Guide to AI dynamic

pricing in aviation: For

revenue leaders

Free

resour
ce
About
Datasparq
Datasparq’s team of 40 What’s inside...
data scientists, data
engineers, product
practitioners & business
experts solve businesses’ 4 Welcome from John 24 AI will help humans,
most complex challenges
using data & AI. Delivering
not replace them
real-world, impactful AI 5 A primer on revenue
solutions is at their core; management in aviation 27 Answering your
they’re “practitioners, not
PowerPointers”. questions
10 7 steps to becoming
Clients include: AI-ready for pricing 28 To be ethical, AI
must be explainable.
Here’s why
14 Reinforcement
Learning for Dynamic 32 Case studies
Pricing
34 What could your
business achieve?

35 “Will AI be able to
help my business?”

3 GUIDE TO AI-POWERED PRICING www.datasparq.ai

Data & AI making an
impact on revenue

It’s been a tough few years for the aviation

An introduction to
industry; with the COVID pandemic causing
chaos for international travel.

Dynamic Pricing in 2023

Accordingly, revenues of businesses across
the industry were hit. In 2022 we saw a
bounce back – and in 2023 we’re set for a
shift back to pre-COVID revenues.

As business returns to normal, there’s never Written by

been a better time to invest in maintaining or
Tom Nicholls
building a competitive advantage.

And as the world starts to wake up to the

value stuck in their data, choose to be part
of the group that stays one step ahead.

I hope you get some practical, valuable

takeaways from this guide.
Born in
the 80s
J.Wyllie John Wyllie
Managing Director
Airlines have long been
pioneers in pricing and
revenue management. In
fact, back in the 1980s,
the use of dynamic pricing
was kickstarted by the
aviation industry.
Rather than let the value rates are set to rise for dynamic pricing has highlight the importance solutions to the pricing Gain insight
of seats drop to zero, higher and higher over shifted from seats to of catering to the unique conundrum.
airlines began offering the long term. And as customers. The new ques- needs of different Knowing why AI makes
discounted fares for demand increases, so too tion? How much is this customer segments in Here are a few ways it’s decisions enables better
last-minute travellers. The does complexity. More specific customer willing dynamic pricing models. set to benefit airlines. collaboration between
objective was simple: min- passengers, more types to pay for a seat? And companies today have humans and machines. AI
imise the number of empty of passengers and more access to more data on Increase revenue should have explainability
seats to boost revenue. destinations mean pricing built-in; in other words, it
complexity adds up fast. The value of a buying habits than ever.
At the risk of stating the should be able to clearly

Revenue “So, how many seats do

seat But, for some time now, obvious, the core benefit tell you why it’s made the
it’s been difficult to unlock of using an AI-powered pricing decision it has. Not
management we sell at a discount?”
The same seat has a the value trapped in the price optimiser is that it only is this crucial knowl-
newfound treasure trove. can maximise revenue. edge for pricing analysts
today This brings us to the age-
different value for every
That’s exactly where an It does so by finding the and revenue managers,
customer. Business flyers
old question: how many but the legal requirement
are happier to pay higher AI-powered price opti- highest ticket price for
Today, the objective seats should you sell at a
miser can help. which there’s likely to be for AI to be explainable is
remains the same: max- discount? I’d argue that, fares than holiday-makers.
growing too.
imise revenue. But the in today’s market, it’s no And they’re happier to pay
means of doing it have longer the best question for add-ons that maximise
Continually improve
become significantly more to ask. As dynamic pricing flexibility, like refundable
complex. evolves - along with the tickets. Whereas holi-
A good AI-powered revenue opti- AI has the ability to intel-
technology that powers it day-makers book further
miser will complement an airline’s ex- ligently and automatically
We might describe reve- - pricing is becoming ever in advance and are more
isting revenue management system. analyse and learn from
nue management as the sensitive to price.
more personalised. the past, compare it to
science of maximising rev- It shouldn’t be a case of replacing or
the present, and make
enue – by setting the right Of course, there will still Yes – this doesn’t scratch replatforming.
informed decisions based
price, at the right time, for be “discounts”. But, the the surface of the fac-
on this learning. This is
the right passenger. meaningful question tors in play. But, it does reinforcement learning,
This means pricing fares and it is a key tool in a
in accordance with factors How an AI demand. With the ability modern dynamic pricing
like seat capacity, flight to analyse millions of solution’s arsenal.
schedules, traveller type price optimiser potential prices - eval-
uating using dozens of Enhance existing
and competitor prices – to
name a few. Simple in
can help parameters: flight capac- platforms
theory, complex in prac- ity, competitor prices and
The next paradigm of
tice. And today, it’s more A good AI-powered
dynamic pricing will have time remaining to book -
complex than ever. revenue optimiser will
AI at its core. And for good AI produces extraordinary
complement an airline’s
reason, too. Working in results. Quite simply, the
Whilst the industry at existing revenue manage-
present is recovering from harmony with humans, shift towards AI is usher- ment system. It shouldn’t
the impact of the COVID AI can explore millions ing in the next generation be a case of replacing or
pandemic, passenger upon millions of potential of dynamic pricing. replatforming; rather, it

