continue to candidate homepage




Continue to client homepage

Machine Learning and Customer Service

Luke Parkinson
Data Science Specialist

It's crazy to think about how Machine Learning (ML) is being used in a wide variety of different industries.

On our recent Data World Tour, we stopped off in Copenhagen where we were host to three, very interesting speakers.

I've written an overview on all of their talks for you.

Our first speaker, Gopal Karemore, discussed how AI has had an impact on the healthcare industry as well as drug discovery research for many different diseases, such as breast cancer.

Our second speaker, Kim Falk, gave us an overview of recommender systems and problems that they bring but offered us multiple solutions to these problems.

Our third and final speaker, Phillip Jarnhus, gave us an insight as to how ML helps to provide better customer service in banks.

Philip is a Senior Data Scientist at Danske Bank.

He specifically looked at how ML is used by Danske Bank to improve their customer satisfaction, provide predictions, see whether these predictions are accurate, and how they need to be altered if not.

Even though Philip discusses ML in relation to Danske Bank, it’s important to recognise that the insights he offers can be applied to other real-world situations and companies.

He didn’t discuss a specific ML model, but rather focused on the general challenges faced when working with ML models at Danske Bank.

The group Philip works for is one of 3 centres of excellence in terms of advanced analytics.

They focus on credit, credit risk and money laundering.

As a Data Scientist, Philip’s main work includes predictive analytics where he works to develop predictive models; these predictions can take place anywhere from 3 months to 3 years.

He states that if they can’t measure an effect within a couple of months, then they’re not carrying the project out well enough.

Here’s the deal:

Advanced analytics can help create customer satisfaction

From the analytics collected, Philip’s group can create customer values by linking the analytics to the strategy that’s in place.

Part of the strategy is to make the customer happy, so they stay with the bank. This is key to growth over time.

This is why customer satisfaction is so important to the bank.

Philip also tells us that the group starts with as little data as possible, so that they have room to grow, move upwards and with this comes a lot of education.

The group can start to build user stories which will contribute to the education.

This is beneficial as the rest of the business can learn to understand the strategy and what the group is working toward.

They need to see that the process is working before it starts being adopted.

Finally, the group needs to execute the plan, so they can measure and communicate the impact it has on the bank’s customers.

Philip explains that out of a 6-month product cycle, you’d spend at least 1 or 2 months at most doing the product modelling, while the rest of the time is spent on understanding the business and going back to measure the impact.

The bank needs to make sure that it has everything mapped out, so it can clearly view what’s working and what isn’t.

This also points the bank in the direction of what needs improving.

No doubt about it:

Relevance is key

The bank uses commercial analytics to make sure it's relevant to the customer.

Commercial analytics lets the bank use the current surveys asked the customers to see what’s irrelevant and what’s relevant in customer satisfaction.

The bank doesn’t want to provide a customer with some sort of contact or communication that they don’t see as relevant; this is likely to leave them less satisfied than if we hadn’t contacted them at all.

You see, relevance affects customer satisfaction.

Advisors are a good asset to the bank and have been proven to have a positive and powerful impact on customer satisfaction.

Advisors are able to deliver good customer experience and fix model shortcomings in real-time; however, they are expensive in comparison to digital channels like a chatbot.

Philip points out that every time the bank contacts a customer, they’re taking a gamble.

But, the group does everything it can to chase so it can be as relevant as possible to the customer.

If the bank is in a purely digital world, it can have a good feedback loop in place which puts the resolved approach in motion.

This means issues get resolved quickly and feedback can be given right away.

All the bank’s sales go through an advisor, so essentially if they didn’t get the approach right, the advisor can fix all the shortcomings and steer in the right direction afterwards.

Since they are expensive, the advisors are only needed for relevant cases.

Philip explains that for other cases, the bank can afford to be hit and miss in a digital channel.

Ideally, the bank would have an advisor available for every customer, however, they can both fight and benefit from having one due to costs but benefit from the customer service it provides.

As it turns out,

Relevance can’t be measured

It can be perceived, but Philip gives us a breakdown on what it includes.

There are 4 main areas we can break it down into.

The first area is the channel.

This looks at our choice of communication: does our choice involve a letter, person, phone call or another digital channel.

There are also factors to consider in this:

Is there a preferred channel by the customer?

Should we respond in the same channel as the customer contacted the bank in, say they sent their message in an email, do we then call the customer? This will possibly give the bank a chance to get feedback.

The second area looks at the content of the message.

This also considers other factors such as tone of voice and how we should talk about our products.

The third area is timing.

The bank needs to consider what it is exactly what the customer is looking for and the timing of communication in regard to this. So, the bank needs to consider what it is they are showing to the customer and when they are showing them this. For example, the bank would question whether they are able to get more product sales on a Saturday morning compared to a Tuesday afternoon say when the customer is at work?

The fourth area is familiarity.

This looks at how well the bank identifies itself with the customer because this gives them more or less leave on how they talk to the customer, how they build a relationship and what subjects are acceptable to talk about.

If the bank wants to build up a level of trust with the customer, they may use a message such as ‘welcome to the bank’.

However, if the customer has been with the bank for years, they won’t want a generic welcome message, so they have to hit the right level of familiarity.

Also, the bank needs to be wary that they aren’t talking about information that isn’t needed, they need to make sure that their product information is being discussed in the right order.


