Travel Chatbots

The Bot's Answer Is Your Answer, Even When It Links to the Right Page

Quick answer: Most guides to travel chatbots begin with what the bot can do. Begin instead with the one thing it must never do, because a tribunal has already priced it. In February 2024 the British Columbia Civil Resolution Tribunal ordered Air Canada to pay a passenger CAD $812.02 after its website chatbot told them they could apply for a bereavement fare within 90 days of buying a full-price ticket. The chatbot's message linked to the airline's actual bereavement page, which said the opposite, and the tribunal held that the link did not help: the airline "does not explain why customers should have to double-check information found in one part of its website on another part of its website." That is the design constraint for every travel bot that talks about money, and most of what customers ask a travel business is about money already paid: change fees, refund eligibility, what happens if the flight moves.

Put that beside a number from our own work. Across the fifteen platforms we have reviewed, the lowest hallucination rates we have measured are 6.7 percent (Tidio, on a single-page index, with the bot refusing rather than inventing) and 9 percent (Chatbase and Botpress, on our five-document protocol), and the anchored spread runs to 25. A bot that is wrong about fare rules seven or nine times in a hundred is a liability with a greeting.

Below: what the decision held, the three answer modes that follow from it, what to do when a storm turns your inbox into an airport, the four messages a trip needs, and what our reviews record about travel, which is very little.

Step 1. Read the decision as a spec, not as a news story

The facts, from the tribunal's own text, which refers to Mr. Moffatt as "they"; this page follows the decision's usage. On 11 November 2022 Jake Moffatt's grandmother died, and "that same day, Mr. Moffat [sic] visited Air Canada's website to find and book a flight from Vancouver to Toronto using Air Canada's bereavement rates" (¶13). The chatbot told them: "If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued by completing our Ticket Refund Application form" (¶15). The words "bereavement fares" in that message were "a highlighted and underlined hyperlink to a separate Air Canada webpage titled 'Bereavement travel'" (¶16), and that page "says, in part, the bereavement policy does not apply to requests for bereavement consideration after travel has been completed" (¶17). They bought two one-way tickets for $794.98 and $845.38 (¶18), applied for the fare on 17 November, "well within the 90 days requested by the chatbot" (¶20), and were refused. In February 2023 an Air Canada representative "admitted the chatbot had provided 'misleading words'" (¶22).

Three findings are the ones to build from.

The bot is the website. Air Canada argued it could not be held liable for information "provided by one of its agents, servants, or representatives – including a chatbot," and the tribunal answered: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot" (¶27).

The link did not save it. "While Air Canada argues Mr. Moffatt could find the correct information on another part of its website, it does not explain why the webpage titled 'Bereavement travel' was inherently more trustworthy than its chatbot. It also does not explain why customers should have to double-check information found in one part of its website on another part of its website" (¶28). And: "There is no reason why Mr. Moffatt should know that one section of Air Canada's webpage is accurate, and another is not" (¶29).

The standard is reasonable care over accuracy. "The applicable standard of care requires a company to take reasonable care to ensure their representations are accurate and not misleading" (¶26), and the tribunal found "Air Canada did not take reasonable care to ensure its chatbot was accurate" (¶28). Damages were $650.88, the difference between the $1,630.36 the tribunal records as paid and the $979.48 the tribunal found should have been paid with a $380 base fare per flight plus the same taxes and charges (¶34, ¶36, ¶40), plus $36.14 interest and $125 in fees, $812.02 in all (¶44).

Read as a specification: a link to the right answer is not a substitute for giving it, and the question asked is whether you took reasonable care. Reasonable care can be documented, and the rest of this guide is about how.

Step 2. Put every money question into one of three answer modes

A travel business is asked all day questions whose wrong answer costs someone money: is this refundable, what is the change fee, do I need a visa, am I owed anything now that the flight has moved. Sort every such question into one of three modes before you configure anything, and write the sorting down, because that document is the start of your record of reasonable care.

Air Canada's bot was in a fourth mode, summarize, in which a language model reads the policy and explains it in its own words. That mode produced "within 90 days" from a page that said the opposite, and it is the mode every generative chatbot defaults to. Our numbers say why it cannot be trusted with money. On a five-document knowledge base, the hallucination rates we measured or anchored run from 6–12 percent (Intercom, anchored range) and 9 percent (Chatbase and Botpress, measured) up to 15–25 percent (AiSensy, anchored); the one measured figure below 9 is Tidio's 6.7 percent, on a single-page support index rather than the five-document set, where the bot's habit was to refuse rather than to invent. The Chatbase review, the best figure on the five-document protocol, also records Capterra reviewers reporting "fake URLs and incorrect product details." Nine percent on "what is the fee to change this ticket" is one wrong answer per eleven customers.

