Agentic AI Travel Booking: The New Duty-of-Care and Liability Boundary
Autonomous AI agents can now search, book, change and cancel corporate travel without a human keystroke. The governance question, who is liable when they get it wrong, has no settled answer yet. Here is what travel managers need to know now
In November 2026, Travelport completed a live demonstration of an AI agent purchasing a flight end-to-end, searching inventory, selecting a fare, processing payment through a tokenised card, and delivering a confirmed itinerary, without a human approving each step. The airline remained merchant of record. The customer approved the itinerary and total price without exposing raw card data. It worked. That milestone matters not because autonomous booking is now routine across managed travel, it is not, but because it signals that the governance gap between AI convenience and employer duty of care is no longer theoretical. Autonomous agents are arriving in corporate travel programs faster than the legal frameworks, contract standards, and internal policies needed to govern them. The organisations that navigate this well will not be those that adopted earliest. They will be those that understood what they were actually authorising.
Three Risk Classes, One Label: What "AI Agent" Actually Means
The corporate travel market now contains three materially different types of AI, routinely described with the same vocabulary. Advisory copilots search, summarise and recommend but cannot change a reservation. Supervised agents prepare a transaction but require affirmative human approval before booking, payment, cancellation or data disclosure. Autonomous agents can complete state-changing actions within delegated limits, booking, cancelling, rebooking, and in some configurations, disclosing traveller data to third-party systems. The risk profile of each is fundamentally different, but vendor announcements frequently blur the distinction. A conversational interface, an "agentic" label, or a connection to live inventory does not by itself prove production authority to transact. Before granting any system access to your travel programme, the first question is not "what can it do?" but "what has it actually been authorised to do, and by whom?" The answer determines whether you are deploying a useful tool or delegating corporate authority to a probabilistic model.
The Liability Gap: No Settled Law, Shifting Contracts
No settled, travel-specific body of law comprehensively allocates responsibility among employer, traveller, TMC, platform, model provider, payment network and supplier when an AI agent makes a bad booking. Existing negligence, consumer, privacy, payment and contractual principles still apply, but they were not written for autonomous agents. In 2026, Citi argued publicly that companies building and operating booking agents should bear booking-error liability, potentially as a condition of registration with payment networks. Current technology contracts, however, more commonly place compliance obligations on the airline, TMC, OTA or other deployer, not the model builder. That contractual deployer-burden model is the more immediately relevant operational reality. California legislation effective January 2026 prevents reliance on AI autonomy as a defence. Colorado's AI Act, effective June 2026, imposes requirements including annual impact assessments for covered high-risk systems. The UK Jurisdiction Taskforce's July 2026 statement framed AI harms through negligence: an organisation owing a duty of care may be liable if it fails to maintain the required standard when using AI. The consistent thread across jurisdictions is that deploying an agent does not transfer your duty of care. The employer remains accountable.
What Incidents Already Tell Us
Public reporting has not yet established a well-documented case in which a corporate autonomous booking agent, operating under formally delegated enterprise authority, caused a major booking, cancellation or data-loss event. That absence must not be converted into a claim of safety. Reported adjacent cases show credible failure modes. In June 2026, Airbnb users reported that an AI support assistant treated ambiguous references to a "refund" as authorisation to cancel reservations without a clear confirmation step. An Expedia virtual agent was reported to have cancelled a non-refundable hotel stay when the customer sought a modification. Priceline's Penny chatbot reportedly told a customer that a non-refundable Thailand booking qualified for a full refund; recovery took six months and media intervention. Security research adds further warning: in 2025, researchers demonstrated that prompt injection against Microsoft Copilot Studio agents could expose sensitive information or manipulate transaction prices. A reported "Red Agent" exercise exploited authorisation weaknesses in an airline booking system to access passenger data and perform write actions. These examples establish technical feasibility of failure. They should not be presented as confirmed corporate losses, but they should not be dismissed as consumer-only problems either. The failure modes are the same; the scale of exposure in a managed travel programme is larger.
The Duty-of-Care Test: Seven Questions Before You Authorise
ISO 31030:2021 provides a general travel-risk framework covering leadership, risk assessment, treatment, communication and review. It is not direct agentic-booking guidance, but it is adaptable: automation must fit within the employer's documented and practised travel-risk programme rather than operate as a parallel channel. Before granting an agent authority to transact, ask seven questions. First, visibility: will the confirmed itinerary reach the corporate system of record and risk-monitoring platform immediately? Second, traveller identity: are name, contact, accessibility and profile data validated against authoritative records? Third, risk context: can the workflow detect a restricted destination, unsafe connection or material destination-risk change? Fourth, reversibility: can a human stop or reverse the action, and who pays when reversal is impossible? Fifth, disruption support: can the traveller reach a competent human during urgent mid-trip disruption? Sixth, evidence: can the organisation reconstruct what the traveller requested, what the model proposed, what policy applied and what was executed? Seventh, exception handling: are accessibility, medical, immigration or security exceptions routed to qualified humans? If any answer is "no" or "we don't know," the agent is not ready for that class of transaction.
