Gabriel Straub has been Chief Data Officer at Ocado Technology since 2020. Before Ocado, Gabriel held senior data roles at the BBC, Not on The High Street and Tesco. He has over 10 years’ experience in leading data teams and helping organisations take advantage of data and machine learning.
Gabriel is also a guest lecturer at London Business School and an Independent Commissioner at the UK’s Geospatial Commission. Gabriel was named AI Expert of the Year by Digital Leaders and one of the top people in data by both DataIQ and HotTopics. Gabriel has an MA in Mathematics from Cambridge and an MBA from London Business School.
In our latest interview, Gabriel explains how Ocado Technology’s advanced AI creates a seamless experience for its customers. Online shoppers enjoy an AI-powered storefront tailored to their unique preferences, but behind the scenes Ocado’s AI tech is also a crucial element of automation and logistics, from robotic grocery picking to optimising delivery journeys. This cutting-edge tech, Gabriel explains, has been driving online grocery shopping long before the AI hype reached its peak:
When Open AI launched Chat GPT in November 2022, artificial intelligence suddenly shot to the top of the agenda in every boardroom. LinkedIn was flooded with hot takes about the hundreds of ways that AI would change the way we live, work and interact. Many media outlets portrayed AI as a brand-new technology that had come out of nowhere and taken us all by surprise.
But in the online grocery space, AI isn’t as new as you may think. In fact, at Ocado Technology, it’s a crucial element of the e-commerce, automation and logistics technology that we develop and deploy for retailers all over the world. Shoppers have been experiencing the benefits of AI working behind the scenes for many years without even realising.
Wide-scale and free access to generative AI tools such as ChatGPT is definitely an exciting development. If your job involves content creation or processing large amounts of documents, I’m sure you will have experienced a change in how you do things.
But for the online grocery space, generative AI is just one tool in our toolbox. We use a vast array of AI techniques to solve business problems for our clients. From deep reinforcement learning to train robot arms to pick and pack thousands of different items, to optimisation algorithms which can make sure deliveries arrive on time using the fewest number of vans.
When Ocado was founded, many believed that the vision of building a feasible online grocery service was an almost unsolvable challenge because of its complexity.
Handling tens of thousands of orders that contain over 50 items every day is complex enough as it stands. When you combine this with the fact that the products need to be in and out of the warehouse very fast for freshness, and you need to arrive perfectly on time so that your customer can refrigerate their groceries, the logistics challenge gets bigger and bigger.
Without getting smart with automation, data and AI early on, we were never going to achieve a viable model. We quickly found that technology from companies who specialise in the less complex general merchandise space could not be repurposed for the online grocery use case. The tech simply couldn’t handle the need for managing different temperature regimes and perishable goods. There was no template for what we wanted, and no off-the-shelf product that would help us, so we built the technology ourselves.
A technology-powered online grocery operation creates a lot of data. Shoppers log millions of data points as they interact with the webshop, and every shipment that comes in from the supply chain, every movement of goods around the warehouse, and every move our logistics vehicles make create flows of valuable information.
We use artificial intelligence to find patterns, insights and optimisations in this data that allow us to improve the customer experience and make efficiency gains.
Using large centralised fulfilment centres that store products in giant grid structures, means that the retailers we work with don’t have the same capacity constraints that a supermarket store has. This means they can offer a much wider range of products to their customers. For example, Ocado.com in the UK has a range of around 50,000 items. This means, we need to make it simple to navigate that large range for customers. We don’t want shoppers to ‘get lost’ in the virtual aisles.
On average, 70% of shoppers abandon their shop after adding products to their baskets (source: baymard.com/lists/cart-abandonment-rate). AI is critical to overcoming the friction that leads shoppers to abandon. We need to make their lives easy.
For example, many shoppers use our AI-based ‘instant shop’ that automatically populates a whole order with items based on their data and history. It means a shopper can go from login to checkout in just a few clicks.
We also use AI to inspire shoppers with ML features that recreate the fun of spontaneously grabbing something new in the supermarket with personalised recommendations.
An example of this is flash sales, where AI technology triggers price drop sales of products with a risk of purge advertised to shoppers. This not only helps us reduce waste, but helps inspire our customers to try (and continue with!) new products. It turns out new-to-you products bought through flash sales have a significant chance of being repurchased in following shops so clearly it’s a proposition that shoppers benefit from for inspiration.
We also use AI in our search function, where our models are designed to improve the accuracy of matching, which massively increases the regularity of preferred brands and products appearing at the top of customers’ search results. AI models can suggest search terms that closely align with a customer’s search history, enabling them to find what they’re looking for without needing to type a letter.
Thanks to AI, every single customer has a different experience and interacts with a unique storefront tailored to their needs.
Customers want their shop to arrive in full, at the right time and all for a competitive price. One of the consumer frustrations with online grocery shopping is that when their order arrives, the retailer has made substitutions that they are not happy with.
For us, this isn’t an acceptable customer experience. We set up data flows through our platform so that we know exactly what’s in our warehouse, when new stock will be arriving, and which supplier orders can still be modified. That means we only show customers products on the webshop that will be in stock at the time that their order is prepared so we rarely need to make substitutions.
But we can only offer a top-level service if the model is operationally efficient. AI plays a crucial role in eliminating waste and driving down cost.
