British web grocer Ocado Group Plc. receives about 2,000 customer service emails a day on average. That number can easily double or triple during the holiday season or when other issues come up such as bad weather that may delay a customer’s order, says Dan Nelson, head of data for Ocado, No. 23 in the Internet Retailer 2016 Europe 500.
Nelson says the problem until recently is that retailer, which sells solely online and was launched in 2000, used to task its customer service staff with going through and sorting each email that came in.
“A lot of customer service reps’ time was spent filtering emails,” Nelson says. “They would have to sort thank you emails, feedback on delivery or the website and complaints.” Another problem beyond the time spent by employees sorting messages is that emails were handled in the order in which they were received. That means that an email expressing thanks for wonderful service that hit Ocado’s inbox at 9:01 a.m. would be addressed and answered before a 9:02 a.m. message about a critical issue, such a customer complaining that she couldn’t make a purchase, Nelson says.
About six months ago, the retailer’s Ocado Technology division began working on a way to manage emails more efficiently using machine learning and Google Inc.’s TensorFlow, an open source software library for building machine learning frameworks. Machine learning is a process by which computers can learn over time when they are exposed to new data, essentially modifying the computer’s initial programming.
The new system, which went live three weeks ago and is still being tested by Ocado, takes emails as they come into the contact center and determines if they are positive or negative. Next, it assigns a strength of the positive or negative sentiment of the email. Finally, it tags the message with a description of its content, such as a request for website help, a complaint about delivery, a product related issue or a cancel order request. Based on all of that data, each email is immediately prioritized on how quickly it should be read and answered, Nelson says.
In developing the system, Ocado employees used TensorFlow to train the computer system using a backlog of three years of customer service emails Ocado had saved. Ocado then looked at which tags the system applied and examined how closely the assigned tags described the the e-mail’s content and the corresponding decision made by the company’s customer service employees. “We trained the system using our own data and then manipulated it for our own needs,” Nelson says. Ocado also had to filter the email data, stripping certain personal data, such as billing or address information, while leaving enough data for the system to correctly categorize and tag messages. “We had to remove specific names like ‘Daniel Nelson’, but it was important the system recognize that a name had been mentioned,” Nelson says.
Nelson says Ocado looked at out of the box solutions that used natural language search and machine learning but in the end decided to develop the technology in house. “Our emails are unique to our business,” Nelson says. Even though he says some systems could analyze email text to determine the sentiment of a message, Ocado found they couldn’t easily and accurately assign accurate description tags, he says.
Nelson says an in house team of about four data scientists and two software developers worked on the new program. He adds that Ocado likely could have developed the program with fewer data scientists working full time on it, but the retailer wanted the employees to go through the process of developing and honing the machine learning technology to aid them in future projects.
Ocado is now going through the results of how the program has handled emails in the first few weeks since the system went live, so Nelson says he doesn’t have results to share.
Ocado, which posted a 14.7% increase in 2015 web sales to $1.4 billion pounds ($1.74 billion), and has more than 500,000 active customers, has long focused on using advanced technology to boost sales. The retailer has employed for more than eight years a team of data scientists which has also worked on other projects, including the development of a system called Instant Order that will predict a returning shopper’s order based on her purchase history. It also has used data science to improve warehouse operations, Nelson says. For example, the retailer regularly rearranges the location of items in its warehouses based on orders it received that are slated to be picked, packed and shipped out the next day in an effort to get orders out the door more quickly, Nelson says. “A really good example of that is over the Christmas season,” Nelson says. “We look at where the best place is to put product in order to move it more quickly.” Ocado, which uses both machines and humans to pack and ship out items from its warehouses is also working on a project to optimize the routes machines take in its warehouses.
Ocado operates two warehouses, one in Hatfield, in Hertfordshire, UK, which is just north of London and another built in 2012 near Birmingham. It also is currently building two additional warehouses near London.
The Ocado Technology unit of Ocado develops robotics, machine learning, simulation, data science and forecasting and routing, systems for the retailer. It also markets its e-commerce platforms and technology services to other online retailers including U.K.-based supermarket Morrisons, which sells both online and in stores. Ocado says Morrisons is the fourth largest grocer in the UK.