The ubiquity of artificial intelligence is evident in the
fact that the abbreviation “AI” is now a common and recognizable term.
A Google search for “AI” turns up more than 820 million
results, with articles from sources as diverse as The Wall Street Journal, the
Verge and Vanity Fair and topics such as AI for healthcare, astronomy and human
resources. An article in Harvard Business Review calls AI “the most important
general-purpose technology of our era…”
In travel, AI is starting to touch every business sector and
every step of a traveler’s journey, from the ideas and images in online
searches to the pricing of flights and accommodations to experiences
in-destination.
This month we’re digging into artificial intelligence to
learn how it's being used in travel and to understand its future potential.
In part three, we talk to Fred Lalonde, founder and CEO of
Hopper.
Using AI, Hopper predicts future flight prices and recommends if users
should “buy now” or “wait for a better price.” Since its launch in 2015, the
app has sold more than $600 million worth of flights and is currently selling
about $1.5 million every day.
If we could pull back the curtain and look inside Hopper, what would we see?
Fundamentally underpinning Hopper, we are a big data company.
I know the buzzword “big data” is not as in favor as AI is, but fundamentally
AI tends to be built on very, very large data sets.
In terms of the flight database, we
take in about 300 billion prices a month. This would be how much can it cost
on a particular set of flights to go from New York to Rome roundtrip, for
example - that’s what we call a price. We’ll be taking in 10 to 15 million of
these per day. I think we have an archive that must be close to six to seven
trillion historic prices. This is the biggest flight database in the world.
It’s an order of magnitude larger than anything that’s been
accrued. It’s larger than what Google has with its ITA purchase, for example, because
we take in basically every provider on earth, ranging from the global
distribution systems to low-cost carriers. And we’ve been storing this data longer
than anybody else.
So just the raw scale at which we operate the data layer gives
us a significant advantage.
In simple terms, how does your predictor engine work?
We observe the actual demand data. We’re not just spinning
up a matrix of prices, we are looking at - in real time - what is being shopped
in every travel agency, on every OTA, every metasearch site through this data
feed we’ve been building with our partners.
Subscribe to our newsletter below
It is by having algorithms that are capable of looking at
what people are shopping that we are able to do our predictions. Until we
started getting this data stream from the providers – I think it’s five or six
years ago now – nobody had every analyzed the demand aspect of flight pricing.
Everybody had focused on what is the airline pricing at now
or some other aspect of it, but nobody had actually focused on well, regardless
of what the price is, how many people are asking for what. And it turns out at
a very high level that the demand - what people are asking for - precedes the
airline adjusting the rates. In mathematics it’s called a leading indicator.
We spent six years and $12 million building this database.
Nobody had ever invested that kind of energy on this problem. Literally
everybody was throwing away the shopping history in air. We can replay like a
ticker tape every shop of every flight that’s happened in the world, for all
intents and purposes, in the past five years. Nobody had ever stored that
history.
Why is this “ticker tape” of value?
It would be like the archeological record of all flight
shopping. And up until we started streaming it and storing it, nobody had
thought that that dataset was of any interest.
The idea would have been, well flight prices changes all the
time, what is the point of knowing what something was priced at previously,
that price is no longer bookable. Everybody was focused on bookability. Nobody
had realized that if you actually looked at what people are asking for and
stored it and analyzed it, you could predict the future.
The GDSs themselves were not storing this data. We were the
first to do it, and I think today we might be the only people storing it still.
Along with price predictions, how else are you using artificial
intelligence?
The second aspect of how we use AI is the one I’m most
excited about. I think it’s going to be much more significant than the forecasting: It’s how we apply our user data to the pricing.
The platform has been used to plan about 60 million trips. This
isn’t data that exists outside of Hopper. It’s the proprietary usage that
people are doing with Hopper.
What’s fascinating about that is today about 25% of every ticket that we sell on Hopper was a ticket that the person did not ask for. It’s the AI that sold it.
Fred Lalonde - Hopper
What is very different about how Hopper works is people
don’t search the app. It’s not like a Kayak or Travelocity where you go to it
and you search whenever you want a fare. We send push notifications. So the
flow is you tell us where you want to go - it can be an entire country, you
don’t even need to enter dates anymore - you just basically express what you need
and Hopper starts notifying you.
We’ve sent over two billion push notifications in the past two
years.
What’s interesting is that when somebody is doing this on
their phone instead of a website, they come back to the app on average every four
to five days. At Expedia when I was there, if we saw you twice a year we were
happy.
That means over the course of the trip planning – and on Hopper
people plan trips almost 120 days before departure - so for 120 days, every four
to five days, the users come in and they click on things and they open
notifications. If you imagine the richness of that data set that we now have.
We have a better understanding of what people want to do in
terms of traveling than Google or Facebook, because neither of those companies
have that frequency of interaction with customers around air travel and hotel.
With that data set we’ve actually been using AI for over a
year now to predict what people will buy even though they have not asked for
it.
Can you explain how that works?
Let’s say you are watching a trip from New York to Rome. It
is very possible that 80,000 other people are doing that on Hopper right now.
It’s possible that of the 80,000, let’s say 40,000 are also watching New York
to Milan simultaneously. And it’s possible you don’t know flights to Milan are
much cheaper, which they happen to be most of the time.
If you are watching for a trip from New York to Rome on
the first week of June, it might start talking to you about Milan a week later.
It will do this because the AI is judging that the cost savings are
sufficiently large that it predicts you are willing to change destinations.
