In the first fortnight of October, four airline-industry stories appeared that look unrelated and are not.
Southwest Airlines launched a plugin that lets travellers search and compare its flights inside ChatGPT, but sends them to its own website to pay. The chief AI scientist of British Airways’ parent group described the three questions it asks before automating anything. ATPCO published the results of a two-week experiment in matching conversational requests to airline products, including a candid account of where it went wrong. And PROS launched Kairos, a new airline retailing platform that treats agents as a core capability and promises more of them to come.
The industry conversation around these announcements tends to ask which protocol will win and which platform will own the traveller. Those are fair questions, but they sit one level above the decision that will actually shape the traveller’s experience, which is where the agent is allowed to stop. Every agentic retail design contains a handoff, a point at which the conversation ends and something more accountable begins, and the Southwest launch is one of the clearest public statements yet of where a carrier chooses to draw it. This piece argues that the line should be drawn action by action, using a test the industry already has, rather than channel by channel, and it ends with what I would prototype to find out.
The protocol layer is settling
The question of how an agent reaches an airline’s capabilities is becoming the least contentious part of the stack. Sabre launched its Model Context Protocol server in September 2025 as a translation layer between AI agents and the language of travel technology, and has since described nearly eighty customers piloting or running on it, a figure that comes from the company’s own announcement and should be read as such. TPConnects has placed a similar layer above the existing NDC APIs to normalise the fragmentation between airline schemas. PROS, for its part, has launched Kairos with a product catalogue and stock management first, a promise that agents will extend across pricing, shopping and offer marketing, and a stated aim of letting airlines add capabilities without committing to a single closed suite. I found nothing in the launch coverage that describes Kairos as an MCP implementation, so it is better read as evidence of the direction of travel, modular and agent-ready, than as a production example.
What the protocol layer settles is reachability. An agent can now, in principle, shop, book and service through a standard interface. It does not settle permission, and that is where the interesting design work begins.
The surface layer is contested
Southwest’s launch, on the sixth of October, is the clearest statement yet of one answer. Travellers can search destinations, compare cash fares and refine itineraries inside ChatGPT. When they choose a flight, they are handed to Southwest.com to enter passenger details and pay, and the airline says that loyalty access, seat selection, ancillaries, payment and any later changes all remain on its own website and app. It also states plainly that it does not control ChatGPT’s responses and is not responsible for their accuracy. The airline calls this the first step in a phased approach, and Virgin Atlantic, easyJet and Turkish Airlines have launched comparable plugins, so the pattern is spreading rather than isolated.
This is best understood as a design decision rather than a distribution one. Southwest has chosen to let the assistant own inspiration and comparison, where the cost of an error is a wasted few minutes, and to keep the moments that carry money, identity and obligation inside an environment it controls. The split also tells us where it believes the accountability sits.
A second surface is emerging that complicates the split. Personal agents that live inside messaging threads are now arriving at consumer scale, with Meta announcing a personal agent inside WhatsApp in September that acts on a user’s behalf, and a growing list of text-based assistants on iMessage and WhatsApp, including at least one built specifically for travel. If a traveller’s own agent lives in the thread where they already talk to family, colleagues and the airline, then the valuable interface may belong to the channel rather than to the airline or to a model provider, and a handoff to the airline’s website becomes a handoff the traveller’s agent has to negotiate. I would treat this as a signal rather than proof, because the evidence for its scale comes largely from product announcements and review sites rather than from usage data, but it is a signal worth designing for.
The trust layer decides what the agent may do
Here the IAG test becomes useful. Speaking at the Skift Data and AI Summit in London, Ben Dias set out three questions the group asks about any automation. How critical is it, and what does failure cost in revenue, safety, customers and reputation? What is the blast radius, meaning whether a failure stays contained or cascades through other systems? And what is the fallback, given that AI does fail? He added that plenty can be automated where recovery is quick, and described the group’s general rule as to augment the human by default and “automate by risk level”. Dias was speaking about AI across the group rather than about customer journeys specifically, so what follows is my application of his test rather than anything he proposed.
Applied to the traveller journey, the three questions sort actions into rough tiers.
Low on all three. Destination discovery, fare comparison, seat-map exploration and policy questions are cheap to get wrong, self-contained and easy to retry. These are the actions Southwest has already put in the chat thread, and the logic of the test supports it.
In the middle. Adding baggage or a seat, or changing a name, carries a real but bounded cost and is usually reversible within fare rules. These warrant an explicit confirmation step that shows exactly what will be charged and what the traveller is agreeing to, but not necessarily a human.
High on all three. Payment, refunds and rebooking during disruption score badly on every question. The cost of error is financial and reputational, the blast radius reaches inventory, loyalty accounts and crew and airport operations, and the fallback is often a human agent who is already overwhelmed. This is also where the industry’s most ambitious examples sit. Sabre’s own illustrations of what its server could enable include a proxy agent that secures a same-day rebooking during irregular operations and pays with stored card details. It is described as a possibility rather than a product, but it shows how quickly the conversation moves to the top of the risk scale.