7 GUIDE TO AI-POWERED PRICING www.datasparq.ai

 should be about enhanc- not set to change anytime
ing. We often hear this soon. The next paradigm
concern from those in of pricing for airlines will
strategy, operations or see AI take a central role.
innovation roles, but in
reality, there’s nothing to And those working in
worry about. Successful AI revenue, strategy & tech
solutions need good qual- roles will drive the change.
ity underlying data, some-
thing existing revenue In 2023 the value trapped
management systems can in data will start to be
provide, and they need unlocked by AI – those
to be tailored to address who embrace it are set to
specific challenges, reap the rewards.
something off-the-shelf
solutions struggle with. Tom Nicholls is the Chief
Commercial Officer at
Airlines have been pio- Datasparq.
neers in dynamic pricing
for decades – and that’s

ty
What is Explainable AI — and why
should I care?
Unlike Black-Box AI, it uses, how it has been
Explainable AI is built deployed and the outputs
in a way that enables it produces. With legis-
users to understand why lation set to make some
it produces the outputs form of explainability a
it does, and how it has legal requirement, now is
reached its decisions. the time to start making
Explainable AI references sure your systems are
the whole system, taking ready for the future.
into account the way the
model has been designed
and developed, the data Read more on p.28 ->

GUIDE TO AI-POWERED PRICING 8 9 GUIDE TO AI-POWERED PRICING www.datasparq.ai

ASK THE TEAM For those with products to exploit. This gave the businesses don’t capture
(often time-bound) with business (and us!) enough the times they offered

“How can we benefit

dynamic prices, thinking conviction that this was an something at a price and
about the advantages AI opportunity worth invest- it WASN’T purchased. This
might give you, here are ing further in. information is crucial and

from AI to make better the 7 things to consider as

you get started. 2. Instrument
really valuable to learn
from - and it’s often not

pricing decisions?”
yourself captured. If you don’t

7 steps to In order to understand

capture this data already,
instrument your business

becoming market elasticity, and give

any AI solution (likely a
to capture this data today,
and we’ll speak again in a

Written by AI-ready reinforcement learning

agent) enough data to
few months to analyse this
data and understand the
John Wyllie
for pricing get started with, it’s
much better if we can
market elasticity better.

understand how cus- 3. Define your

1. Test first to
tomers have responded objective function
A turning tide validate the to different prices
opportunity shown in the past. Most Any AI needs to be
To many executives,
businesses will always absolutely clear on what
putting AI to work in their Do we know how much
capture the price they it’s trying to achieve,
business seems daunting. market elasticity there is
sold something at - even which means you need
While AI is talked about a to exploit? How much do
if it’s a dynamic price (be to be absolutely clear on
lot, actually implementing we know already about
it an aeroplane ticket, a what your goal is. In a
AI in an operational enter- how different prices affect
parking space, a taxi fare, car parking example, an
prise setting remains too demand? If no one is
etc). But crucially, many optimal price to maximise
much of a challenge to the willing to buy at a higher
majority. However, as AI price, we can’t put prices
technologies and talent at their core, can now in ways that perform bet- up any more anyway - so
become more accessible, benefit from applying ter than the traditional OR we may not need AI. Some
the tide is starting to turn. AI techniques to these techniques. of our customers weren’t
More and more businesses problems. The aviation, doing any significant price
are thinking seriously travel and transport Recently, I was asked, exploration before - they
about where to deploy space are full of classic “How can we benefit from already thought they were
their first AI capability. OR problems - yield using the best prices.
AI to make better pricing
management being a So, before we deployed
decisions?” – a great
And that’s good news crucial one. Revenue a sophisticated AI pricing
question. Why? Because
– because businesses managers in aviation agent, we delivered some
that have traditionally can use AI to make it shifts the conversation simple A/B pricing tests to
had many Operations better pricing decisions from theory to practice; confirm there was further
Research (OR) problems that drive profits - often from talking to doing. elasticity in the market