Control Architecture: Models Propose, Deterministic Systems Decide
The safest architecture separates the language model from the transaction engine. The model interprets intent and proposes an action. Deterministic software, not the model, validates policy, authorisation, inventory, price and payment before calling supplier systems. This separation is not a technical nicety; it is the difference between a system that can be audited and one that cannot. Every agent should have its own identity rather than sharing a human credential. Permissions should be just-in-time, task-specific and no broader than those of the delegating person. Sub-agents must receive narrower, not expanded, authority. The audit record should be append-only and independent of the agent. At minimum it should capture the human and agent identities, instruction, data accessed, tool calls, policy-rule identifier, model and workflow version, offers considered, approval, timestamp, payment token reference, supplier response, overrides and a tamper-evident decision hash. Monitoring should detect "authority drift," repeated retries, unusual destinations, rapid transaction sequences and changes in model behaviour. High-impact agents need a pre-authorised kill switch. Use tokenised wallets or single-use virtual cards restricted by merchant, amount, currency and time. Apply per-trip and cumulative limits, block split transactions, and prevent duplicate bookings or refunds through idempotency keys.
Allocating Responsibility Across the Chain
When an AI agent makes a bad booking, responsibility is distributed across a chain of parties, and contracts rarely allocate it clearly in advance. The employer or programme deployer decides whether the agent may act, defines policy, approves data access and must maintain duty-of-care continuity. The traveller or approver confirms intent where approval is required, but should not be made responsible for defects the organisation could not reasonably expect the person to detect. The TMC or booking platform controls workflow design, fulfilment, servicing and often the connection to inventory. The agent or model provider controls model behaviour and updates, but commonly seeks contractual limitations on liability. The supplier or merchant of record fulfils the purchased service and ordinarily controls supplier-side charges and refunds. The payment provider authenticates and processes payment but should not be assumed to validate the suitability of an itinerary. Contracts should allocate responsibility for incorrect supplier data, duplicate transactions, booking failures, cancellation errors, model changes, privacy incidents and recovery costs. They should also require audit access, prompt incident notification, service continuity, data-use restrictions and evidence preservation. IATA has not supplied a comprehensive travel-agent liability standard for autonomous passenger booking. Its documented positions instead emphasise responsible AI, verification, human oversight and control of aviation data. A company must confirm data-use rights rather than assume that access to travel content permits model ingestion.
The 90-Day Implementation Checklist
Days 1–30: map and contain. Inventory every AI touching search, booking, servicing, payment, expense or risk data. Classify each as advisory, supervised or autonomous. Identify merchant of record, data controller, system owner and incident owner. Disable shared credentials and unapproved state-changing integrations. Review contracts, insurance and applicable law with counsel. Convert critical travel rules into machine-readable controls. Days 31–60: test and evidence. Run the agent in recommendation-only mode against human baselines. Test cancellations, duplicate requests, refunds, unavailable inventory and disruptions. Conduct prompt-injection, authorisation and data-leakage testing. Implement immutable logging, transaction limits and a kill switch. Verify itinerary ingestion into traveller tracking and emergency-support systems. Train approvers and support teams to recognise automation bias. Days 61–90: controlled release. Permit only low-risk, reversible transactions within narrow limits. Require explicit approval for payment, non-refundable purchases and cancellations. Monitor policy exceptions, error rates, reversals, complaints and unrecovered funds. Review every incident and near miss across travel, security, finance and privacy. Expand authority only after demonstrated compliance and successful recovery testing. Reassess permissions and vendor changes on a fixed schedule.
Agentic AI Governance: Six Non-Negotiables
Agentic booking can improve convenience, policy execution and disruption response, but its defining feature is authority, not conversational fluency. The November 2026 Travelport demonstration proved that end-to-end autonomous booking works. It did not prove that it is safe to deploy without governance. The prudent 2026 operating model is supervised autonomy: models propose, deterministic controls validate, humans approve irreversible or safety-sensitive actions, and independent logs preserve accountability. Until liability law and industry standards mature, employers should assume that deploying an agent does not transfer their duty of care, and that unclear contracts will leave them carrying the operational cost of failure. The organisations that get this right will not be those that moved fastest. They will be those that understood exactly what they were authorising.
Contact TRSS to build an AI governance framework that keeps your duty of care intact as autonomous booking agents enter your travel programme
Contact TRSS