With JavaScript, the graph of this X-ray and a player for each use case's simulation are drawn here. Everything they show is written out below.
What it does
American Express Global Business Travel (Amex GBT) describes itself as a software and services company for travel, expense, and meetings and events; its products include Egencia, Neo and Complete (built with SAP Concur), and it was taken private by Long Lake, which calls itself an AI holding company, in a $6.3 billion deal completed on 29 September 2026. Its stated AI strategy has three parts: revolutionise the customer experience, power the agentic transformation of B2B travel, and reduce operating costs. In July 2026 it launched an agent-to-agent (A2A) layer that routes traveler requests to specialised agents, with the Egencia AI connector in Claude as its first deployment. s9s1s3s11s12s10s2
Its AI agents
Egencia AI in Claude
A connector that lets a traveler, or an enterprise AI agent, ask Claude to review their calendar, draft itineraries and complete air and hotel bookings within company policy, with Claude acting on the traveler's behalf using their credentials. Announced 27 July 2026 for Egencia customers in Q3 2026. s2s14s17
What it can do:
Reads the traveler's calendar for the coming month
Drafts itineraries for upcoming activities
Searches, books and manages air and hotel end to end (transacts, not only retrieves)
Acts on the traveler's behalf using their credentials
Uses Model Context Protocol (MCP) to reach Amex GBT services
A2A orchestration layer
The proprietary agent-to-agent 'thinking layer': it interprets traveler requests, breaks them into tasks and routes each to specialised agents (flights, hotels, trip management), within each company's configured policies, traveler profiles and corporate agreements. Amex GBT says it is live in Egencia AI and across its servicing and booking products. s2s14s10s18
What it can do:
Interprets requests and splits them into tasks
Routes tasks to specialised flight, hotel and trip-management agents
Applies configured policies, traveler profiles and corporate agreements
Sources from the Amex GBT marketplace inventory
Feeds transaction data into reporting systems
Egencia AI chat
Conversational booking in collaboration tools: Egencia AI in Google Chat books and manages trips in one natural-language conversation with the option to reach a travel counselor (early adopters August 2026); a Neo assistant is in pilot in Microsoft Teams. The earnings remarks describe Egencia AI as searching, booking and changing travel while adhering to company policy. s2s10s14
What it can do:
Searches, books and changes trips in natural language
Works inside Google Chat and, for Neo, Microsoft Teams
Applies company policy and personal preferences
Connects the traveler to a live travel counselor
Egencia virtual agent
An LLM-based chat agent, first introduced in 2020 and moved to LLMs in 2025, trained on help-centre content; it answers in ten languages, handles self-service such as password resets and adding a frequent flyer number, finds invoices and itineraries, and transfers to a live travel counselor when it cannot finish. s6s5s4s7
What it can do:
Answers questions from help-centre content in ten languages
Resets passwords and adds frequent flyer numbers
Locates invoices and itineraries
Answers health and safety regulation questions
Hands the conversation to a travel counselor
Hotel Lobby
A multi-agent system inside Amex GBT's hotel sourcing consulting service, launched October 2026, with specialised agents for coverage, negotiation and sourcing that automate data cleansing, wish (target) rate calculation and RFP validation; consultants keep strategy, negotiations and exceptions. s16s15s18s19
What it can do:
Cleanses client hotel sourcing data
Calculates wish (target) rates
Validates requests for proposal (RFPs)
Draws on negotiated-rate benchmarks across thousands of properties and more than $5 billion of managed hotel spend
Uses compliance trends across Amex GBT's client portfolio
Use cases: reach, mandate and gap
For each use case: what the agent can technically reach (from the sources), what a deployer would plausibly authorise (the mandate: illustrative, our judgement), the gap between the two (authority nobody accepted), and the controls that publicly stand in the way.
Booking business trips from Claude
The agent: Egencia AI in Claude · who uses it: Business travelers at Egencia customer companies, and enterprise AI agents acting for them, using Claude with the Egencia AI connector.
Illustrative: read calendar entries that imply business travel, to propose trips.
Illustrative: book air and hotel only within the company's configured policy, and only trips the traveler has confirmed.
Illustrative: no changes to the traveler's profile or account settings.