What’s fascinating about that is today about 25% of every
ticket that we sell on Hopper was a ticket that the person did not ask for.
It’s the AI that sold it.
It’s looking at all of the prices as they come in. Let’s say
for your trip, 100 million fares might be relevant over the course of a day, which
would be a good estimate. And it’s predicting the probability that you would
buy every one of those 100 million fares, even though they don’t match exactly
what you asked for. In a way it’s acting like a really smart travel agent,
because it’s doing more than just fulfilling your initial request. It’s
thinking for you, but it’s doing it at a scale that no human being could ever
do.
What we found – and this is a fascinating part - when we
send a push notification that comes from the AI, meaning it is talking to you
about something you did not ask for - the conversion is almost three times
higher than when we talk to you about the thing you asked for.
Where do you see this going in the future?
We believe upwards of 75% of all planning in travel is
probably going to end up being algorithmic, meaning you will express some vague
or specific intent and the algorithms will guide you to a better outcome.
The same way algorithms decide what shows up in search
results, algorithms decide how your social media feed changes up over time and
even algorithms can drive the car that you’re in. We believe leisure planning
is exactly that kind of complicated problem where the AI will do a better job
over time.
We are already seeing signs that people are traveling more
when they use Hopper. Fundamentally if you had no work to do to plan any of your
travel, would you travel more often? Most people that I ask that question to,
the answer is yes. It’s a lot of work to put together leisure travel.
If you had an all-knowing engine that was continually
sending you interesting things that you could go through in five-minute
sessions over your phone - basically micro-moments or snackable content - would
you pull the trigger more often on a weekend getaway or even a trip to Europe
for a week? I think the answer is yes.
I think that not only are we reducing the cost of travel because
we are actually pushing to better outcomes, I think we are also increasing the
frequency of travel.
I think AI will have a fundamental role in reshaping how
travel is distributed, consumed, purchased that is going to be as big or
greater as putting the inventory online in the 1990s.
Hopper began testing a hotel price prediction engine last
fall starting with New York and just a few weeks ago you added Los Angeles,
Miami, Chicago, Las Vegas and San Francisco, with Boston, New Orleans, Orlando
and Denver rolling out soon. What have you seen with that product so far?
Fifty percent of the people that bought hotels from us in
the past weeks had never purchased air. So we know that the purchase is very,
very different.
Also when you look at the amount of trips being planned on
the platform - not necessarily what people purchase because not everybody buys
from Hopper; a lot of people use us as a recommendation engine but don’t
purchase from us, they might go directly to the airline website for example - when
you look at that, Hopper is being used to plan travel for about 400 million
room nights a year.
I think AI will have a fundamental role in reshaping how travel is distributed, consumed, purchased that is going to be as big or greater as putting the inventory online in the 1990s.
Fred Lalonde - Hopper
Whether those are with friends and family, in hotels or
Airbnb is a different question. But we already have a huge footprint in lodging
that’s not being used, because we know exactly when and where people are
traveling.
What’s fascinating about this problem is we’ve learned
that what people are watching in air is an 85% predictor of what lodging they
will purchase when they do purchase on the platform. That means the air signal is
going to be an incredibly relevant element to give to the AI to help make
recommendations.
The fact we have such a deep and frequent relationship with
our users is undoubtedly going to let us do things that nobody else has ever
been able to do.
When you look at how other companies - I’ll name Booking.com
because they’re famous for optimizing their things using technology - they’re
just trying to sell another hotel room. They do things like “four rooms left”
and all this pressure selling. We believe that the way to use AI in hospitality
is to help you find the perfect outcome or the perfect property, instead of
just getting you to buy whatever we can sell you as quickly as possible.
Will your hotel predictor cover the broad range of
accommodations that now exist around the world?
We are attacking everything. Right now we are hyper-focused
on hospitality in the traditional hotels, independents, chains, branded properties,
partially because there is a ton of inventory that is easily accessible, and
it’s still the majority of what people stay in.
But we fundamentally believe
over a long period of time, Hopper will automate all manner of travel planning,
which will include all manner of lodging. But that’s not in the app right now.
Hopper sells more than $1.5 million worth of flights every
day in the app. Now you’re offering hotels too. Is Hopper now an online travel
agency?
We definitely transact in the platform. We don’t lead-gen
out. There is no advertising in the app and there never will be as long as I’m
alive and running this company. You can’t click out. Everything starts and
stops in the app. The other thing we don’t do is we don’t forward distribute anything.
You’ll never find a Hopper fare available on any metasearch site, we don’t buy Google
Adwords.
In a way Hopper is a completely isolated universe and
part of what’s starting to happen now in air is that we are developing exclusive
pricing. This has to do with data and the AI. Our platform is able to predict
what somebody will buy before they buy it. It’s a logical extension that we can
get carriers to file fares that are only available to certain people.
Today about 12% of all the airfare on Hopper on a route
level is cheaper than the airlines websites, and these are fares that are only
available on Hopper.
If you look at what we’re doing in terms of the AI but also how
we built our business model, our goal is to decrease the price of travel
overall. We believe that this marketplace - data-driven, completely isolated,
no pay-per-click - is actually going to create a lower cost of product.
Because think of it: Every time that an OTA pays a metasearch
site for a click, which is hundreds of billions a year in revenue if you add up
all the players, that ultimately is a cost that’s being added on the price of
the ticket or the room. If you had an ecosystem where none of that was happening,
it would be much cheaper to travel.