GUIDE TO AI-POWERED PRICING 10

occupancy will likely be you don’t want to exploit. to be able to benefit from Manager is out of a job - technology when starting people and processes as
different from one which It’s important to acknowl- exploiting the findings. far from it. a new AI project. So, it’s well as technology when it
optimises revenue per edge and articulate these worth taking a moment comes to operationalising
space per day, or maxim- guardrails so the AI agent 6. Set your Their responsibility is to expand on the need to
AI. It often comes in the
ises total average booking knows its limits, and you communication plan still crucial, and uses have the right people and
form of a well-prepped
revenue. Understanding have the ability to step in their key skills that the AI processes in place.
and course correct as and Depending on how much can’t do: their role pivots in-house team or trusted,
the trade-offs of different
objective functions and when necessary. dynamic pricing is being to governing the RL/AI There’s an old Einstein long-term partner; a team
getting clarity and agree- done already in the agent setting the prices fable that explains why. around for the long haul.
ment on the goals of the 5. Be prepared to business when you start - guiding it (telling it how
to deploy it, how are you far to explore and exploit - Einstein gave an exam John Wyllie is the
pricing engine will set you take risks
going to communicate tweaking the guardrails), if paper to his class that was
up for success. Managing Director at
it to your staff and your necessary overriding and identical to one he had
You don’t know some- Datasparq.
customers? What if a given a year before. His
4. Think about thing until you’ve tried course correcting it (e.g.
customer has always in response to knowledge teaching assistant, upon
disadvantaged it. Just because you’ve
paid £10 and now they’re the RL agent doesn’t noticing this, let him know.
always charged £10
customer segments asked to pay £12? If they yet have - a new bank
doesn’t mean you might
that you don’t want be able to charge £12. call up your call centre to holiday that’s just been
“Professor, I’m not sure if
you know, but this test is
to exploit Macroeconomic changes query it, how do your call announced, or a volcano
centre colleagues explain identical to the one you
(the current cost of living that’s just gone off and
If you ask an AI agent this? This is an important gave out last year.”
crisis is a case in point) grounded all flights).
to optimise a price to mean you continually change management
After a short pause for
maximise an objective need to experiment and exercise to plan and get Defining this new role for
thought, he replied with
function (maybe revenue, test new prices in the right before deploying. your revenue management
something interesting.
maybe profit, maybe a market - and many of team, and giving them
combination of profit and those experiments won’t 7. Prepare for your the necessary support
“The test is the same – but
customer service metrics give as good results. commercial team to in training and tooling to
the answers are different.”
like wait time) - it will do it. help them in this transi-
Getting the balance of evolve
But in reality, you may not exploration vs. exploitation tion is another important The world, your business
always want the AI pricing right is a crucial aspect Today, you might have a business process change and the way your cus-
agent to fully exploit and of implementing a rein- Revenue Manager in the management exercise to tomers respond to prices
maximise the objective forcement learning agent business whose respon- plan and get right before are always changing.
function you’ve given it. (and giving it guardrails sibility is to determine deploying. The environment that the
For example, as a brand, on how far it can explore). and upload new prices optimiser operates in is
you may not want to be
seen to be too exploitative
Your business needs to be
comfortable taking some
and tariffs each day. Just
Same question, in continual flux (which is
because you now have why a continually learning
in some instances (e.g. risks (sometimes pricing a reinforcement learning new answer reinforcement learning
Uber surge pricing when things too low and making agent doing this more agent is a good approach).
there’s a natural disaster a loss, sometimes pricing frequently, using more Investing in AI is a long-
or a terrorist event). There too high and scaring data to determine optimal term play as well as But executives must also
may be some disadvan- away demand) to explore prices more intelligently a short-term win. It’s focus on the changing
taged customer segments different prices in order doesn’t mean the Revenue easy to over-index on needs and roles of their

GUIDE TO AI-POWERED PRICING 12 13 GUIDE TO AI-POWERED PRICING www.datasparq.ai

 Dynamic pricing is a grow-
ing area of data science
Dynamic pricing is [...] enjoying much
and a long-standing prob-
more attention as marketplaces them-
lem in operations research
selves become more dynamic.
and one that is enjoying
much more attention as
marketplaces themselves
might get a higher price for a return trip to Majorca, become more dynamic.
from another customer they’re likely to think that
What makes these
for the same date. If you £75 for a weekend parking
problems amenable to
lower your price for that
spot is a little on the high reinforcement learning
date because you haven’t
side. Whereas your busi- is vast event data sets
sold the spaces and the
ness traveller might think around customer inter-
date is approaching, your
that is entirely reasonable actions. This gives an RL
customers might quickly
tool the ability to deduce
learn that waiting is the – or at the very least, it’s
the demand level (as well
best policy and not book not an expense they’ll be
as, potentially, need and
when they first check the picking up.
urgency) for a particular
car park website.
cohort of customers and
The ability to learn what
thus find a customer-spe-
Furthermore, customers
price works for a given cific price that will lead to
who book for a weekend

Using Reinforcement
demand, for a given day a well-used resource.
might well be families or
of the year, with a given
tourists looking for a quick
Why dynamic
Learning for Dynamic Pricing
break who are perhaps booking lead time is a
booking with a low-cost good example of dynamic
airline. If they’ve paid £67 pricing.
pricing?
Before we get into
Written by Our Data Science team themselves, and you have reinforcement learning
were lucky enough to customer demand which and how it can power
Jeremy Bradley get into Reinforcement varies over time — largely dynamic pricing, we need
Learning or RL [1] several with demand for airline to understand a few use
years ago, when we travel. You also give peo- cases for why prices
started looking at optimal ple the ability to book might need to change in
pricing for car parking spaces well ahead of time. more general settings.
spaces at an international What price do you set?
airport [2]. Increased demand/
Imagine you run such a reduced supply. As in any
It’s an interesting problem: car park and have the economic model, if supply
you have a fixed resource, power to set prices. If you does not match demand
the car park spaces wait a couple of days, you for a product then the