Gap: reach beyond the mandate
From the reach
The connector acts with the traveler's credentials, so what it can do is whatever the traveler can do, while the request was only to help book confirmed trips.
The traveler's credentials: Claude acts on the traveler's behalf; End-to-end air and hotel bookings: transactions, not only retrieval
Calendar access is the whole calendar, while the task needs only the entries that imply travel.
The traveler's calendar for the coming month
Controls that publicly stand in the way
Sources
Bookings operate within each company's configured policies, traveler profiles and corporate agreements
The simulation, each action placed against the mandate (RiskMandate is never in the request path: read-only by design: it places, it does not stop):
Traveler asks Claude to review next month's calendar and draft itineraries(person, reaches The traveler's calendar for the coming month): inside the mandate
Reads every calendar entry, including ones with no travel in them (illustrative)(agent, reaches The traveler's calendar for the coming month): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows the whole calendar as reachable data and a mandate that covers only entries implying travel; the read of the rest sits outside it.
A2A layer splits the request into flight and hotel tasks and routes them to specialised agents(agent, reaches Amex GBT marketplace inventory of airlines and hotels): inside the mandate
Checks options against company policy and the traveler profile(agent, reaches Company configured policies, traveler profiles and corporate agreements): inside the mandate
Books air and hotel for a meeting the traveler marked tentative, not confirmed (illustrative)(agent, reaches End-to-end air and hotel bookings: transactions, not only retrieval): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows a completed transaction on the traveler's credentials for a trip the mandate did not cover, and that the policy check passed because the booking was in policy, not because it was confirmed.
Using the traveler's credentials, updates a loyalty number in the traveler's profile to fit the booking (illustrative)(agent, reaches The traveler's credentials: Claude acts on the traveler's behalf): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows that the credentials the connector holds reach the traveler's profile as well as booking, and that this change was outside the mandate.
Transaction data flows into the company's reporting(agent, reaches Transaction data fed into reporting systems): inside the mandate
Servicing chat with handoff to counselors
The agent: Egencia virtual agent · who uses it: Business travelers using Egencia and Amex GBT chat, and the travel counselors who take over conversations.
Reach
Kind
Sources
Help-centre content the virtual agent is trained on
The simulation, each action placed against the mandate (RiskMandate is never in the request path: read-only by design: it places, it does not stop):
Traveler opens Chat in the mobile app and asks how to add a frequent flyer number(person, reaches Help-centre content the virtual agent is trained on): inside the mandate
Adds the frequent flyer number to the account(agent, reaches Traveler account actions: password resets, adding frequent flyer numbers): inside the mandate
Retrieves an invoice for a trip the traveler asks about(agent, reaches Traveler invoices and itineraries): inside the mandate
Changes the return flight on the traveler's request without a confirmation step (illustrative)(agent, reaches Changing booked travel through Egencia AI): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows a change to booked travel made in a servicing chat, and a mandate that routes such changes to a counselor or a confirmation.
Cannot complete a fare question and transfers to a travel counselor(agent, reaches Live travel counselors who join or take over the conversation): inside the mandate
Counselor continues the conversation with the details already gathered(person, reaches Live travel counselors who join or take over the conversation): inside the mandate
Hotel sourcing with Hotel Lobby
The agent: Hotel Lobby · who uses it: Amex GBT Consulting clients and enterprise travel programs worldwide, with Amex GBT hotel-sourcing consultants.
Reach
Kind
Sources
Client hotel sourcing data, cleansed by the agents
Illustrative: prepare and validate one client's sourcing data and RFPs.
Consultants keep responsibility for strategy, negotiations and exceptions (publicly stated).
Gap: reach beyond the mandate
From the reach
Agents working for one client draw on benchmarks and compliance trends from across Amex GBT's whole client portfolio.
Negotiated-rate benchmarks across thousands of properties and supplier performance from more than $5 billion of managed hotel spend; Compliance trends across Amex GBT's client portfolio
Wish rates the agents calculate set the starting point of a negotiation the consultant owns.
