary airline changes score highest, 6 of 8, because a published rule decides the outcome and ticket or refund records show what happened. Hotel relocations score lowest, 3 of 8. No segment has published buyer utility results, so every score is a hypothesis to test.
Does SourceX publish travel data supply or deal figures?
No. Supply, deal and commercial figures stay off its public pages and are shared with investors privately. SourceX's raise runs under Rule 506(c) and admits only verified accredited investors; materials are available on request at sourcex.si/investors.
Sources
- The Schema-Guided Dialogue dataset contains more than 20,000 annotated assistant conversations across 20 domains; this article's travel counts come from its train, dev and test files. Source
- DOT's April 2024 final rule requires automatic refunds for cancelled or significantly changed flights when passengers decline alternatives, defines significant change, and sets refund deadlines. Source
- DOT's December 2024 advance notice reported carrier commitments on partner rebooking and hotels for controllable disruptions and noted airlines can remove commitments at any time. Source
- Regulation 261/2004 gives passengers on cancelled flights a choice of reimbursement or re-routing and releases carriers from compensation when they prove extraordinary circumstances. Source
- On June 15, 2026 the Council announced an agreement with Parliament that sets a three-hour deadline for offering rerouting, with self-rerouting if the airline misses it. Source
- Under Regulation (EU) 2021/782, rail passengers may self-reroute after 100 minutes without an offered alternative, and the force majeure clause excuses compensation only. Source
- The Commission's 2026 evaluation found the rail self-rerouting right inconsistently applied and the force majeure exemption given up by main railway undertakings in around half of member states. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.