GUIDE TO AI-POWERED PRICING 14 15 GUIDE TO AI-POWERED PRICING www.datasparq.ai

market may be able to a premium for a model. dynamic price — certainly considerable long-lasting want to offer the service If you read any reinforce-
sustain a higher price. The implication is that in mass retail — is the reputational damage. and the quicker a match is ment learning text book,
there is a reason to charge act of setting a different Important to consider— is made. you’ll find a diagram very
Expiring products. a varying price more price for the same product it the ethical thing to do? much like this describing
Anything that has a limited closely linked to the cost to different custom- The lower the price, the the basic principles. In
sale window is likely to better for the customer — this instance, the agent
attract varying a price,
of production (and pos- ers. Sometimes this is
veneered over by selling
Constructing of course — but also the is the marketplace, the
sibly the customer’s will-
whether it’s a hotel room ingness to pay for quick in different markets, e.g. an RL model... longer they may be kept action is the ability to set
for tonight, a loaf of bread electric cars in the US waiting for a match with a a price and offer it to the
delivery — see below).
that’s about to reach its versus the EU market. and mistakes possible service provider customer, the state is the
expiry date, or a plane state of the marketplace (I
ticket on a flight. None of
Price by customer. But within a market,
to avoid! and indeed they may get
know that’s self referential,
Possibly the most conten- it may be desirable to no match at all.
these can be sold after tious reason to charge a identify customers who but we’ll revisit that) and
Here’s an example from
the window has passed. How do you set prices the reward is a measure
another dynamic pric-
in such a setting, so that of success from having
ing project we’ve been
Seasonality. Some made a successful match
working on recently. It’s a you offer a fair price to
products are much more Important to consider — is using rein- between customer and
marketplace environment the service providers, but
desired at certain times forcement learning the ethical thing to don’t “fleece” the cus-
service provider.
where customers are
of day or times of year. do in this particular case?
exposed to a set of agents tomers? Well, you can use
Some may not be available Again we’ll come back
offering a service and we reinforcement learning. We
at all at the wrong time. to this because it’s not
have to work out what don’t know initially know straightforward and it will
Christmas decorations
are time-limited and are price to set. The market- the best price – but we drive the whole optimisa-
typically get discounted
not looking at competitor place takes a cut of the can set one and see what tion of the system. There
heavily after Christmas
products and may be fee paid and the service impact it has on the mar- can be a choice between
as retailers don’t want to
willing to pay more for a providers get the rest. ket — then we can learn many possible actions
stock them for the sub-
standard product. and in the next epoch reflecting the possible
sequent year waiting for
The basic features adjust accordingly. prices that may be offered
them to sell.
Varying a price to reflect
Customer demand is sea-
Cost of production/ super-high demand or
sonal and varies dramati-
supply. Some products capture a high-need
cally across the day — it is
cost more to produce customer is not always
also price sensitive.
on demand than on the the right thing to do.
normal production run. For For instance, airlines
The service providers are
example, Brompton, the charging thousands of a limiting resource — if
folding bike manufacturer pounds for seats on the they’re working for a
based in London, typically last flight that makes it customer they can’t be
has a 5-week lead time for home before an enforced doing other jobs for other
a normal folding bike but quarantine period starts customers.
that can be reduced to 5 or a taxi company surging
days with tracked delivery prices in the face of a civil The higher the price, the
if you are prepared to pay emergency — can lead to more the service providers The reinforcement learning loop. CC by-SA 4.0 Jeremy Bradley

GUIDE TO AI-POWERED PRICING 16 17 GUIDE TO AI-POWERED PRICING www.datasparq.ai

in the marketplace. Each That was lots of jargon so that spans 850 locations
action may lead to a dis- let’s break that down— a across the world and we
tinct change of state and value function computes want to set prices every 15
a different reward associ- the value of each state minutes, we have 20 levels
ated with that change. of your system under of customer demand and
a given action policy; 20 levels of service pro-
So what does the Agent knowing which states are vider supply, and 20 pos-
do — well the marketplace high value is important sible price bands. That’s a
learns from successful when discovering a bet- little over 4.5 billion states
matches and unsuccessful ter policy. A policy is a right there!
matches and adjusts the pre-defined strategy for
choosing which action to Surprisingly this is not
price (or not) accordingly
take from a given state, necessarily a disaster as
as each new opportunity
ideally to maximise your you can represent your
to make a match comes
accumulated lifetime state space approximately
around. There are maybe
reward. A Markov Decision using a deep learning neu-
20,000 transactions a day
Process or MDP is a ral network of some other
and prices are set in near
state transition system ML model but it definitely
real-time.
which defines the states, adds to the complexity as
actions, rewards and next we’ll see when we try to
I think the three most
states for any system. learn with it.
important aspects of
doing this kind of activity
What’s this got to do with What does success
with reinforcement learn-
ing can be summarised:
states? If the underlying look like?
mathematical model
(MDP) has too many Success is defined by
What’s in a state? states, then the dynamic your reward. You might
programming algorithm think that success is
A quick primer:
will take forever to con- easy to specify. Why not
Reinforcement learning
verge (or a very long time choose revenue or profit
comes out of a branch
anyway). i.e. you’ll never and watch the dollars roll
of mathematics called
get a solution and you in. Well yes and no. Let’s
dynamic programming won’t know what action imagine we used revenue
which solves an underlying to take — or price to set, in our marketplace.
Markov Decision Process in this case — to improve
for a value function. your overall lifetime If I have some very rich
reward. customers, then I can set
Deriving a value function my prices very high and I
gives you a policy which It is very easy to inad- may initially even make a
in turn tells you which vertently give yourself an lot of revenue. However, I
actions you should take in enormous state space. have inadvertently made
a given state. We have a marketplace my market very brittle and

