From Search to Booking in Seconds: How One OTA Built an AI Layer on Top of Amadeus Enterprise

A mid-size B2B online travel agency serving corporate accounts across Europe and the Middle East reduced average booking time from over nine minutes to under ninety seconds by layering an AI-powered natural-language search interface directly on top of their existing Amadeus Enterprise GDS integration.

Explore The Platform Read Full Case Study

An Established OTA Platform With a Search Problem

The client is a B2B online travel agency headquartered in Amsterdam, operating across six regional offices in Europe and the Middle East. Their platform serves corporate travel managers and travel agencies who book frequent business travel: predominantly multi-leg itineraries, tight connection windows, preferred carrier programmes, and negotiated corporate fares that sit outside standard public inventory.

By 2024, they had a mature Amadeus Enterprise setup: a fully certified GDS integration, BSP settlement in place, and a booking desk staffed by 240 travel agents. The platform worked. It just moved slowly. The average time from a travel manager's request to a confirmed PNR was nine to eleven minutes per booking, even for experienced agents who knew the system well.

The bottleneck was not the Amadeus infrastructure; it was the interface sitting in front of it. Agents were navigating traditional GDS command-line inputs, cross-referencing results manually, and building itineraries through a sequence of lookup and confirmation steps that had not meaningfully changed in over a decade. For a team processing 4,000 to 6,000 bookings a month, the cumulative time cost was measurable and growing.

The CTO approached OneClick IT Consultancy with a specific ask: don't replace the Amadeus stack. Build an AI-powered natural language interface on top of it that lets travel agents describe what a corporate traveller needs in plain language and get back a ranked, actionable set of fare options in seconds, ready to book with a single confirmation step.

Problem Statement

Key Pain Points

Structured queries, unstructured requests: Corporate travel requests arrive in natural language but required agents to manually extract parameters and run multiple GDS queries. Converting a single request took two to three minutes before any fare had been seen.
Manual fare ranking and policy checking: Amadeus returned results in default price order. Agents had to manually cross-reference corporate policy rules (fare caps, carrier restrictions, seat class limits) against every result before presenting options.
No contextual memory across a booking session: Every modification to a request required a full new query sequence. On complex itineraries with four or five legs, this multiplied booking time significantly and increased the rate of agent error.
Ancillary steps handled outside the core booking flow: Seat, meal, and baggage requests were a separate post-booking process. This split flow accounted for 20 to 25 percent of total booking time and was the most error-prone part of the process.

Key Metrics

Avg. Booking Time9.2 mins

Policy Compliance Rate~81%

Ancillary Attach Rate~34%

Agent Error Rate on PNRs~4.2%

Monthly Booking Capacity4,000-6,000

Agent Onboarding Time6-8 weeks

The Solution

Natural Language Intent Engine

A fine-tuned LLM-based intent parser converts free-text travel requests directly into structured Amadeus API query parameters. Trained on 14,000 real anonymised booking requests, it extracts origin/destination, dates, cabin, carrier preferences, and connection constraints. It handles ambiguous inputs by asking a single clarifying question rather than failing silently.

Policy-Aware Fare Ranking

Search results from the Amadeus Flight Offers Search API are passed through a ranking engine that applies each corporate account's active travel policy rules before presenting options to agents. The best policy-compliant option appears first. Non-compliant fares are flagged for override rather than hidden. This eliminated one to two minutes of manual policy cross-referencing per booking.

Session Context and Itinerary Memory

A session state layer holds the full context of an active booking across multiple search iterations. Date changes, alternate routings, or additional legs are handled in natural language without restarting the search. The system preserves confirmed legs, retains passenger details, and re-queries only the modified portion. Multi-leg itinerary queries reduced from 12-15 GDS queries to 3-5.

In-Flow Ancillary Attachment

Ancillary services (seat selection, meal preferences, extra baggage) are surfaced within the booking confirmation step rather than as a separate post-booking workflow. Based on the traveller's stored profile, the system pre-populates ancillary selections for agent confirmation before the PNR is finalised. Attach rate moved from 34% to 61%.

Our Approach

Discover

Amadeus API audit
Booking workflow analysis
Agent session recordings review
AI feasibility scoping
Policy data structure review

Deliverable:

Architecture spec, training data requirements, API integration map

Define and Build

LLM intent engine development
Training corpus preparation (14,000 records)
Amadeus API query builder
Sandbox testing of search and price flows

Deliverable:

Intent engine v1, query builder, sandbox validation report

Rank and Remember

Policy ranking engine
Policy ranking engine
Session context layer
Multi-leg itinerary state management
Agent UI beta

Deliverable:

Policy ranking engine, session layer, agent UI beta

Ancillary and QA

Ancillary integration (Seat Maps API)
Traveller profile pre-population
PNR confirmation flow
QA and load testing

Deliverable:

End-to-end booking flow, ancillary attach live

Rollout

Pilot group: 30 agents, Amsterdam
Feedback iteration
Full desk rollout (240 agents)
Agent training programme

Deliverable:

Production go-live, training materials, 30-day post-launch review

Get Your Free Consultation Today!

Unlock the potential of your projects with our expert guidance. Contact us now!

