Every Major TMC Is Quietly Becoming an AI Router

Every major travel management company is quietly becoming an AI router. Not metaphorically. Literally. Amex GBT, BCD Travel, Navan, and Identity Travel built the same capability this year: an orchestration layer that hands a traveler’s request to whichever AI assistant the traveler happens to be using. This is becoming the default architecture across the major TMCs, and the real question is not which AI model wins. It is whether your travel policy can survive contact with it.

What Is Actually Happening Across the TMC Market

In late July, American Express Global Business Travel launched an Egencia AI connector inside Claude, letting a traveler check their calendar, build an itinerary, and complete a booking without leaving the conversation. A Forrester study commissioned by Amex GBT describes this approach as “graduated autonomy”: the AI handles the busywork, the traveler stays in control. It runs on Amex GBT’s own agent to agent architecture, built on Anthropic’s Model Context Protocol.

Amex GBT is not first. BCD Travel launched an MCP framework across Tripsource in May. A month later, France’s CDS Groupe separately opened its hotel booking platform to Copilot and ChatGPT, unrelated to BCD. Navan built its own MCP layer for querying spend and policy data through Claude, ChatGPT, and Cursor. Identity Travel announced in July that its Vibe-powered platform can take a request from Copilot, ChatGPT, or Claude at once. Different TMC, different vendor stack, identical bet.

Google is taking a different path. Rather than plugging into a TMC’s orchestration layer, it launched its own Universal Commerce Protocol at Google I/O 2026, letting travelers book hotels directly inside Gemini and Google Search, Marriott among its first named partners. Today that is a consumer-facing play, not the policy-governed corporate flow this piece is about.

The urgency is not confined to TMCs either. A June 2026 study by Aven Hospitality and h2c, published via Skift, found only 11 percent of hotel organizations have deployed an AI agent that can complete a booking and price inventory in real time. A router only matters if what it routes to can actually transact.

Why “Which AI Model” Was Never Your Real Question

Here is the uncomfortable part. You were never going to choose between GPT, Claude, and Gemini directly. Unless your company is large enough to build its own AI front end, that choice belongs to your TMC, and increasingly your TMC is not choosing just one. Egencia is live in Claude and expanding into Google Chat and Microsoft Teams within months of each other.

The question I keep coming back to is not which assistant wins. It is what your vendor built on top of whichever one it picked, and whether that construction protects your program the way your old single interface used to.

Where Policy Enforcement Actually Lives

Every one of these architectures breaks an assumption most travel policies were built on: that the booking tool enforces policy because the traveler has to go through it. A request now gets interpreted, split into tasks, and routed to specialized agents several steps removed from what a travel manager can see. A travel manager needs real answers:

Where exactly does enforcement happen: the orchestration layer, each specialized agent, or only at final ticketing?
What is the audit trail when an AI agent, not a human, made the actual selection?
Does policy travel with the traveler across front ends, or does enforcement vary by channel?
Who is authorized to deploy an agent on a traveler’s behalf, and does duty of care change when an agent started the request?

The Governance Problem This Is About to Make Worse

None of this is hypothetical. It is live for any organization using Egencia today, and soon for BCD, Navan, and Identity Travel customers too. It’s the same issue at the center of Amazon’s fight with Perplexity: in August, the Ninth Circuit ruled it’s the user, not the AI agent, who legally “accesses” a site on someone else’s behalf, vacating the injunction that had blocked Perplexity’s shopping agent from acting on Amazon.com. That’s a precedent worth watching if it survives further appeal, and it’s directly relevant here: it points toward the traveler, not the AI agent or its vendor, bearing legal responsibility for an agent-initiated booking, which only sharpens the audit-trail and accountability questions travel managers need answered now. I covered that fight in Amazon vs Perplexity: The Lawsuit That Will Change Corporate Travel.

Even regulators are behind schedule. The EU AI Act’s high-risk AI system obligations, its own risk-management and human-oversight requirements, were set to take effect August 2, 2026. Weeks before that deadline, the EU pushed them back to December 2027, sixteen months, on one of the world’s most resourced AI governance efforts. If Brussels cannot finish its own audit-trail rules on schedule, corporate travel policies keeping pace on their own does not hold up either.

I saw this everywhere at GBTA this month. Every vendor’s booth had some version of an AI feature on display, and buyers were genuinely curious, stopping to ask what was actually happening under the hood. But nobody had a clear answer when I pushed on how policy enforcement actually works underneath. My read is that the industry did not fail to keep pace with AI. TMCs are moving faster than most policy frameworks can track. The gap is not adoption. It is governance.

The Bigger Risk

Stop evaluating AI vendors on model quality. Ask your TMC where policy enforcement lives, and get a specific answer, not a reassurance. Request the audit trail format for AI agent initiated bookings before your first traveler uses one. Treat multi front end parity as a requirement: if a vendor supports Claude, Google Chat, and Teams at once, guardrails must work identically across all three. Update your policy documents now, most still assume a human made every decision.

My take is that the organizations that get ahead of this will not be the ones that picked the best model. There will not be a meaningfully best one for long. They will be the ones that rebuilt their guardrails before an agentic booking flow exposed the gap for them.

References

Rajeev Goswami is CEO of WWStay and a member of the GBTA Technology Committee.

Enjoyed this post?
Subscribe on Substack to receive my latest writing directly in your inbox.

Notes by Rajeev Goswami

Insights on AI, Business Travel & Leadership