So the rule for a travel bot follows from the standard the tribunal applied: on any question about money already paid or owed, the bot quotes or computes or hands off; it never summarizes. The chatbot confidence policy guide sets out the general guess-ask-or-hand-off design; this is the travel-specific version, with "guess" removed from the list. Our guide to reducing hallucinations covers the scope-first triage that implements it on a platform, and the AI guardrails entry covers where a "never summarize policy" rule can actually be enforced, which is the system prompt at best and a retrieval filter at worst.

Step 3. Know the thresholds you will be asked about

The question a disrupted traveler asks first is "am I owed something," and a bot in compute mode can answer part of it only if it knows the rule. In the United States the text is the Department of Transportation's refund rule, 14 CFR 260.2 as it stands on 31 August 2026, which defines a "significantly delayed or changed flight" as one where, among other triggers, the consumer is scheduled to depart "three hours or more for domestic itineraries and six hours or more for international itineraries earlier than the original scheduled departure time," or to arrive "three or more hours for domestic itineraries or six or more hours for international itineraries after the original scheduled arrival time," or from a different airport, with more connections, in a lower class, or, for a passenger with a disability, via different connections or on aircraft lacking a needed accessibility feature. The same section defines an "automatic refund" as one issued "without waiting to receive an explicit refund request, when the consumer's right to a refund is undisputed."

For a travel agency this cuts two ways. A bot with the booking record and the new schedule can compute whether a change crosses the three- or six-hour line and say so. Whether the customer is in fact owed a refund, from whom, and when, is a hand-off, because it depends on the ticketing carrier, what was sold, and the carrier's tariff, which in the Air Canada case the airline asserted but never produced (¶31: "it is not enough in a legal process to assert that a contract says something without actually providing the contract"). A bot that says "your flight moved four hours, which the US rule treats as a significant change; here is what that means and here is a person" has quoted, computed and handed off in one turn. A bot that says "you are entitled to a full refund" has summarized, and is the Air Canada bot again. Outside the US the thresholds differ (the EU's passenger-rights regulation is built around arrival delay and distance bands), and a business selling across borders should hold each jurisdiction's text as its own dated document rather than let the bot reconcile them.

Step 4. Design the disruption mode before you need it

A travel chatbot has two workloads. On an ordinary day it fields pre-trip questions at a trickle. On the day a storm closes a hub, it receives in an hour more messages than it saw in the previous month, nearly all of them "what happens to me now," from people who are frightened and at an airport. This is the case that justifies the bot and the case most builds never test.

Build a disruption mode as an explicit switch. When it is on: the greeting names the disruption and what you know ("Flights through Denver are affected today; if your booking is, we will message you directly"); the first turn asks for a booking reference and goes straight to compute mode against the record; the bot volunteers the thresholds from step 3 and nothing beyond them; any mention of a missed connection, an overnight, a medical need or a claim is a hand-off at once, with transcript and record attached; and the escalation queue shows its true wait time, because "an agent will be with you shortly" to a hundred people at a gate is a second misrepresentation. Test the mode by replaying a real disruption day's messages and counting how many conversations reached a person with the booking already attached.

The proactive side is where the money is saved. If you know which bookings are affected, message them before they message you, with the new time and one action. On WhatsApp that is a utility template (order management, specific to the user, non-promotional), and it removes the customer from the queue before they join it. Our WhatsApp channel guide covers the messaging-limit ramp a new number hits when it tries to send a thousand of these at once, a constraint to discover in rehearsal, not in a storm.

Step 5. Write the four trip messages, and keep them in the right category

A trip generates four business-initiated messages, and each is a template that has to be written before the trip, categorized correctly, and approved. By our reading of Meta's two-part utility test, all four sit in utility if they stay factual and specific to this booking, and any of them is repriced as marketing if a promotional sentence is added; since April 2025 that reprice is silent, because a utility submission judged to be marketing is approved as marketing rather than rejected. The WhatsApp message template entry has the rules and the approval-time audit; the appointment reminder guide has the copy discipline (one action, one reply key, under a segment).

1. Confirmation, at booking: what was bought, for whom, the total paid, the supplier reference, and the change and cancellation terms as a link to the dated document, not as a summary. This is the message a tribunal would read first. Our confirmation message entry covers the read-back pattern.
2. Pre-trip check, at T-7 and T-2: documents needed, check-in opening time, the meeting point. Quote mode for the documents line: entry requirements are the second most dangerous thing a travel bot can summarize, and they change without notice.
3. Disruption notice, when the record changes: old time, new time, whether the change crosses the step 3 threshold, one action.
4. Post-trip, at T+2: a request for feedback, which Meta's categorization page allows as utility only when tied to the specific transaction, and which becomes marketing the moment it mentions the next trip.