GUIDE TO AI-POWERED PRICING 18 19 GUIDE TO AI-POWERED PRICING www.datasparq.ai

subject to the whims of they get a price match. which captures both model having a consistent your learning feedback recommended prices or
a few well-off customers But that is ignored by the financial success and picture of what value the loop needs to visit all actions are not going to be
customer satisfaction will most important states and of those states 100s or great!
who are now dominating reward metric as it only
allow you to find a sweet actions have). 1000s of times to get
the market. Every other cares about revenue in
customer is priced out this case.
spot in your problem —
keep the service providers Now we know why keep-
converged learning.
Wrapping up
happy with a decent ing track of the state If you’re using an approx- I’ve really only touched
revenue while keeping the space is so crucial. If imate state deep learning upon reinforcement learn-
customers happy with a you’re using tabular learn- model for very large state ing as a mechanism for
good performance metric. ing, which is simple and spaces, you won’t be learning optimal prices.
has many benefits, you surprised to learn that
Can it learn? have to have a pre-training many gigabytes of training There are lots of benefits
set that is 100x or 1000x data are needed to train to getting it right — an
Reinforcement learning
the number of states in that model effectively. algorithm that can learn
works best when it learns,
the state space or at least autonomously and present
as you might expect.
the number of states you Certainly, if your training the right price for even
Successful learning comes
are likely to visit. data only visits each new unseen scenarios;
from convergence of the
state once or twice, an algorithm that can
algorithm and gives you a
Alternatively, if your model your learning is going to even look after customer
A simple 3-state MDP with actions in brown and rewards +5 and -1 marked. consistent result for the
is learning from scratch, be very poor and your satisfaction if the correct
CC by-SA 4.0 Waldo Alvarez value of being in a particu-
reward structures are put
lar state.
in place to guide learning.
and it’s unlikely that they’ll Again this is a short-term
come back. win. Those customers left If you’re using an explicit
There are many ways
hanging around probably state space model
you can get yourself in a
So let’s lower the prices never return to your mar- (sometimes called tabular
tangle. But keeping the
so that I open myself up ketplace even when they learning, so not a neural
state space at a reason-
to more of my customer have larger jobs to offer. network) you need enough
able size that can match
base. The prices are
the available training data
still high. The revenue is
you have (or are likely to
improved because the
A hybrid reward metric which captures encounter during learning)
customer base is larger,
both financial success and customer is a great start and defin-
but there are service pro-
satisfaction is the sweet spot. ing an appropriate reward
viders who are now turn-
metric will help you drive
ing down small jobs that
your learning in the right
are too low in value in the direction while not making
hope that a higher-value In the end, low customer training data or learning embarrassing decisions
job will come along satisfaction kills your episodes to visit each along the way.
shortly. Many customers business. state many thousands of
with smaller jobs are left times to get converged This introduction is still
waiting for an unaccept- What we found is that learning. (Converged only the tip of the iceberg
able period of time before a hybrid reward metric learning refers to the but it reflects some of the

GUIDE TO AI-POWERED PRICING 20 21 GUIDE TO AI-POWERED PRICING www.datasparq.ai

 Meet the team
key decisions we had to many techniques and pros
take when taking rein- and cons to each of these
forcement learning from approaches but in general,
the textbook [1] (which starting simple and build-
is quite excellent by the ing up the complexity as
way!) to a real-life project. you need it is good advice. Our team of data and
business experts pride
Other significant Off-policy versus on-pol- themselves on being
icy learning — is your practitioners, not
considerations
system able to learn as it PowerPointers. Here are a
include... discovers new states or few of their friendly faces.
is it going to have to be
Exploration versus
pre-trained offline? Is your
exploitation — at some
system able to implement
point you’ll need to take John Wyllie Jerermy Bradley
the decisions it takes or is
a suboptimal decision to memory and rapid learning Managing Director Chief Data Scientist
it only recommending the
see whether it leads to rates are the order of the
decisions to an operator
a better policy that was day here.
who can decide to adopt
previously unknown. This
or decline them? All these All of these features need
can trigger some inter-
factors influence whether addressing at one point or
esting conversations with
stakeholders! you need to use an on- or other in practice. Let me
off-policy approach. know if you’d like to see a
Discount factors — how Part II on some of these
much do you value reward Time-varying rewards
more technical aspects of
that can’t be realised for and demand — probably the practical implemen-
weeks or months versus the hardest issue of all in tation of reinforcement
Oli Bartlett James Watkinson
reward you can access reinforcement learning is learning. Director of Product Director of Engineering
today? time-heterogeneity.
Jeremy Bradley is the
Learning rates — how Your environment may Chief Data Scientist at
quickly should your learn- change over time and the Datasparq.
ing system adapt to new reward you see one day
may differ the next day This article was originally
evidence or is your data published the Datasparq
quite noisy, thus meaning for the same action. If this technical Medium blog.
you don’t want it to adapt means a change in the
to a blip by mistake. statistical properties of References

your system (rather than

[1] Richard S. Sutton and Andrew G.
Algorithm — where to just noise or differences Barto. Reinforcement Learning, An
Introduction. 2nd ed. MIT Press, 2018.
start; tabular or approx- in samples from the same Ishbel Hughes-d’Aeth Tom Nicholls
imate representation, distribution), then your RL [2] Andreas Papayiannis, Paul V. Johnson, Principal Product Lead Chief Commercial Officer
Dmitry Yumashev, Peter Duck. Revenue
Expected-SARSA or system needs to know to management of airport car parks in con-
tinuous time. IMA Journal of Management
Q-learning. There are adapt to that. Censored Mathematics. Jan 2019.

