“…the ability to control the whole flow of products from the warehouse to the end customer…that gives us a lot of ability not only to control the flow of the product, but also flow of information…I think you’ll see it, too, as a customer where you’re starting to get more precise estimates of delivery. You’ll get notes that say, hey, you’re eight stops away from your delivery, et cetera….we see a lot of benefit from that…because we pretty much have perfect information between the order placement allocation to warehouses where we’re going to pick and box up the product and send it on its way. So lots of advantages. We are continuing to invest, and we’ll see a large investment in this area through 2021 as well.” – Amazon.com CFO post Q1 results
‘With employees at Amazon’s fulfilment centres protesting poor working conditions during the pandemic, the tailwind for automation hardware looks strong – better conditions/wages are a pragmatic interim measure. The global market for warehouse/logistics automation was $53bn last year but could double by mid-decade with retailers following the Amazon/Ocado lead post pandemic. Amazon already has more than 200,000 mobile robots working in its warehouses, mostly to deliver products from shelves to workers packing items in boxes for shipment, eliminating the need for workers to leave the conveyor system. The next stage of industrial automation for Amazon involves robotic systems that pick and pack items, combining AI software, computer vision sensors and suction/grasping arms. As for Uber, humans are the weakest link in the retail business model, and further automating their tasks will be a priority after this, not only for Amazon but across the grocery sector, much of which still relies on workers picking goods off retail outlet shelves to put in a van. That worked (at a high margin cost) in the pandemic to allow rapid scaling of grocery deliveries by limiting store opening hours to allow them to become part-time fulfilment centres. However, once this passes every food retailer board will be forced to face up to the classic incumbent ‘creative destruction’ dilemma and reinvent their online business models.’ Weekly Insight, 7th May 2020
‘…COVID may be accelerating a profound change that gets robotics to the upside of that J-curve. Plug-and-play” systems are much more mundane than the advanced robots that dominate the headlines. Most don’t even rely on artificial intelligence. But they embody sophisticated new automation that can be put to practical use with relatively modest preparation and disruption. Plug-and-play systems include automated guided vehicles, which use laser-based LIDAR sensors to navigate around factory or warehouse floors. They also include computer-controlled conveyors and sorting machines and even automated baggers. Brynjolfsson and Beane say the surging demand for plug-and-play systems signals a potentially profound shift toward robots that are highly sophisticated on the inside but appear simple on the outside. As robot manufacturers catch on to the rising demand for plug-and-play robots, spurred in part by the COVID pandemic, they will come closer to delivering the real promise of robot productivity. “There’s an instructive lesson here from the introduction of electricity,” Brynjolfsson says. “When factories became electrified [at the start of the 1900s], they didn’t have any significant productivity increase for 30 years. It only happened when they realized that electricity allowed them to change from having one huge steam-powered motor to having smaller motors in everything. Then you had a doubling or tripling of productivity. The same is true today of advanced robotics.” From the Stanford Human Centred AI blog
- We have long been highlighting the prospect of radical service sector automation over the next decade, replicating that in manufacturing from the 1990s onward, a trend with profound political implications. That has been expressed over the past few years in a thematic global stock basket including names like Cognex and Rockwell in the US, Shenzhen Inovance in China, Daifuku, Yaskawa and Keyence in Japan, Delta Electronics in Taiwan etc., comprising makers of various automation vision/motion sensors, components such as actuators and full systems. The long held view has been that automation capex growth would shift from robots in safety cages performing specific functions on factory production lines to ‘cobots’ working alongside humans in fulfilment warehousing and logistics depots.
‘Automation as a Service’ Via Cloud Toolkits Will Be Transformational…
- The ‘automation of consumption’ theme, including the rising use of robots across the logistics and fulfilment sectors has proved a good bet and is still in its early stages, with precise geolocation and developments in self-driving technology converging with robotics and accelerating the uptake of cobots, autonomous delivery vehicles and drones etc. Indeed, because the underlying technologies overlap, following the spate of M&A from commerce, car and robot companies will increasingly merge – Honda and Hyundai (which bought US walking robot maker Boston Dynamics last year) are early examples of ‘systems integration’ among auto OEMs. Amazon and Alibaba now extensively use robots developed in-house to bring the shelves to the pickers, who stand stationary instead of walking around the warehouse, with the average worker picking 3x as many items in the automated system.