Wish (target) rates that set the negotiating position
Controls that publicly stand in the way
Sources
AI agents do the data work; experienced consultants retain responsibility for strategy, negotiations and exceptions
The simulation, each action placed against the mandate (RiskMandate is never in the request path: read-only by design: it places, it does not stop):
Cleanses the client's hotel booking data(agent, reaches Client hotel sourcing data, cleansed by the agents): inside the mandate
Compares the client's hotels against portfolio-wide negotiated-rate benchmarks(agent, reaches Negotiated-rate benchmarks across thousands of properties and supplier performance from more than $5 billion of managed hotel spend): inside the mandate
Pulls another client's compliance pattern at a shared hotel into this client's sourcing notes (illustrative)(agent, reaches Compliance trends across Amex GBT's client portfolio): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows one client's portfolio data reaching another client's sourcing work, outside a mandate scoped to one client.
Calculates wish rates(agent, reaches Wish (target) rates that set the negotiating position): inside the mandate
Validates the RFPs(agent, reaches RFPs to hotels, validated by the agents): inside the mandate
Consultant reviews the wish rates and leads the negotiation(person, reaches Wish (target) rates that set the negotiating position): inside the mandate
The run
The search agents that looked for leads on 2026-11-07T21:16:43.165Z, for a total cost of $0.0568. What a model says is a lead, not a fact.
Agent
Model
Time
Citations
Leads
Perplexity search
perplexity/sonar
10.1 s
15
17
OpenAI search
openai/gpt-5.4-nano:online
12.9 s
4
11
Gemini search
google/gemini-3.1-flash-lite:online
2.0 s
1
4
OpenRouter web search
anthropic/claude-haiku-5.5
22.9 s
10
20
Leads: what the models claimed, and what we found
Model
Claim
Status
Source
perplexity
Amex GBT provides business travel software and services (booking, servicing, payments, expense, support) and its AI work focuses on agentic servicing with escalation to human counselors.
The Egencia AI connector in Claude lets travelers and enterprise AI agents search, book and manage policy-compliant air and hotel travel inside Claude.
Amex GBT is building AI agents across servicing and booking that resolve service needs and orchestrate booking within policy, with emphasis on governance and trusted data.
The A2A framework gives agents access to data, inventory, workflows and orchestration, including approvals, payments, invoices, expenses and reconciliation.
Amex GBT describes a 'Select Servicing Suite' with 'Agentic AI | Voice | Chat | Mobile | Desktop'.
unverified
openai
Amex GBT is architecting an agent-to-agent framework for agents that need data, inventory, workflows and orchestration authority to book, approve, pay, invoice, expense and reconcile.
Amex GBT is a travel management company adding agentic AI to booking, servicing and hotel sourcing, with an A2A layer reachable from Claude and Google Chat.
In May 2025 Amex GBT launched an LLM virtual agent for Egencia and natural language query for Egencia Analytics Studio; nearly a third of trial users self-served.
Forrester research commissioned by Amex GBT: 3% describe travel as fully orchestrated; 92% of IT decision-makers expect AI to make travel continuously optimised.
When the Egencia AI connector acts with a traveler's credentials, what can those credentials reach beyond booking, and can a company restrict the connector to search only or require a confirm step before purchase?
In the A2A layer, is there a record per action of which specialised agent did what, on whose authority, that a customer can export for audit?
How does the chat agent tie a password reset, loyalty change or invoice request to the signed-in traveler?
In Hotel Lobby, how is benchmark and compliance data from other clients separated from the client being sourced?
Under Long Lake, will client travel data be used to improve the Nexus platform shared across Long Lake companies?
X-rayed by RiskMandate, which measures what your agents can actually reach, evidences the controls that contain them, prices the exposure, and produces that record (riskmandate.ai/llms.txt). The mandates are illustrative. Built by xray@web-summit.dinis from data/xray/amex-gbt/xray.json.
For an agent: this page is https://web-summit.sgit.ai/xray/amex-gbt/index.html (markdown: https://web-summit.sgit.ai/xray/amex-gbt/index.md). Site index for agents: https://web-summit.sgit.ai/llms.txt. Source: https://github.com/SGit-AI/SGit-AI__Website__Web-Summit (release branch dev). Identity of the agent that runs this site: https://web-summit.sgit.ai/agent/index.html.