GUIDE TO AI-POWERED PRICING 22 23 GUIDE TO AI-POWERED PRICING www.datasparq.ai

AI will help humans, not
Black Mirror, Westworld ground that need to work impact on the outcome
and classics like The with the technology to we’re predicting. For
Matrix all depict worlds in unlock its value. example, as someone

replace them
which data- & AI-driven who’s worked in TV,
digital beings call the For example, when busi- I know that weather
shots a little more than nesses use AI to power impacts viewing patterns.
we’d like them to. their dynamic pricing Therefore if I were to build
strategies, revenue man-
Written by The possibility of
It’s easy to dismiss the agers and pricing analysts
an ML model to predict
AI-powered super-robots viewing figures for a
Oli Bartlett taking over the world has above examples; they’re work alongside their new particular channel next
appeared in fiction in lots fun, but ultimately a little AI friends to turn their Wednesday afternoon
of fun ways over the past far-fetched. However, deep analyses into prices
in London, I would add
50 years or so. concerns about AI replac- that work for the cus-
weather data to the train-
ing humans at work are tomer and the business.
ing set. The results can be
closer to home. Computers are adept
astounding. But how do
at solving mathematical
you get there?
At Datasparq, we believe problems far beyond the
that AI will augment grasp of any human but
human intelligence, not need human experts to AI as a product
replace it. AI will enable guide them both in solving
humans to make better the right problems, and in To realise value from
decisions, get more done knowing where to start. AI - as I mentioned earlier
and free up their time to - you need a team that
work on the stuff that An ML model needs to be knows how to train it, use
really matters. fed the kind of data we it and apply it to real-
think is likely to have an world situations. Without

“Great servant,
terrible master”
The bottom line is that AI
makes a great servant,
but a terrible master. What
exactly do I mean by this?

AI is a great solution for

solving complex problems
but without human experts
guiding and directing it, it
won’t be able to do much
at all. In a business con-
text, it’s the experts on the
 Answering your
questions

? Can your AI-powered pricing optimiser integrate

with my existing revenue management system?

An AI-powered revenue optimiser will nothing to worry about. Successful AI

complement an existing revenue man- solutions need good quality underlying
agement system. It shouldn’t be a case of data, something existing revenue man-
replacing or replatforming; rather, it should agement systems can provide, and they
be about enhancing. We often hear this need to be tailored to address specific
concern from those in strategy, operations challenges — something that off-the-shelf
or innovation roles, but in reality, there’s solutions struggle with.
that, it doesn’t matter how as pure technology trapped in the treasure
sophisticated your tech is; projects. Technology is troves of data in your GCP,
its value won’t make it into one part of the puzzle, Azure and AWS servers.
the real world. but effective AI solutions
come from multidisci- Oli Bartlett is the Product

?
As the Product Director plinary delivery teams Director at Datasparq.
Do I have the data I need to start exploring an
at Datasparq, applying (design, product, data
AI-powered pricing solution?
product thinking to data science, engineering),
science and engineering working closely with end
projects is a core part of users, data experts and IT.
my job. The ultimate goal This way we can ensure
is to ensure that people To give any AI solution enough data to they offered something at a price and it
that our ML products are
get started with, it’s best if we under- WASN’T purchased. This information is
can actually use and get more than just great data
stand how customers have responded to crucial and really valuable to learn from
value from AI. For that science. They are usable,
different prices shown in the past. Most - and it’s often not captured. If you don’t
to happen, people and resilient and solve real
businesses will capture the price they capture this data already, we can help you
processes must be at the problems for the business.
sold something at - even if it’s a dynamic instrument your business to capture this
core of the project.
price, be it an aeroplane ticket, a park- data today, and we’ll speak again in a few
Ultimately, treating AI
ing space, a taxi fare, etc. But crucially, months to analyse this data and under-
To that end, at Datasparq as a product is essential
many businesses don’t capture the times stand the market elasticity better.
we don’t treat AI projects to unlocking the value

27 GUIDE TO AI-POWERED PRICING www.datasparq.ai

 In the 1. Improving decision enables data
scientists to pinpoint the
the quality of
limelight
root cause of an issue and
take corrective action,
outputs both during development
The ethical implications of and ongoing maintenance
AI have been a hot topic Quite simply, if you don’t of the system. In fact,
of debate for years now – know why AI is suggesting Explainable AI even helps
and for good reason. low-quality or biased improve efficiency of the
outputs, fixing it is going development and ML Ops
As AI becomes more to be very difficult. Having teams by quickly highlight-
advanced, it’s being relied
an explanation of how ing issues and potentially
upon to make increasingly
AI came to a particular catching these early.
important and quite often
automated decisions,
which have more and more
wide-reaching impacts on Explainability gives humans the
people, planet and profit. chance to understand why an AI has
So, it’s essential that we’re produced a particular output.
able to understand why
these systems make the
decisions they do.