What our fifteen reviews record about travel

Almost nothing, and we would rather say so than pad it. Searching the fifteen reviews for "travel," "airline," "flight" or "itinerary" matches four files. In SendPulse every match is the vendor's customer list, which names LOT Polish Airlines and Radisson Blu Hotels among enterprise references, attributed in the review to the vendor's own pages. AiSensy lists travel among six verticals on its about page. Wati's template library, in the alt text of a screenshot, has a travel category. Botpress's two matches are the phrase "during travel," describing agency owners on the road, and are not about the vertical at all. No review built a travel flow, tested a fare-rule question, or watched a disruption.

What the corpus does have is the number this guide is built on. Thirteen of the fifteen reviews report a hallucination rate from Scenario D of our methodology, most on a five-document knowledge base, and the lowest measured figures are Tidio's 6.7 percent (single-page index) and the 9 percent of Chatbase and Botpress. For the travel-specific question, which platform will hold quote mode on policy text, the reviews that report per-source citation are the place to start: Chatbase (88 percent citation accuracy, measured), Tidio (82 percent citation rate on grounded answers, measured, and a refusal posture that is what quote mode wants) and Intercom (Fin surfaces source articles with each response, anchored range). Whether any of them can be made to print a policy chunk verbatim and refuse to paraphrase it is a question our reviews did not ask, and one to ask in the trial before you load a single fare rule.

FAQ

What did the Air Canada chatbot case decide?

That a company is responsible for what its chatbot says the way it is responsible for any page on its website, that a link to the correct policy inside the wrong answer does not shift the burden to the customer to double-check, and that the standard is reasonable care to ensure the bot's representations are accurate. The tribunal awarded CAD $812.02 in damages, interest and fees. It is a small-claims decision from British Columbia; its reasoning is the clearest statement we have of how a bot's words are read.

Can a travel chatbot answer refund and change-fee questions?

Yes, in two modes only: by quoting the governing text verbatim with its date, or by computing a value from the booking record and a rate table. It should never explain a policy in its own words. Anything involving a claim, a circumstance, or a dispute goes to a person.

Should a travel chatbot use generative AI at all?

For tone, language and understanding the question, yes. For the content of any answer about money, dates, documents or eligibility, the generated part should be the framing around a quoted or computed fact, not the fact. Our measured hallucination rates, 6.7 percent at best and usually above 9, are the reason.

How do I handle a mass disruption with a chatbot?

Switch to a rehearsed disruption mode: a greeting that names the event, a first turn that takes the booking reference, compute-mode answers on the schedule change and the refund thresholds, immediate hand-off on connections, overnights, medical needs and claims, and honest queue times. Message affected bookings proactively with a utility template before they message you.

Which chatbot platform is best for a travel agency?

Our reviews cannot say; none tested a travel flow. Start from the platforms whose reviews report per-source citation (Chatbase, Tidio, Intercom) and confirm in the trial that the bot can print policy text verbatim rather than summarize it. The best chatbot for customer support list is the shortlist to read alongside.

Does a disclaimer that the bot may be wrong protect me?

The decision did not consider one. Air Canada's only contractual defense failed because it never produced the tariff it relied on (¶31), so whether terms or a warning would have mattered is an open question, not a settled one; what the tribunal did apply was a standard of reasonable care over accuracy. We are not lawyers; treat a disclaimer as courtesy rather than defense, and put the effort into quote mode and the twenty-question test.

Related guides

• Hotel chatbot guide — the lodging half of travel: the accessible-reservation rule and what a booking bot must not fail to do.
• Reduce chatbot hallucinations — scope-first triage and the weekly check, the general method behind step 2.
• Chatbot confidence policy — guess, ask, or hand off; this guide removes the first option for money questions.
• AI hallucination — why models make things up, and what reduces it.
• Chatbot QA testing protocol — the checklist the twenty-question policy test belongs to.
• WhatsApp message template — categories, review and the silent reprice that step 5 warns about.
• Chatbot escalation playbook — the queue and the triggers that disruption mode leans on.
• Appointment booking chatbot — availability, the double-booking race and timezones, all of which a tour booking inherits.

