Improving customer communications with data science at Ocado
A lot of the discussion around machine learning is very theoretical, with thousands of research papers published on the topic per month. Marcin Druzkowski, senior data scientist at Ocado told the audience at DataIQ Future, how he used machine learning to solve a tangible problem for his company – speeding up communication between customers and the contact centre.
Online supermarket Ocado is all about saving time for its customers but found that there was often a delay in its contact centre staff responding to customers emails, one of only two points of contact the retailer has with its customers. This was a problem that senior data scientist, Marcin Druzkowski and his team set were tasked with solving.
When Ocado’s contact centre manager approached Druzkowski for help, he responded by bedding into the communications hub so that he could properly understand what the problem was, where the model would be applied and how the data would be measured.
“For one whole day, I became a proper contact centre advisor with my own headphones. I listened to customer feedback and that showed me how the whole environment looks,” said Druzkowski. He realised that with the volume of emails received, it was difficult for the staff to know which to prioritise for a speedy response. For example, a customer may have written in to say that she loves the website but at first glance the contact centre advisors would not know if the content of her email was a complaint or praise.
“The business goal for the project was pretty simple. We would like to have the situation where the urgent emails are responded to very, very quickly and low importance emails can be handled after a little longer,” he said. He added that an objective was essential because having a proper and very clear business goal is crucial to success in machine learning.
The first thing he and his team had to do was clean the data as the emails came in with a lot of personal identifiable information. “In GDPR times we need to know how to process the data, so we had to find all telephone numbers, all post codes and anonymise the data properly before processing,” he said.
With the data cleaned, they built a simple classifier with a few lines of code, but the accuracy was very poor, so they decided to build a much more sophisticated framework with proper neural networks to analyse the content of the emails better. Fortunately, they had a lot of data to work with.
“We built neural networks on top of our huge data set because we have more than 12 years of conversation history between us and the customers,” Druzkowski said. When a new email comes into the contact centre, the content of it is sent to his team and machine learning - which is deployed in the Google platform - returns a list of appropriate taxonomy. They changed words into vectors in such a way that words with similar meanings will be very close to each other and used sentiment analysis to determine whether the customer was happy.
One in every ten emails is still checked by contact centre staff as Druzkowski said it is important to have people involved at the end of the process - they can improve the model and let the data science team know if they have missed anything. Sarcasm, said to be the highest form of intelligence, is still too sophisticated for machines to identify and so humans are needed in this case to respond appropriately.
An unhappy customer who emailed to say that his birthday cake was damaged, said thank you and that he would have a great birthday. The machine learning algorithm classified that email as positive feedback.
He said that having the staff spot checking the process helps to build a data set that they can monitor but it also builds trust as staff can see that they are essential to the process. “Every organisation should think about how we can augment people with machine learning, so it should be people plus rather than versus machine learning,” he said.
When you finally deploy the model avoid the temptation to crack open the bubbly and celebrate said Druzkowski, because deployment is only the beginning and most data science projects fail just after deployment - usually due to a lack of monitoring. The data science team also has to ensure that the software is tested, code versioned and has full reproducibility.
The benefits of the deploying the system at Ocado were tangible. The urgent emails got a response four times faster. They also saved money as they detected that 7% of emails – such as a spam - were not worth reading.
They could also track customer sentiment over time and noticed a significant change in sentiment over the Christmas period - they pumped this information up to the CRM systems to improve the marketing campaign.
These results saw Ocado Technology scooping the silver award in the Best Innovation in Technology category at the European Contact Centre and Customer Service Awards in late November.
According to the Gartner Hype Cycle 2017 machine learning is at the peak but with results like these it may be riding high for a while still.