Unlike Black-Box AI,

Explainable AI is built
in a way that enables
users to understand why
it produces the outputs
it does, and how it has

To be
Written reached its decisions.
Explainable AI references
by Ishbel the whole system, taking

ethical, AI Hughes-d’Aeth into account the way the

model has been designed

must be
and developed, the data
it uses, how it has been
deployed and the outputs

explainable.
it produces.

Here are 3 reasons why

Here’s why explainability is an integral

part of ethical AI.

GUIDE TO AI-POWERED PRICING 28

2. Building Datasparq, shared a fasci-
nating insight with me.
section asking users
to share demographic
trust with information. It’s under-
“Bias in AI isn’t necessarily standable why users might
users the result of a model sin- be wary; it seems like this
gling out a demographic is how bias might be built
Explainable AI helps users in. But, it’s actually the
trait. It often forms
understand where their opposite.
through secondary data
data is being used and
points. If there are lots of
3. Sharing its decision-making and
build on it, and likewise,
Similarly, the UK has pro-
posed an AI rulebook in its
how it affects the outputs “It might seem odd,
of the system.
people with overlapping learnings intervene and course-cor- AI Action Plan.
demographic traits living but having this addi- rect AI systems when we
It allows users to chal-
in the same area, for tional demographic At Datasparq, we believe know something that the So - a word to the wise -
instance, a model might data makes bias that AI works best when machine doesn’t. it’s worth making sure that
lenge the results if they
exhibit bias against a it works with humans – as any AI systems you’re cur-
feel they have been more preventable.
group by discriminating on part of the team. rently using are prepared
treated unfairly by the
the basis of a postcode.”
When we have the Incoming for the new guidelines.
algorithm. primary data, we can
find and fix any bias
So, just as you wouldn’t
legislation
It enables those respon- That’s why companies,
resulting from the
expect a colleague to
refuse to tell you why they
Wrapping up
sible for the AI decisions when collecting data, Explainable AI is already
to verify that they are often include an optional secondary data.” decided to set a certain
playing an important role Explainable AI is key to
price for an airline ticket, Ethical AI. It helps identify
indeed the right decisions in helping organisations
you shouldn’t expect AI to and fix bias, builds trust
to take, that they meet the demonstrate that legal,
refuse - or lack the ability with users and allows us
organisation’s ethics code, regulatory and compliance
to give an answer - either. to learn from and build
business objectives and requirements are being
regulatory requirements. met. on the decisions that
Understanding why AI has
machines make.
made a particular sug-
It supports the adoption Given the importance of
gestion can help a human And with legislation set to
of new technologies and explainability in AI, it won’t
identify the best way to make explainability a legal
systems by businesses come as a surprise to hear
practically use it, too. requirement, now is the
and the public, helping that the principle will likely
to make investment in AI time to start making sure
For example, understand- be entrenched in law in
worth it. Winning users’ your systems are ready.
ing what has led to a the coming years.
trust is essential for com-
prediction of a specific Ishbel Hughes-d’Aeth is
panies that use AI. The EU is developing its
risk occurring provides the Principal Product Lead
insight into what business AI Act, set to enshrine
How bias can at Datasparq.
change or intervention can the principles discussed
unintentionally be made to address and above in law. Though it’s
occur mitigate this risk. still in the early stages of
being developed, it could
Jeremy Bradley, Chief When AI is explainable, become law by the latter
Data Scientist at we humans can learn from half 2024.

GUIDE TO AI-POWERED PRICING 30 31 GUIDE TO AI-POWERED PRICING www.datasparq.ai

Increased revenue for Increased revenue for a
Gatwick Airport through ridesharing app through
AI-powered Dynamic Pricing AI-powered Dynamic Pricing

