The Booking Assistant
Building the Booking.com Virtual Assistant.
Phil Parsons & Maarten Versteegh
Booking.com 101.
Founded in 1996Employ 18 500 204 office in 77
different countries
World leaders connecting travellers with the widest variety of great places to stay
Joined Priceline Group 2005
28,373,691 bookable listings
With 131,269 unique destinations
In over 227 countries and territories
Serving 43 languages
1.55 million Number or room nights reserved each night
851,000 Population of Amsterdam
851,000 + 600,000 = 1.41 million Koningsdag (Kings Day) in Amsterdam
Planning Search and discover
Decide Book
The booking experience.
Holiday confirmed
But what happens after booking?
People have questions and change their plans… a lot!
Typical questions that we get.
Cancellation
“I can’t get the time off of work as
promised, can I cancel my
reservation?”
Pets
“My mother-in-law has invited
herself along for our holiday, is it
possible to get an extra bed?”
Parking
“Can I book a parking space for
the duration of our stay?”
Room related
“I have booked a twin room for
two, can it fit six men?”
How can we automate workload for our customer service and accommodation partners whilst providing a seamless experience for our
guests?
Meet the Booking Assistant, a
virtual customer service agent
in training.
End-to-end guest timeline.
Planning Search and book Pre-stay build up In-stay experience Post-stay
This stage is where we chose to focus to provide clear customer value
What can it do?
Automate answers
Answer questions automatically
using ever-evolving Artificial
‘Narrow’ Intelligence
Offer direct self-service
Empower customers through a
variety of contextually served
self-service options
Offer proactive assistance
Deliver relevant info proactively.
Push help with checking in,
finding things to do
Connect humans
Talk to Booking.com CS or
Partner for help when the
Assistant can’t answer
What type of queries can it deal with automatically?
● Arrival and departure time● Parking queries● Transport questions● Payments information● Add extra beds
● Confirm booking details● Directions to and from properties● Manage room cancellations● Manage date changes
You may have realised we haven’t called it a chatbot?
Human Robot
‘Cyborg assistant’
Why the Assistant needs humans.
True AI is still a long way away
Human-in-the-loop training is
constantly required to train the
assistant
The assistant is hungry for data
Even simple prediction models
need tens of thousands relevant,
clean and annotated data points
The assistant can’t think
It isn’t able to understand context
or accurately define intent
The assistant can be unpredictable
Even the most precise prediction
models make decisions that cannot
be understood fully
What does the guest need?
What can the bot give?
Agile Bot Development
Check-in/Check-outPaymentsDirections/public transportShuttle pickupsCancellationDate changeLuggage dropHotel facilitiesRoom facilitiesCar parkingBreakfast dealsRoom type requestsBed type requestsPetsWi-fiThank youGreetingsSmoking preferenceLanguage detection...
Start from guest intent
Aim for guest satisfaction
Prefer simple systems
DesignLean machine learning
Ride on winners
Find value through experimentationDouble down on areas of impact
Fast Iteration
Pareto Principle80% of effects from 20% of causes
Impact over accuracy
Pragmatism beats smartsFix it in the copy
Customer facing success metrics
Find value fast Abandon line on diminishing returns
Customer message
CHANGE DATESPARKING
PAYMENTS
Modular Architecture
More automated replies
Deliver more answers automatically,
without need for human support,
allowing for scale
More users in the assistant
Grow the audience of guests with
access to the Assistant, to generate
more messages & learn faster on topics
most relevant to guests
More training data
Messages serve as training data. CS agents annotate topics
to improve the models for topic detection and routing,
allowing for better automation
Booking
Assistant
strategy
Listen to Spanish conversation
Ask for meaning of unknown words
Get explanation of meaning
Learn from explanation and expand vocabulary
Active Learning
Monitor incoming messages
Select unknown messages for annotation
Receive annotations
Learn from annotations
Active Learning
Efficiently put humans in the loop
in model improvement
process
The Booking Assistant
● Scalable, personalized help on the go● Human-bot hybrid for best guest experience● Design-driven development● Leverage machine learning and human expertise
for continuous improvement
Thank you!