Deep learning models are able to forecast what customers will buy. Retailers have found that our deep learning models are up to 50% more accurate than their traditional forecasting systems. The AI recommends the optimal amount of stock to buy from suppliers to guarantee availability of products for shoppers without creating waste from unsold products. For our partners using our supply chain AI, over 95% of stock is ordered automatically without the need for any manual intervention.
The models benefit from the latest developments in transformer models, which are at the heart of the current Iarge language model and generative AI phenomenon. This innovation means that in the UK, only around 0.7% of stock we order goes unsold compared to an industry average of 2-3%.
But once we have inbounded our stock from the supply chain, we need to get it in and out of the warehouse in the most efficient manner.
In our automated customer fulfilment centres that we provide for retailers worldwide, hundreds of retrieval robots whizz around giant storage grids to fetch containers containing the items that people have ordered.
These ‘bots’ are orchestrated by an AI ‘air traffic control’, which controls every move the bots make. AI helps make decisions like: which stock storage box to use for the next order, how to efficiently path the bots around each other, and where to put boxes down once we’ve finished picking from them.
Compared to a typical manual warehouse, the storage and retrieval bot system and advanced automation reduces the end-to-end labour time to complete one order from over an hour to under 10 mins.
Our fleets of bots travel tens of millions of km every year, and occasional mechanical wear can mean there is a failure. AI plays a crucial role in minimising the impact. If a bot ceases to provide updates to the air traffic control system, we use computer vision on the 360-degree fisheye cameras above the grid to locate the bot for maintenance attention. This makes our engineering operations more efficient.
The bots deposit stock next to robot arms to be picked into customer bags. Our retail partners can stock up to 50,000 different products each. Robot arms need to be able to deal with a huge variety of weight, size, shape, surface, fragility and packaging.
We use the latest AI and machine vision algorithms to solve this challenge. The system takes inputs from the cameras and sensors on the robot and outputs the motor-control instructions to move the robotic arm and the end effector.
We used imitation learning techniques where human pilots demonstrate how picking and packing should be done with a controller and the robot arms learn from this data, generalise, and improve over time.
Now that robotic pick is live in operations, it’s supervised by remote teleoperation crews who can assist the arm in identifying grasp points for unfamiliar products or scenarios. If the AI gets ‘stuck’ it calls for a person to assist remotely. Every time we use teleoperation we gather data which, combined with the information of the failed pick attempt, helps train the AI to handle similar scenarios in the future.
Unlike more traditional courier services, grocery orders cross multiple temperature regimes, take up significant space, and are expected within narrow time windows. As a result, a grocery delivery van may deliver 20 or so orders in a shift, whereas a van delivering parcels might deliver hundreds. This makes the labour cost in the last mile far more significant on a per-order basis.
Minimising delivery costs is therefore vital to achieve the margins we want. Without AI, it would be difficult for us to deliver groceries to your door on time for the right cost.
Firstly, AI helps us minimise the number of bags we use, and therefore vans and routes. 3D packing algorithms mean we can densely fill shopping bags so they fit into a lower number of vans, whilst ensuring bags aren’t too full so fragile items might be damaged.
Our AI delivery optimiser calculates the optimal distribution of all of our orders onto available vans taking into account the order volume, weight, availability of vans and drivers and much more. It also plans the optimum path for all of these vans to take in order to minimise fuel consumption, number of drivers needed, and ensure everything gets delivered on time.
It takes into account traffic conditions, and learns from past drive time and stop time data for every road and every customer address.
Data from vans is used both for the real-time monitoring of the routes and for feeding into our routing systems so that the routes we drive tomorrow will be even better than the ones we drove today – e.g. the best place to park on a Wednesday afternoon during school term time may be very different to the best location on a Sunday morning.
AI is embedded throughout our tech estate. We have over 130 distinct AI use cases across Ocado. Many of our robotics products deal with highly unstructured environments. We can’t pre-programme a robot to deal with the many millions of scenarios it could be faced with. Due to the high number of unknown factors and almost infinitely variable environments, we must have AI systems based on data that can learn, adapt and generalise.
But any AI we use is always augmented with great engineering in other crafts. ML engineers and data scientists work alongside mechanical engineers, controls, mechatronics, electronics and software engineering. Our data team is embedded in our product streams, rather than being a totally separate function.
Despite our wide-scale use of AI, our mantra is to always choose the right tools for the job. We need to remain critical when we use this technology. AI is not the right approach for every product domain. There are many areas where we will always have hardcoded software systems.
We take a similar approach with the latest generative AI tools. Generative AI, or any sophisticated AI approach, can be very exciting, but there is only a small subset of business problems that are best solved by very sophisticated techniques.
More often, the most important thing is good data and an understanding of what it means. It might be very trendy to develop complex algorithms to achieve certain tasks, but often there’s a simpler maths approach which will be just as effective. And likewise, small, faster and cheap-to-run machine learning models beat massive parameter large language models at a lot of tasks.
With so much AI technology embedded across our platform, it’s critical that we develop and deploy these technologies as robustly as possible. We want our partners and their users to feel confident about using our AI and robotics products and services. They want to know we’ve thought about performance, safety and privacy. This is even more important as we scale internationally and support new businesses beyond the online grocery space.
This is why we have committed to design and deploy AI systems responsibly – always considering fairness, accountability, transparency and explicability throughout the AI lifecycle. This will ensure systems function as intended and prevent mistakes, meaning responsible AI is essentially successful AI.