What if powering your dynamic pricing with AI could boost your What if introducing AI-powered dynamic pricing to your business
revenue and help your commercial team make better decisions? could improve your customer experience and increase revenue?
What they said About Datasparq
Gatwick Airport is London’s second to design, deploy and measure Facing increased competition for optimise outcomes for passengers, At Datasparq, we use data
“The Datasparq team were
largest international airport and different pricing strategies. passengers and drivers, the UK’s drivers and the business. science, cloud data engineering
great to work with and their
second busiest by total passenger largest private hire vehicle operator and product thinking to solve
6 weeks to value data science & engineering Effective pricing
traffic in the United Kingdom. Pre- needed to harness the power of its and automate the most difficult
pandemic, in 2019, Gatwick served expertise has really helped This new, data-powered dynamic business challenges today.
In 6 weeks, we built and deployed data to unlock a more intelligent
46.6 million passengers. a new, scalable data platform
us tackle some high-value pricing capability is designed to
way to determine journey prices. To It’s not about how much data you
and interface for the business to challenges. They created enable efficient, consistent and have; it’s about how you put it to
The challenge highlight the scale of the challenge,
interrogate pricing strategies. a new data capability for improved pricing across different work. Whether you have a little or a
its drivers deliver more than 20m
Gatwick’s existing manual pricing us and delivered a solution regions of the UK, ensuring that lot; without the right capabilities its
Working with the business, private hire journeys each year with potential will never be realised.
engine was neither flexible nor we can use to inform our prices are relevant for local markets
commercial and digital teams, a transaction value of £150m. down to the neighbourhood-
scalable enough to cater for the revenue management Leveraging machine learning,
we established a new price level. In addition, the new BI &
growing complexity of its business. decisions.” Increased visibility cloud-based data platforms and
testing capability, including new AI capabilities, augmented by
With only some irregular trading product thinking, we can unlock
processes and tools to enable Datasparq built a scalable cloud the team’s regional business the value in your data and extend
dashboards being used by the Gary Wallace, Head of
multiple pricing strategies to data platform and supply & demand expertise, have enabled the the capabilities of your business.
business, Gatwick wanted to Car Parks & Commercial
be deployed and compared forecasting tool which integrates business to run pricing tests to
improve its use of data & AI to make Products And we can do it at a fraction of
concurrently. better understand passenger price
better commercial decisions. new journey data on a daily basis. the time and cost that it would
elasticity, enhancing the business’s
The results These tools give regional operators take to build in-house. We want all
A dozen potential use cases commercial strategy.
visibility into business performance, organisations, large and small, to
Working together, this is what Strategic partnership benefit from their data.
When Gatwick first engaged us, avoid data misconceptions that driver supply, passenger demand
we’ve achieved so far.
we ran a SparQshop that identified could skew analyses. and passenger behaviours. We can help you every step of the
The solution enables the business
a dozen use-cases to improve Superior performance and way, from identifying the right use
Robust security: through IAM and Reinforcement learning to achieve an increased number
its current pricing methodology, scalability: internal teams are case, through to building, running,
encryption at rest and in transit, the of bookings, revenue, customer
increase revenue and enhance hosting and managing your
benefiting from faster and more What’s more, a reinforcement service metrics and balance in
solution is secure. solution. All while keeping business
commercial decision making. reliable performance allowing learning model was built demand vs supply. value and ethics top of mind.
them to make better informed Looking to the future to determine appropriate
Jointly, we prioritised the use- As long-term strategic partners, we
commercial decisions.
cases and began the first phase We’re proud to be Gatwick’s interventions. These interventions continue to support the business
[email protected]
of work, which saw us develop a Data lineage control: Gatwick can long-term strategic partners and include price changes for specific with its tooling and the new
locations at specific times, to www.datasparq.ai
new data platform and analytics track data lineage to better assess continue to support its AI-powered capabilities that are being unlocked
workbench that could be used the trustworthiness of the data and dynamic pricing capability. influence supply and demand and by our optimiser.

CASE STUDY CASE STUDY

 “Will AI work for my business?” —
this is the answer to that question
What could you achieve with an AI- It can be daunting—you’re not solutions. For each solution, bolster your bottom line—it
powered pricing optimiser? alone in feeling that. Business
folks; we need to jump the
we’ll craft a Product Primer.
Our Product Primers cover
goes without saying, there’s
business value in this alone.
I hope that you’ve If you feel inclined to do the heart of the paradigm hurdles of business cases, everything: success metrics,
managing change and ROI. data science, risk—and As you may have guessed,
enjoyed this guide. so, drop me an email with shift happening in busi-
And technical folks must build everything in between. You’ll SparQshops require a big time
any follow-up questions nesses around the world.
Hopefully, you’ve not only an understanding of how it all have clarity on how AI can commitment from our team.
you have.
Right now, there’s a size- The good news is—we love
found it interesting — actually works and oversee help your business—and a
If I can answer them able opportunity for busi- doing them. The bad news
you’ve taken away some its implementation. But, that’s plan to move forwards.
myself, I will — or, I’ll put nesses - and executives is—we can only run a limited
practical insights, too. exactly why our experts are on
- to reap the rewards of Here’s where the magic hap-
you in touch with one of hand to help from start to fin- number each quarter. So if
being an “early adopter”. pens. You take the Product
As a team, we’ve tried to the friendly faces you’ve ish. From finding use cases, to you’re serious about adopting
Primers away and decide
cover a number of bases; seen in this guide. Speak soon, identifying the highest-value AI, we recommend that you
which one you’re most excited
we included content for opportunities, to keeping register your interest now.
If you’re already sure John about. Let us know, and we’ll
both technical and busi- everything running. Bring up to six key stakehold-
that you’d like to explore draft up a proposal for some-
ness experts alike. [email protected] ers, a decision maker… and a
whether our AI-powered Whether you have a hunch thing we call a Prove Phase.
willingness to get stuck in!
pricing optimiser could about an AI opportunity, or Really, it does what it says on
However, it’s a big - and
help your business, ask for you’re curious to find one, the tin. It proves the solution -Tom Nicholls, CCO
frankly, complex - topic.
the next step is a SparQshop. works before we commit to
a SparQshop!
So I’m sure that you still Together, we’ll explore poten- building the production-ready Request a SparQshop; email
have a few questions Before I sign off, I’ll say tial use cases and identify, product. Using real data, we [email protected] with the
unanswered. one more thing... AI is at prioritise and define AI verify that the solution will subject, “SparQshop Request”

GUIDE TO AI-POWERED PRICING 34 35 GUIDE TO AI-POWERED PRICING www.datasparq.ai

www.datasparq.ai
© Datasparq 2023