Amex GBT · 2025 – now
An AI traveler assistant for enterprise travel
Summary: At Amex GBT I lead an AI assistant for business travelers, now in pilot. It's a multi-agent system: an orchestrator routes requests to specialized agents, and conversations hand off to human travel counselors when needed.
What I own
- Product direction for the pilot: which traveler requests the assistant should handle first, and which it shouldn't.
- A Trip MCP server, which I defined and delivered, so the assistant answers trip questions from live itinerary data instead of generic content.
- The evaluation framework: LLM-as-judge scoring of containment, deflection and human handoff, used to decide when a release is ready for the pilot.
The hard part
Business travel is high-stakes and time-sensitive. A wrong answer about a flight is worse than no answer. So the central product question is the same one as in my e-commerce case study: when should the assistant answer, and when should it bring in a person?
How I approach it
- Treat the handoff as a feature, not a failure. A good handoff with context is a success.
- Ground answers in live data through MCP, so the assistant answers from the actual trip.
- Gate every release on evals, not on demos.
The details of this system are confidential. For a deeper look, the e-commerce case study shows the same way of thinking applied to a scenario I can describe fully.