Direct answer: before a UK tour operator, OTA, airline retail desk, hotel group or travel tech product lets an **AI assistant draft, personalise, swap or “explain” booking terms, cancellation notices or price-variation wording**, treat that output as a **consumer contract term or consumer notice** under Part 2 of the Consumer Rights Act 2015. Unfair wording is not binding on the consumer. Transparency is a duty, not a style preference. An AI that can invent a softer cancellation rule, bury a unilateral price change, or summarise T&Cs in a way that shifts the meaning is a control-stack problem before it is a model-quality problem.
That is the operator question Kaize is answering here. We are a UK AI product studio for travel and hospitality. This sits beside — and does not replace — our DMCC drip-pricing / total-price controls, CMA guidance on AI agents under UK consumer law, and ATOL AST 14 booking-assistant disclosure. Those articles cover invitation-to-purchase price display, agent behaviour, and ATOL packaging. This one is narrower: **what must already be true before AI owns the words that bind the booking**.
Primary evidence is first-party UK public sources: the Consumer Rights Act 2015, Part 2; the CMA’s unfair contract terms guidance (CMA37) (updated 22 July 2026); and the CMA’s dynamic pricing project update (20 June 2025), which explicitly flags AI as a reason dynamic pricing may become more complex in travel-like markets.
What Part 2 actually covers in a travel booking stack
Section 61 applies Part 2 to trader–consumer contracts and to **consumer notices** — announcements and other communications that relate to rights or obligations between trader and consumer, whether or not they sit inside a formal T&Cs PDF. Chat replies, quote footers, “price may change” banners, cancellation explainers and AI-generated summaries are in scope when they behave like notices.
Section 62 is the fairness test: a term or notice is unfair if, contrary to good faith, it causes a **significant imbalance** in the parties’ rights and obligations to the detriment of the consumer. Unfair terms and notices are **not binding** on the consumer. Section 68 requires written terms and notices to be **transparent** — plain, intelligible and legible. Section 69 resolves ambiguity in the consumer’s favour.
Section 64 excludes assessment of the main subject matter and of the appropriateness of the price *only when* those terms are transparent and prominent. A buried AI-generated price-variation clause does not get a free pass because “price is core”. Schedule 2’s grey list still matters for terms that allow unilateral variation without a valid reason or without a genuine right for the consumer to dissolve the contract.
- **Must already exist:** a human-owned master wording set for booking, change, cancellation, refund and price-variation terms — versioned, legal-reviewed, and the only source the AI may cite or render.
- **Must already exist:** a gate that blocks the model from inventing new obligations, softening statutory rights, or contradicting the master set.
- **Must already exist:** logging of every AI-rendered notice shown to a consumer, tied to booking ID and wording version.
Where AI breaks CRA Part 2 in travel ops
Travel stacks already generate high volumes of consumer-facing text: quote emails, chat, WhatsApp, post-booking change journeys, disruption messaging. Teams then ask AI to “make it clearer”, “personalise the cancellation policy”, or “explain why the fare moved”. That is exactly when unfairness and opacity creep in.
- **Unilateral price variation after a quote or hold:** if AI is allowed to raise a shown price mid-journey without a transparent, prominent term and a controlled reason code, you are in CRA unfair-terms territory and often also in DMCC invitation-to-purchase territory (see our drip-pricing article).
- **AI-rewritten cancellation or change fees:** a model that “helps” by softening refund language for retention can create a notice that no longer matches the binding term — or that misstates the consumer’s rights.
- **Opaque dynamic-pricing explanations:** the CMA’s June 2025 update says dynamic pricing is not automatically unlawful, but flags harm when consumers are unaware prices can change, feel pressured by sudden rises, or face disadvantage for vulnerable groups. AI that invents a persuasive story for a price jump, instead of rendering approved material information, is a trust and fairness failure.
- **Buried exclusions in long generated answers:** if the only path to a material limitation is inside an AI paragraph the consumer never saw as a prominent term, transparency under sections 64 and 68 is at risk.
Control stack: what AI may prepare vs what it must not own
Kaize’s product stance for travel AI is consistent across this journal: AI can prepare; humans and systems of record own commitment. For CRA Part 2 that means:
- **AI may:** retrieve the correct approved clause by product, channel and jurisdiction; render it verbatim; flag when the consumer asks for a change that would require a different approved variant; draft an internal summary for a human agent.
- **AI must not:** invent or paraphrase binding rights; accept a booking on terms that were never approved; apply a price variation without an approved reason code and consumer-facing notice template; tell a consumer that statutory rights are waived.
- **Humans / systems must own:** the master terms library, legal sign-off, prominence design (what is shown before pay), and the decision to change price or refuse a booking when inventory or risk rules fire.
Wire this into the same agent-control pattern we describe in how travel companies should control AI agents and the booking/CRM system-of-record discipline in connecting booking systems and CRM. Terms that live only in a prompt are not a control stack.
Operator checklist before AI sits on travel terms
- Map every AI surface that can emit booking terms, notices, cancellations, refunds or price explanations — chat, email, voice, quote PDF, agent desktop.
- Bind those surfaces to a versioned terms library; refuse generation when no approved clause matches the intent.
- Separate **price display at invitation to purchase** (DMCC / drip-pricing controls) from **fairness of variation and notice wording** (CRA Part 2). Both gates must pass.
- For dynamic pricing, decide in advance which material information is always shown (that prices can change; what drives change; any published range) and stop AI from improvising beyond that pack — consistent with the CMA’s June 2025 considerations for businesses.
- Run a fairness review on grey-list style variation, termination and limitation clauses with counsel; do not treat “the model said it nicely” as transparency.
- Keep an audit trail: wording version, model output hash, human override, booking ID.
- Train agents that an unfair AI notice does not become fair because a customer clicked “accept” in a hurry.
What this is not
This is not legal advice for a specific contract, not a claim that dynamic pricing is banned, and not a substitute for the CMA’s full unfair-terms guidance or your own counsel. It is an ops boundary article: if AI can change what the consumer thinks they agreed, you do not yet have a safe travel AI product.
If you want help designing the control plane — terms library, agent permissions, booking-system truth and measurement — start at what we build, read more of our thinking, or contact Kaize.