There are issues with the classic use of the ML model

Philip looks at how we’re using the ML model and states that the way we’re using it is perhaps not the best, however it's not the worst since the model is used as a strong algorithm.

The classic use of the model prioritises which customers to contact, the predictive model sorts all the relevant customers to the left.

As suggested, there are issues with the classic approach.

Predicted categories are often too broad.

Advisors are forced to spend excessive time preparing for a call or left searching for the right product to offer the customer.

The classic approach has also been criticised for having a too-short time horizon for its prediction since three months is not sufficient enough to capture the customer early in the decision process.

Leads are only sorted and not filtered, advisors often cherry-pick the customers they know in advance to ensure a positive conversation.

For example,

Say the bank wanted to run a campaign for an investment product, using the classic approach they'd see who'd want a care investment product in 3 months.

They’d then take the top 1000 people and call them, which is essentially a sorting algorithm, there’s no sort of threshold that doesn’t filter in any way.

Using this classic approach, the bank ends up having so many customers compared to how many they actually call, which isn’t the best way to use all the information.

Philip suggests that the bank needs to improve on this.

Since there are some other issues to consider.

Say the investments spoken about in the example above are home finances.

The bank would tell the advisors to talk to these people, within that there’s a myriad of products.

There’re different ways you entail that service and every customer in the bank has their own unique, tailored product, depending on who they are, what they’re rated for and what their needs are. So, this essentially gives the advisor a blank slate to say something with this.

The key is to find advisors and not salespeople, so they are giving genuine advice.

For many machinery problems, you need a long time to predict ahead.

The original prediction period is way too short, if you do say home finance and a customer wants to buy a house, they usually have a year or two to think about the process.

The bank aims to get customers on board as early as possible so that they can help steer the bank.

This means that the customer will end up having a home finance mortgage with the bank, but also receives a good customer experience with the bank.

The last issue is that all leads become available to all devices.

There’s no filtering method, just a sorting method, so the model can go through what ones are most likely to get a good customer experience.

So, they choose customers they already know, essentially carry out collaborative filtering, but they filter the customers based on their own experience and what they’re trying to solve.

This also means that they’d had to evolve the text to include more modern content.

Philip explains that Python is prevalent in many areas of the bank and they are going into Deep Learning (DL) within the next year.

DL hasn’t been relevant in the bank before until now because of noise levels in other stages of the process.

DL couldn’t provide enough for the bank to use and they would rather have a rapid prototype to use in seeing what customers had said, rather than spending a year or two building an architecture of a neural network.

The types of problems the bank is dealing with, can’t do a lot of transfer learning, so that’s why DL has been sitting back.

Philip takes us through four different examples of how they’re trying to ‘break the curse’ of the issues they had previously.

  1. Providing early warning signals for customers leaving early

So, one of the issues the bank had is that identifying people leaving the bank 3 months ahead, was way too short for anyone to do anything.

By that point, the bank had had their last plateau where you have that number one in the image above, which is where most of their churns came from.

But, customers had already started moving their mortgages etc. even when getting back in touch with them, they have no save rate.

The bank could tell they were churning in the next three months, but couldn’t do anything about it, so they needed to extend further back, in order to be able to estimate proxies.

The bank needs to be 3 months ahead of the customer either moving their mortgage or unit number transactions, anything that acts as a predictor.

So, the bank moves the prediction time back, giving themselves 6-9 months of heads up for someone actually leaving them, without having to extend the data.

If the bank were to go more than 3 months back, they’d need a lot more data to actually make their predictions.

  1. Predicting product level rather than category level

The bank moves on from a model based on internal business definition to the customer situation.

They attempt to predict the product most likely to interest the customer.

This leads on to the bank giving an advisor an entry point into the conversation with the customer.

The advisor starts the conversation by asking if the customer has heard about the product.

 It might make more sense to start the conversation this way, rather than straight into finding out if the customer has enough money to manage one of the investment products.

This is because that process usually takes a fair bit of capital to get into, so this way the advisor can actually get into the product level predictions and begin the work in progress.

  1. Predicting first time home buyers

The predictive model catches patterns in the behaviour of previous first-time buyers in Danske Bank and uses these patterns to identify potential first-time homebuyers in the existing customer database.

This aims to capture customers at earlier points in their home-buying journey.

What matters to customers most is getting the product, in the case of this example, their home.

When buying a home, the customer generally thinks about the process of getting a home, explores their options through other websites and decide on which bank they want to get their mortgage with.

For the customer to get into the bank, they start receiving cards and pushing out a few other products, so they can start being educated on their home mortgage.

It is a long process to buying a house.

  1. Breaking down the customer journey

The bank think of the customer's journey as a funnel, where at each step, less and less people move forward.

There’s a significant number of churners which is an issue.

The solution to this is to focus on individual steps and not end-customers' problems.

Advanced analytics can also be used to fix pain points.

By fixing these issues, the bank can hope to drive more customers to success by increasing customer satisfaction.

Most people change banks, which is a prime opportunity and a ripe market to get new customers from.

Danske bank needs to make sure that they predict every step of the way where they can, so they can keep in check with every touchpoint.

If a prediction looks like it's about to go off scale, then extra attention is paid to it.

So the bottom line:

Philip knows that Danske Bank still has room to develop and that their ML models will become deeper as the bank moves forward.

Though advisors are a great asset to the customer service, it’s impossible to use them for every customer.

This is where Philip sees the bank improving the most, as they continue to grow.