Sources

• Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, Tribunal Member Christopher C. Rivers, decision issued 14 February 2024, file SC-2023-005609. Read in full at decisions.civilresolutionbc.ca on 2 September 2026 (the CanLII copy returned an access-denied page to us). Quotations by paragraph: ¶13 (the 11 November 2022 visit), ¶15 (the chatbot's bereavement-fare text), ¶16 (the hyperlink), ¶17 (the policy page's contrary statement), ¶18 (the two fares, $794.98 and $845.38), ¶20 (application "well within the 90 days"), ¶22 (the "misleading words" admission), ¶25 (the elements of negligent misrepresentation, citing Queen v. Cognos Inc., 1993 CanLII 146 (SCC)), ¶26 (duty and standard of care), ¶27 ("a remarkable submission"), ¶28 (the double-check passage and the finding of no reasonable care), ¶29, ¶31 (the tariff not produced), ¶34 ($1,630.36 paid), ¶36 ($380 accepted as the bereavement fare), ¶40 ($650.88 damages), ¶42 ($36.14 interest), ¶43 ($125 fees), ¶44 (the order: $812.02). All amounts are Canadian dollars. Note for anyone checking the sums: the two fares in ¶18 add to $1,640.36, while ¶34 and ¶40 use $1,630.36; the $10 discrepancy is in the decision itself and is reproduced here, not corrected. The decision does not say whether the chatbot was removed; the statement that it was no longer on the site by April 2024 appears in trade-press reporting (CMSWire, 2 April 2024) that we could not confirm against a wire service, and is therefore not made in the body.
• 14 CFR Part 260, §260.2 (Definitions), read at ecfr.gov on 2 September 2026, eCFR currency stamp "up to date as of 8/31/2026," source note 89 FR 32832 (26 April 2024) as amended by 89 FR 65536 (12 August 2024): the "significantly delayed or changed flight" definition, items (1) through (7), and the "automatic refund" definition, quoted in step 3. A further DOT rulemaking, "Airline Refunds and Other Consumer Protections," was published in the Federal Register on 5 December 2025; the eCFR text we read postdates it and still carries the three- and six-hour thresholds, but we have not read that document and flag it for the next refresh. The EU reference in step 3 is descriptive and cites no article; readers selling in the EU should hold Regulation (EC) 261/2004 as a dated document of its own.
• Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), searched 2 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. Travel matches: grep -liE 'travel|airline|flight|itinerary' sample-reviews/*-review.md returns 4 (aisensy, botpress, sendpulse, wati); per-file counts from grep -ciE on the same pattern are 1, 2, 6, 1. Hallucination-rate coverage: grep -liE 'hallucination rate' sample-reviews/*-review.md returns 13 (all but SendPulse and Tars). Figures quoted: Chatbase "9% hallucination rate" and "88% citation accuracy" (measured, English, five-PDF knowledge base) and its Capterra note ("fake URLs and incorrect product details"); Botpress "hallucination rate 9%" (measured); Tidio "6.7% (1/15) hallucinations" and "82% citation rate (9/11) on grounded answers" (measured 28 May 2026 on a single-page support index, tidio-review.md line 205); Intercom "6-12%" (anchored range, Fin); AiSensy "15-25%" (anchored). The remaining rows read from the thirteen files: Voiceflow 10, Botpenguin 13, Typebot 12, Wati 12 and Manychat 12 (all measured); Landbot 12–20 (anchored range); Chatfuel 15 (anchored, pending measurement); and Blip "~10%" as a projection not counted as a measurement. "The spread runs to 25" is the top of the AiSensy range; "6.7 and 9 percent" are the two lowest measured figures. The statement that no review tested a travel flow is an argument from absence across the fifteen files and is stated that way in the body.
• Ahrefs Keywords Explorer, US overview and volume-by-country, queried 2 September 2026 — the demand, difficulty, CPC, global-volume, parent-topic and country-split figures in this page's keyword note.
• Chatbotscape evaluation methodology, including the six-scenario protocol under which the Scenario D hallucination figures were recorded. /methodology (continuously updated).

About this guide

Chatbotscape launched in 2026 as an independent review site for chatbot platforms. This guide is part of our SMB chatbot Academy and is written for the owner or operations lead of a travel agency, tour operator or airline-adjacent service deciding what a bot may say about money. It names no best platform for travel, because our reviews have not tested one, and it reads the tribunal decision it is built on rather than the coverage of it.

Methodology

The tribunal decision and the federal regulation were read in full on the day of drafting and every quotation carries a paragraph or section number so that a reader can check it. The hallucination figures are our own published measurements and anchored assessments, with the evidence class of each stated in Sources. Every corpus search was executed in the form the page states. No travel deployment was run for this guide and no figure on it is a measurement of a travel business.

The editorial judgments on this page, listed rather than flagged line by line:

1. Reading a British Columbia small-claims decision as a design constraint for businesses outside its jurisdiction. The decision binds no one but the parties; its reasoning is what we rely on, and we say so.
2. Treating "summarize" as forbidden for money questions on the strength of hallucination rates measured on a five-document knowledge base that contained no fare rules. The transfer from our test corpus to policy text is a judgment.
3. Advising that a disclaimer be treated as courtesy rather than defense. The decision does not discuss one and its contractual defense failed on evidence (¶31), so this is our reading of the standard applied, not a holding.
4. Placing all four trip messages in the utility category. That is our reading of Meta's two-part test, and the silent reprice described in step 5 is the consequence of being wrong.
5. Publishing the corpus gap (four mentions, zero tests) instead of a shortlist.

See our methodology for how platform facts are verified.

Last updated

3 September 2026.