- These residual human roles will likely be automated by end decade, as challenges like engineering ‘haptic feedback’ into robot grippers are solved. Food delivery has boomed over the past year but its preparation is now migrating to industrial scale, automated ‘dark kitchens’ that can prepare thousands of meals daily across a range of cuisines. The DoorDash acquisition of food service robotics start-up Chowbotics (which replaces humans in customised salad preparation) is a sign of things to come across the sector, with dozens of companies aiming to automate specific kitchen functions such as flipping burgers in fast food outlets. The company also entered into a delivery automation JV with GM and acquired self-driving start-up Scotty Labs. Delivery robots and even drones are going to become ubiquitous by end decade and transform the economics of order fulfilment – Starship has developed small 45kg pedestrian robots that can carry items within a 6km radius with a cargo bay can only be opened with the recipient’s smartphone and are similar in design to Amazon’s Scout and FedEx’s Roxo same-day delivery robots.
Robotic Process Automation (RPA) Gaining Traction Across Professional Services…
‘Perhaps the most compelling reason for thinking a disruption lies ahead, though, is a realisation on the part of many professionals themselves. Over the past five years, I have heard doctors worry about the implausibility of seeing patients online, been told by teachers that students cannot learn properly unless they are in the same room and listened to lawyers insist that court work could never be done virtually. And yet, in a matter of weeks, telemedicine, online learning and virtual courts have become the norm. The main barrier to transformation in the professions was never technological. Many of these technologies have been around for years. The main barriers were cultural. Most people are resistant to new ways of working, white-collar workers particularly so. With the Covid-19 crisis, those cultural barriers have largely disappeared, leaving many preferences for traditional ways of working looking more like indulgences.’ Daniel Susskind, fellow in economics at Balliol College, Oxford and the author of ‘A World Without Work’
- While the retail supply chain is now rapidly automating, even more profoundly, ‘deep automation’ is now heading to offices. Back in 2003 when AI had largely been abandoned as unworkable, the economists David Autor, Frank Levy and Richard Murnane published a study on technological change noting that machines perform tasks, a narrower unit of work than most human jobs. They classified these task as “routine” or “nonroutine” (constructing a financial model is typically routine, cleaning a hotel room involves a series of non-routine tasks and therefore hard to automate). This approach helped explain why technology could transform the nature of many jobs without eliminating them in large numbers and also why a ‘barbell’ jobs market developed, with low-paid and high-paid jobs resilient while the middle was hollowed out. In his book ‘A World Without Work’, Oxford University economist Daniel Susskind argued that the second part of that analysis is now becoming questionable as reinforcement learning software encroaches upon non-routine skilled tasks.
- He points out that what Google DeepMind’s AI supercomputers can do today will likely be achievable on laptops and even phones using edge computing within 10-15 years. The Artificial Intelligence Index project, based at Stanford University, which tracks a wide variety of benchmarks, highlights rapid progress at symbolic achievements from playing poker to translation, speech recognition, and classifying diseases such as skin cancer and diabetes via image pattern recognition. As machine learning adapts to ever more tasks, the reorganization of human work in response will be wrenching across the cognitive distribution curve. Only fringe US Presidential (and now more plausible NYC mayor) candidate Andrew Yang has focused on this key policy challenge, but it will likely redefine politics over the course of this decade and have profound investment implications.
- Over the next several years, language models will likely become far more general purpose, encompassing an unimaginable range of problem types. Being able to have a world described through language and rendered as an image or video, or even asking text-based questions about the world with answers based on a system’s understanding of our nuanced reality sounds like science fiction but is now within our grasp by mid-decade. Selene is NVIDIA’s own AI supercomputer used for experimental and production operations. These services have been valuable for NVIDIA, but first in entertainment and gaming where text can be used to feed into image or video synthesis (to create worlds from description) and then more broadly to corporate applications, companies globally will have a toolkit to replicate a far broader range of insight previously assumed unique to human intelligence. Language modelling is one of the most expensive done in machine learning – we are at the point of asking the language model to solve a new problem that it’s never been before and it is generating an accurate answer by contextualising, a true (if at this stage, expensive) breakthrough.