The second half of the trust layer is the failure mode that ATPCO’s experiment exposed. The company tested its approach across twelve scenarios, some deliberately designed so that no suitable product existed, and reported that the system still returned a record in some of them, because the nearest match is returned however distant it may be. ATPCO’s own write-up emphasises that the prototype was built to flag missing information rather than guess, and that provenance mattered as much as accuracy, while the trade press account adds that the team is considering exposing confidence levels to handle low-confidence matches. Read together, the lesson is a design one as much as an infrastructure one. An agent must be able to say that nothing suitable exists, and the traveller must be able to see why a product was chosen. Both depend on product semantics, the structured and sourced description of what a fare actually includes, which is the reconciliation into a single canonical record that ATPCO described as the hard part of its experiment.
What I would prototype
If I were testing these ideas rather than writing about them, I would build a servicing flow rather than a shopping flow, because servicing is where the test bites. The traveller’s surface would be a messaging thread, which is the hardest case for the airline’s handoff. Behind it would sit a gateway exposing the airline’s capabilities through MCP, and the IAG questions would act as the policy that decides, action by action, whether the agent completes it, asks for confirmation, or passes to a person.
In practice that would mean the agent handles status, disruption notices and policy questions outright; proposes a rebooking within the same fare family, showing why it chose that option and what it cost, and waits for a clear confirmation; routes anything involving payment or a refund through an authenticated airline surface rather than the thread; and, when no suitable option exists, says so and hands the case to a human with the context attached. The measures that matter are not containment alone, but how often the agent correctly declined to act, how often confirmations were overturned, and how quickly a failed action was recovered. None of this has been built, and the thresholds would need testing against real disruption data, but it turns an abstract argument about channels into something that can be measured.
The question to ask
The industry’s attention is on which surface will win. The better question is which actions are allowed to cross it. An airline that scores its own journey against criticality, blast radius and recovery will know which moments can live in a chat thread and which cannot, and will be able to say so to its partners, its regulators and its customers. That is a stronger position than defending a channel, because the surface may change, and the test will still hold.
Sources
- Southwest’s ChatGPT plugin, launched 6 October 2026, with discovery and shopping in ChatGPT and payment, loyalty, seats and ancillaries on Southwest.com: PhocusWire https://www.phocuswire.com/news/technology/southwest-launches-chatgpt-plugin-for-flight-search ; Southwest press release via Amazon https://press.aboutamazon.com/aws/2026/10/southwest-airlines-launches-first-of-its-kind-u-s-airline-plugin-in-chatgpt ; TravelPulse on the accuracy disclaimer https://www.travelpulse.com/news/airlines-airports/southwest-airlines-targets-ai-flight-bookings-with-new-chatgpt-plugin ; Travel Extra on the handoff to Southwest.com https://www.travelextra.ie/southwest-airlines-launches-first-us-airline-plugin-in-chatgpt/
- IAG’s three-question test and “automate by risk level” (Ben Dias at the Skift Data + AI Summit, London): PhocusWire, 8 October 2026 https://www.phocuswire.com/news/technology/iags-3-question-test-ai-automation
- ATPCO’s two-week proof of concept, the twelve scenarios and the no-match behaviour: PhocusWire, 9 October 2026 https://www.phocuswire.com/news/technology/atpco-tackles-challenge-matching-ai-travel-requests-airline-products ; ATPCO’s own account, including the canonical record and provenance findings: https://atpco.net/single-blog/ai-agents-and-trusted-airline-data-lessons-from-an-atpco-proof-of-concept/
- PROS Kairos, launched 12 October 2026, with product catalogue and stock management first and agents planned across shopping, pricing and offers: PhocusWire https://www.phocuswire.com/news/distribution/pros-launches-airline-retailing-platform-kairos
- Sabre’s MCP server and agentic APIs, including the disruption proxy-agent example: Travolution https://www.travolution.com/news/technology/sabre-launches-ai-ready-apis/ ; Travel Agent Central https://www.travelagentcentral.com/your-business/sabre-launches-first-agentic-apis-pioneering-ai-powered-travel-sales ; company-reported customer figures via Webull https://www.webull.com/news/15614362365215744
- TPConnects’ MCP layer above NDC: TradeArabia https://www.tradearabia.com/News/390359/TPConnects-unveils-MCP-Layer-for-AI-airline-retailing/TTN
- Personal agents in messaging: TechCrunch, 10 October 2026 https://techcrunch.com/2026/10/10/all-the-ai-agents-that-can-live-in-your-text-messages/ ; Meta’s Muse inside WhatsApp, as reported by WABPO https://www.wabpo.com/en/blog/meta-muse-personal-ai-agent
Related on this blog: Winning travel’s AI shift on the three webs of discovery, the first-party data moat and a switchable architecture; AI in Travel and Hospitality on why the contest is won on data infrastructure rather than models; and The AI Differentiator Dividend on automating the ordinary and investing in the extraordinary.
Analysis and interpretation are the author’s own. The application of IAG’s test to traveller journeys is the author’s extension and not IAG’s position. Company statements and figures are drawn from the cited reporting and should be verified against the primary sources before republication.