Team Structure

Role	Responsibility
Project Manager	Delivery coordination, stakeholder reporting, sprint planning, risk management, go-live coordination
Amadeus Solutions Architect	Amadeus Enterprise API integration review, endpoint selection, certification compliance, sandbox-to-production transition plan
AI / NLP Engineer	LLM selection, fine-tuning on historical booking corpus, intent parser development, ambiguity handling, prompt engineering
Frontend Developer	React.js agent chat interface, ranked results UI, confirmation and ancillary flow, traveller profile display
Senior Backend Developer	Node.js orchestration layer, policy ranking engine integration, Amadeus API query builder, session state management
Data Engineer	Training data pipeline from historical bookings, anonymisation, corpus structuring, model evaluation framework
QA Engineer	End-to-end booking flow validation, AI accuracy testing, Amadeus sandbox and production testing, load testing, agent UAT coordination

Technology Used

Frontend

React.js

Backend

NodeJS

ExpressJS

Other Services

PostgreSQL
Square API
Twilio API
Google Calendar API

Languages

JavaScript

TypeScript

Framework

Angular Universal
PrimeNG
AngularJS
jQuery

Bootstrap

Angular Material

Frontend

HTML5

CSS3 / SCSS / LESS

Backend Technologies

NodeJS

.Net

API

REST
GraphQL
SOAP

Architecture

MVC
Micro services
Micro front end
Component-Based
Modular
SPA
Hybrid

Database

PostgreSQL

Tools and Build Systems

NPM / Yarn
Gulp / Grunt
Webpack / Browserify
Bower

Version Control Systems

GitHub
GitLab
Bitbucket

Testing Frameworks and Tools

Jasmine
Karma
Protractor
Mocha / Chai
Sinon

Project Management Tools

Jira
ClickUp
Trello
Asana

UI Design Frameworks

Material Design
Bootstrap
Foundation
Semantic UI

Caching and Performance Optimization

Lazy Loading
Code Splitting
Redis
Minimizing Watchers
$digest cycle optimizations
Angular $cacheFactory

Deployment and Automation Tools

Jenkins (CI/CD)
GitLab CI/CD
CircleCI
Docker (for containerization)

Security

Preventing XSS (Cross-Site Scripting)
CSRF Protection
$sce for context based escaping
JWT

Application Performance Monitoring

New Relic
Sentry
Lighthouse

Cloud Servers

AWS
Google cloud platform

Knowledge about Cloud Services

AWS S3
Firebase Hosting
Azure Blob Storage

Experience with Third-Party Services

Google Maps API
Facebook SDK
Stripe
SendGrid

Build and Deployment Tools

Angular CLI
Bower
Vercel / Netlify

UX and Design Tools

Figma
Adobe XD
Sketch

Real-Time App Tools

Socket.IO
SignalR

AI Development Tools

GitHub Copilot
ChatGPT

Why AI Access Changed This Agency's Market Position

Parameter	Without AI Layer	With AI Layer (This OTA)	Business Impact
Search input	GDS command-line queries; agents manually structure every request	Natural language chat input; intent engine builds the query automatically	2-3 minutes saved per booking before a single fare is seen
Booking time	9.2 minutes average; 17+ minutes for multi-leg itineraries	91 seconds average; under 3 minutes for multi-leg itineraries	6x faster per booking across 4,000-6,000 monthly transactions
Policy compliance	Manual cross-reference; ~19% of bookings required rework or re-approval	Automated ranking surfaces policy-compliant option first on every search	Compliance rate from 81% to 96.4%; rework overhead cut significantly
Ancillary revenue	34% attach rate via separate post-booking step; frequently missed	61% attach rate via in-flow confirmation with traveller profile pre-population	+19% ancillary revenue per booking; no additional agent action required
Agent throughput	Hard ceiling of ~38 bookings per agent per day; capacity growth required headcount	31% more bookings processed by same 240-agent team at 6 months	Revenue capacity expanded without a corresponding increase in staff cost
Error rate	4.2% PNR error rate; correction overhead and client complaints growing	0.9% PNR error rate via structured query build and pre-confirmation re-validation	Fewer escalations; lower correction cost; stronger client retention
Agent onboarding	6-8 weeks to full booking competency; high training cost per new hire	~3 weeks to competency; AI handles GDS query complexity agents previously learned manually	Faster team scaling; lower onboarding cost when booking volume grows
Infrastructure risk	Any change to GDS stack risks disrupting BSP settlement and live PNR flow	AI layer deployed as decoupled service tier; Amadeus stack untouched throughout	Full capability upgrade with zero disruption to certified GDS integration

Project Outcomes (Targets)

Booking time reduced from 9.2 minutes to 91 seconds

The average time from agent receiving a travel request to a PNR being confirmed fell from 9.2 minutes to 91 seconds at six months. For multi-leg corporate itineraries, the reduction was more pronounced: from an average of 17 minutes to under three minutes.

+31% agent booking throughput

The same 240-agent team processed 31% more bookings per month at the six-month mark without any headcount increase, translating directly to revenue capacity expansion without a corresponding increase in staff cost.

Policy compliance rate from 81% to 96.4%

Before the AI layer, roughly 19% of bookings required rework or retrospective approval. At six months, the policy ranking engine was surfacing the correct compliant option first on 96.4% of searches.

+19% ancillary revenue per booking

Moving ancillary attachment into the core confirmation step, with traveller profile pre-population, increased the rate at which seat, meal, and baggage services were confirmed at time of booking. The attach rate moved from 34% to 61%.

Agent error rate on PNRs from 4.2% to 0.9%

Structured query building, session state management, and pre-confirmation fare re-validation collectively reduced the rate of PNR errors requiring correction.

4.2x ROI on build cost within six months

The total build and deployment cost was recovered within the first six months through the combination of increased booking revenue, reduced rework overhead, and ancillary revenue growth.

Agent onboarding time reduced from 6-8 weeks to 3 weeks

New agents reaching full booking competency now take approximately three weeks rather than six to eight, because the intent engine handles the GDS query complexity that previously required deep system knowledge.