In October 2025, Expedia, Booking.com, and TripAdvisor launched AI-powered integrations inside ChatGPT. Not chatbots bolted onto a homepage. A shared infrastructure layer - the Model Context Protocol, or MCP - that changes how AI agents connect to external data and systems.
At OAG, we process up to 400,000 flight schedule changes a day. The question I kept coming back to when I looked at those launches was simple: when AI agents become the primary interface for travel decisions, where does the authoritative data come from? And who controls that relationship?
Before MCP, connecting an AI agent to an external system meant custom code for every pairing. Your own authentication, your own data formats, your own error handling - and that cost multiplied with every new integration. Engineers call it an N×M problem. MCP reduces it to N+M. Tool providers build one server. AI applications implement one client. Everything interoperates.
Anthropic open-sourced the protocol in November 2024. By late 2025, support had landed across most major AI platforms.
The architecture question worth asking now is: how do you move fast on MCP without locking yourself in if the protocol evolves or gets superseded?
The October 2025 ChatGPT integrations are genuine demonstrations of what's possible. They also stop well short of the full promise.
Look closely, and users are still bounced out to external platforms to complete transactions. The reasons aren't technical - they're structural. Payment processing requires PCI compliance. Booking confirmation needs legal agreements. Liability for errors - wrong dates, misspelt passenger names - creates risk that AI platforms aren't yet equipped to absorb.
What MCP currently delivers is discovery, comparison, and recommendation. The handoff to traditional systems happens the moment actual money or contractual commitment is involved.
That gap is informative. The AI interface captures attention and intent. The authoritative data layer - schedules, availability, pricing, operational status - remains the foundation everything depends on. Build the most elegant AI experience imaginable on stale or incomplete flight data, and the entire interaction fails. The garbage-in-garbage-out problem doesn't go away with AI. It accelerates.
There's a cost buried inside how AI agents work with external data that anyone supplying that data needs to understand.
Large language models don't know your flight options. They call external tools to fetch them. Each tool comes with a definition including what it does, what parameters it accepts and what it returns. As agents connect to more sources, loading all those definitions gets expensive fast. A flight search returns dozens of options with nested data: fare classes, layovers, baggage policies. Multi-step workflows pass large data payloads through the model repeatedly.
Anthropic published research showing naive implementations can consume 150,000 tokens for tasks that better architecture handles in 2,000 - a 98% reduction in cost and latency.
The implication is practical. The most valuable MCP servers won't just expose raw APIs. They'll deliver structured, LLM-ready context - richer than a raw data dump, leaner than loading everything into the model's working memory. An agent planning a complex multi-city itinerary doesn't need the global schedule database. It needs a scoped subset with intelligent defaults and metadata that helps the AI reason about trade-offs.
The data provider who can deliver exactly that becomes easier and cheaper to integrate. Which, in time, becomes a sales argument, not just a technical one.
The travel tech demonstrations from Microsoft, Google, and others sketch where this is heading: specialised AI agents collaborating on a single traveller query. One agent handles intent. Another queries flight data via MCP. A third searches for accommodation. A fourth synthesises the plan.
In that architecture, the Flight Agent doesn't carry encyclopaedic knowledge of global aviation schedules. It has an MCP connection to a provider who does. Which means the traveller using the agent isn't the customer that data providers should be thinking about first. The developer building the Flight Agent is.
That's a different conversation to the one most aviation data businesses are used to having.
If online travel agencies (OTAs) dominate the consumer-facing MCP layer, data providers are in a familiar position: their data powers the experience, the customer relationship stays with whoever controls the interface. The same dynamic that shaped online travel distribution for twenty years could reassert itself.
The counterpoint is that MCP's core value is disintermediation. It reduces integration complexity for everyone - airlines can expose MCP interfaces directly to AI agents, bypassing aggregators. New entrants can compete on data quality rather than distribution muscle.
The protocol doesn't determine who wins. What determines who wins is who has the best data to generate the best insights and intelligence from, in the format AI agents actually need, with the reliability that mission-critical applications require.
The October 2025 launches confirmed that MCP production deployments are real. They're also primarily discovery experiences that hand off to traditional booking flows. The full promise of seamless AI-mediated travel is still being built.
A few things seem clear to me from watching this closely:
Data provenance matters more in an AI-native world, not less. An agent can't interrogate a stale database and flag its doubts to the traveller. It just returns the stale answer.
Structured, LLM-ready data is a differentiator now. The provider who understands how agents consume data - and designs their APIs accordingly - will be easier to integrate. That distinction will show up in developer choices before it shows up in commercial conversations.
The developers building the agents are the channel. That's the relationship worth investing in.
We've spent decades building the data infrastructure that powers aviation intelligence. The next chapter is about positioning that foundation for how AI agents consume and act on it. The protocol is here.
The Builders Blog is a practitioner-first publication on AI, aviation data, and building things that actually work. Subscribe to The Builders Blog LinkedIn newsletter here.