And it’s been successful. So successful it now sells its Ocado Smart Platform technology to other online grocers including U.K.-based Morrisons, Canada’s Sobey’s Inc. and Groupe Casino in France. The retailer plans to grow its Ocado technology team to 1,200 over the next year, which would mean the unit would account for nearly 10% of the retailer’s total workforce of 13,000, says Greg Cempla, general manager of Ocado Technology, the unit that develops and sells robotics, machine learning, simulation, data science, forecasting and routing systems for the retailer.
Last week, Ocado revealed metrics about the technology it has developed for its highly automated warehouse in Andover in Hampshire, U.K.
The fulfillment center features a three-story-high aluminum grid containing stacks of white storage crates of grocery items. Whizzing on top of the grid is a fleet of robots that picks up crates from each stack and delivers them to pick stations where personal shoppers assemble the customer orders. The robots were designed in house by Ocado Engineering and Ocado Technology.
The retailer also says:
- An average order of 50 items takes five minutes for the fulfillment center robots and employees to pick and pack.
- The warehouse is the size of three professional soccer fields.
- Each robot travels between 30 and 40 miles per day
- Each day, the robot fleet travels the equivalent of 4.5 times around the planet
Ocado says it picks and packs hundreds of thousands of orders a week and delivers them with 98.9% accuracy. The retailer carries 50,000 SKUs. It has more than 645,000 customers and processes 280,000 orders per week. Its average order value is nearly $140.
Such a vast base of sales, orders and customers means that even the smallest efficiency-oriented tweaks can have a great influence on Ocado’s bottom line, Cempla says.
“It’s the economy of scale,” he says. “A tiny savings on the cost of processing an order can have a significant impact on your bottom line when you have hundreds of millions in sales. We are always looking for the biggest bang we can get for our buck.”
For instance, on the technology team’s most recent programs, launched early this year, is a fraud detection and prevention machine-learning algorithm built using Google’s TensorFlow, an open-source software library for building machine-learning frameworks.
The new fraud model collects data from Ocado’s order management system, payments, customer-relationship management system and e-commerce teams to predict if an order is fraudulent or not. Its algorithm also adapts over time based on the data it gathers, Cempla says.
For example, the machine-learning model consumes data from the retailer’s contact center using tools developed by its data engineering, data platform and machine-learning services teams. If an order is incorrectly flagged as fraudulent and the customer calls to complain that it was legitimate, that complaint and the details of the order are noted and used to improve fraud forecasting.
It also helps in other ways. If a customer emails or calls the call center to say she will not be home to receive her order and wants to cancel it, the machine-learning system can take that information and automatically cancel the order, making that cheese or eggs in the order available for other shoppers to buy and saving a warehouse worker time picking and packing.
Ocado tested the program against the manual fraud detection program it had been using and found it to be 15-times more accurate at correctly spotting fraud than an employee. “Analyzing orders for fraud can be very tiring for an employee,” Cempla says. “Doing it for a long time increases the likelihood of a mistake. Machine learning guarantees the same level accuracy applies to all orders.”
The Ocado tech team also spent six months building a system using machine learning to better manage emails sent to its contact center. The system takes email as it comes into the center and determines whether the email has a positive or negative sentiment and 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 that data, each email is immediately prioritized on how quickly it should be read and answered.
Before implementing the system, the retailer’s customer service staff had to scan and sort each email that came in. Moreover, the previous system didn’t have a way to prioritize emails, which meant that messages were handled in the order they were received. And so an urgent message such as “I can’t place my order” might be read after 50 messages asking the retailer to add a new product line.
The new email system saves Ocado 100,000 pounds ($144,233) a year, Ocado says. And, after using the system, Ocado found that 7% of emails don’t even need to be answered. Additionally, response times for non-urgent emails decreased from 19 hours in 2016 to nine hours in 2017. For urgent emails, the average response time is now two hours. Additionally, with the new system, Ocado now answers 95% of emails in 24 hours, an increase from 74% in 2016, the retailer says.
But developing such helpful, cost-saving technology can be tedious. For example, in building its machine-learning-based email system, Ocado employees had to train the computer system using a backlog of three years of customer service emails the retailer had saved. Ocado then looked at which tags the system applied and examined how closely the assigned tags described the email’s content and the corresponding decision made by the company’s customer service employees.
It took Ocado’s in-house team of four data scientists and two software developers six months just to get to the testing phase.
Other noteworthy internal tech projects include a skill for Amazon’s voice-activated Alexa software that can determine, based on a shopper’s order history, when a customer says “Alexa, add milk to my Ocado order,” she probably means add two liters of semi-skim Alpo-brand milk.
To encourage and foster new technological advancements, Ocado hosts a program called hackdays, where data scientists and software engineers explore new features, test new models and analyze and visualize new data.