- The potential applications are endless – Amazon now has a cloud service called Connect to automate customer service call centers for other companies with natural language chatbots, interactive voice response, and automated customer voice authentication. The service is integrated with Amazon’s Lex AI that powers the automatic speech recognition and natural language understanding in Amazon Alexa, offers a “Wisdom” feature that searches across independent databases like Salesforce and ServiceNow repositories to find answers for customer queries significantly faster than any human agent, analyses customer sentiment in real time and can automatically alert supervisors with issues during calls. It looks like sales of Connect jumped 150% to roughly $175m last year – this looks likely to be a billion-dollar software market within 2-3 years.
- Susskind quoted above makes a good point – the past year of remote, virtual working has been an experiment allowing service companies to reimagine the workplace. Ever more sophisticated AI has changed everything, starting with the definition of “routine” work, which has vastly expanded, making it harder for displaced from low-skilled jobs to retrain for more challenging roles. Indeed, it’s been a long-standing theme that office space per capita would decline (with process automation in insurance, banking etc. the structural tailwind). The full implications of the ‘proof of concept’ trial forced upon us by the pandemic will be profound and already being felt in tech and finance hubs like NY and San Francisco, which risk seeing their tax bases and ability to fund services erode as employees flee to low-income tax states like Florida and Texas. The implementation of AI in the professional service sectors is straightforward wherever tasks are well defined and there is a sufficient library of labelled data available to train a neural net to acquire a useable function e.g., automating standard legal contracts.
- Rapidly advancing computer vision algorithms and neural networks allow the software to “see” desktops and understand relevant documents, using AI to understand patterns in the data and how best to break the process down into replicable steps. RPA start-ups have been building automation toolkits so that IT departments with limited coding experience can input their own processes and get those automated – cloud based ‘automation as a service’ to boost white collar productivity is coming.
- The approach is particularly applicable in any industry with repetitive, heavily rule-based process-oriented tasks, such as insurance or accounting e.g. patient records in healthcare, preparing invoices in financial services, digitizing legal documents, or automating repeat searches (files, priors, data to cite) for insurance claim representatives. Across the service sectors, robotic process automation (RPA) combined with AI software is emerging as one of the fastest growth new enterprise software markets. An RPA tool can be triggered manually or automatically, move or populate data between prescribed locations, document audit trails, conduct calculations, perform actions, and trigger automatic responses. It automates rote tasks, handing them off from humans to software algorithms.
- UiPath raised $1.3bn last month via a SPAC listing and is trading at a $37bn valuation (>10x more than in 2018 and approaching a third of SAP’s market value). Its automation clients range from Google and GE to Equifax, HP, and McDonald’s. UK AIM listed Blue Prism , Kryon and FortressIQ in the US have also raised capital to expand their product ranges while Microsoft acquired Softomotive last year (now part of Microsoft Power Platform). The UIPath approach starts with generating a map of workflows to recommend what to automate, and then moves into the data-gathering process e.g., taking screenshots as an employee goes through a task. It’s on track for an IPO this year – start-up competitors include Automation Anywhere, which raised $290m from investors including SoftBank in 2019.
- In terms of sector automation, mid-back office admin work looks to be where warehousing and fulfilment was five years ago – i.e. on the cusp of being radically reshaped as humans are increasingly left doing only those tasks the machines can’t analyse, model and replicate. Even as we face multiple labour market bottlenecks over the next couple of years in the post-pandemic recovery (notably in blue collar sectors from truck driving to manufacturing and construction, where Millennials seem unwilling to replace retiring Boomers), on a five year plus view a major dislocation looms as the cognitive threshold for a middle class life gets pushed ever higher by highly scalable automation breakthroughs. In that context, Andrew Yang may not have all the answers, but he’s been asking many of the right questions…and portfolios should retain thematic exposure to the secular AI/sensor investment opportunity.