In Part 1, I posited that one of
the reasons “data Science” had such low RoI was that the expectations of the
“business” were too low. This brings me to the importance of metrics, both on
the revenue and costs sides of the equation.
I mentioned that the business had
no idea what can be done with “data science” and that 2 ways of bridging the
gap were to get “data scientists” to talk the “business language” (have enough
domain expertise or at least be able to communicate on equal ground with
someone who does) and secondly, the advances of technology should be embraced
since, at least they speed up analysis and should reduce costs.
So what is the cost of “data science” and is it really
that hard to measure it?
First of all there are direct costs to “data science”: the cost of the “data
scientist”, the apportioned cost of the IT infrastructure (that can be low if
you use open source solutions and have proper IT infrastructure roadmap –
something that is beyond me to do, but which is aligned with a “data science”
roadmap which I think is essential) and the costs of running the experiments/campaigns.
I had a very interesting
conversation at Strata in Singapore this year with someone from Swansea
University. They have data analytics and data science programs for all sorts of
different skill levels in the field of health sciences.
It does make sense, right, that in order to be qualified to be a data
scientist in a field where lives may be at stake, you’d need proper background,
domain knowledge? For example, the Master’s programme requires some real
life application in the field by practitioners so they can be awarded the
qualification.
It makes sense because the most
people would agree that the cost of mistakes in the medical field is very high.
Hence domain knowledge is seen as critical. But are the costs of “data science” mistakes in the commercial world
low?
I have a friend who works at one
of the large banks in Singapore who has a Masters in Analytics from the
Singapore Management University (SMU) and has extensive work experience in the
banking sector. She was asked to hand over her model building role to someone
so new that her only knowledge of the banking world was in Human Resource, and
focus on churning out campaigns. Her manager explained the decision by saying
that building models can potentially cause less harm than doing campaign
management (cleansing the lists which ironically may have come from the
models).
This brings me to the idea of potential harm.
My friend’s manager simply
believes that little harm can come from the application of “data
science”/analytics. If that was true, then
“data science” is the Holy Grail: little potential for harm, but huge potential
for good! The adage “if it looks too
good to be true, it probably is” surely applies here.
My friend’s manager is totally ignoring the opportunity cost of contacting a
customer. Many organisations today have relatively well oiled campaign
machines, churning offers to customers on a regular basis, and there usually
are contact policies in place: a customer cannot be “disturbed” that many
times. Basically: any offer you contact
a customer with implies there’s another he/she is not getting.
Furthermore, going back to the
customer contact policy, there is the
classic argument of customer fatigue. Unless the contact policy has been
set ‘properly’, there is a risk of customers simply ignoring the communications
from an organisation if these constantly bring no value. The customer need not
complain, simply ignore. Setting the contact policy itself is thus critical and
is something relatively simply analysis can reveal.
With today’s technology,
behavioural factors are being taken into account more often in “data science”
algorithms, and “real time” offers can be made following triggers based on each
customer’s behaviour. For example, if I know you usually like a good cup of
coffee after your meal, once you pay for a meal using your credit card, I can
almost instantly make you aware of offers for coffee with that card in the
vicinity (combining behavioural pattern and geo-location to make you a relevant
offer). This can even go a level deeper, for example, I may know your child’s
birthday is in a week, and after you finish your meal, I decide, instead to
send you an offer at the nearby toy store. How I decide which offer to make can
depend on the chances of success, and the relative ticket sizes, but all this
can be fed into your fulfilment engine and the ‘best’ offer made on the fly.
It is however important to
distinguish between tactical and strategic moves. A danger when using “real time” offers is these are often purely
tactical, contextual yes, but purely short term. It can be argued that if
that through my offers I manage to make a customer move my credit card to the
top/front of the stack of cards he/she holds, then the customer would use my
card more often with or without offers (free money!). But customers learn, and
their expectations change, may be they will wait for that offer at a toy store
before purchasing the present, after all, the child’s birthday is still a week
away; Doctor Pavlov’s experiments were with dogs, not humans.
Hence costs should be expanded to include opportunity costs, but also a
strategic Customer Lifetime Value (CLV)/Customer journey component. An organisation’s relationship with a
customer should not be a series of short term interactions, but, hopefully a
long term one punctuated by strategic reinforcements[1].
If costs have to take into
account not only opportunity costs, but also a longer term view of the impact
on the relationship with the customer. So should revenues.
Some organisations have attempted to bridge the gap between the short
term tactical approach, and the long term strategic engagement by creating a
series of small journeys a customer engages into, and where the organisation
wants to help, profitably.
For example, in the case of the
child’s birthday, based on past patterns or similar people, the journey can
include the gift, booking a venue for the party with friends, arranging for
caterers for the party, engaging entertainment (may be a clown?)... A bank that
identifies these steps a customer is likely to take within a kid’s birthday
journey can send offers to the customer that make his/her preparation for the
birthday easier while generating more revenue for the bank (using the bank’s
credit cards at participating merchants).
This is a step in the right
direction. But eventually, this is only one micro-journey; it makes more sense for an organisation to work on the macro journey,
from customer acquisition, growth, reaching a plateau, and may be exit, and use
these micro journeys to manage the customer, creating and reinforcing habits.
When you have millions of customers is where “data science” can really shine,
allowing you to personalise every journey based on the strategic decisions made,
and even affect customers via their social network [2].
Hence the costs of not making the right offer at the right
time to the customer go much beyond the simple costs of the hardware, software
and time of the “data scientist”, but only if the organisation is set up to
really take full advantage of data science.
To me, an organisation needs to be properly set up to enjoy
the benefits of “data science” and understanding the actual costs of bad “data
science” is only part of it, but an important starting point. As mentioned above, the best way to do this is to
align the infrastructure roadmap with the data science roadmap so the “data
science” adventure pays for itself, fully realising the costs and revenues.
This will be the topic of my next instalment.
To summarise:
- While many organisations get low RoI out of “data science” the RoI is even lower that they believe because they tend to underestimate the costs of data science
- While I argued that technology should be fully utilised and would be likely to bring down costs, the hardware, software and “data scientist” time costs are just a small part of the costs.
- Many organisations have contact policies and campaign machines, they fail to include the opportunity costs when they calculate the costs of a campaign.
- Furthermore, with technological advances, it is not that difficult to make real time offers to customers, trying to be the “right offer at the right time”, to be contextual.
- However, very often these are purely tactical in nature and do not maximise what “data science” can do for the organisation.
- Some organisations have attempted to build event based customer journeys, and while these are a step up from purely tactical approaches, they are still quite short term in nature.
- To understand the true costs of “data science” and reap its full benefits, an organisation has to exploit “data science” across the whole relationship with the customer, not just in a purely tactical piece-meal fashion.
- How well organisations have been setting themselves up to benefit from “data science” will be the topic of my next post.
