Reinventing the Job to Optimize Work Automation

In a bank teller job, some of the work tasks are very repetitive and require little thinking (counting cash) and are ideal for robotic process automation. Others are highly variable and require a great deal of thinking (collaborating with product designers), where AI could enhance the process. Some involve substituting automation for humans (e.g., verifying account balances), while others augment humans (e.g., recommending that a human do the banking services, but AI can identify the best services and make appropriate recommendations to the human). A routine process (counting and giving cash) can accomplish some tasks, while other tasks are best accomplished by automating the cognition to detect patterns, identify the best option, or make recommendations (predicting customer receptivity to additional services). In yet other areas, automation can create new types of work (e.g., supporting customers remotely as they engage with the bank without ever walking into a branch).

Now you see more clearly how some automation applies better to work elements with certain characteristics and ROIP. You can now more precisely define the cost-quality-risk implications of the combinations of tasks, ROIP, and automation. Table 3-1 shows how you might define the optimal work automation combinations.

You can clearly see a reinvented job that used to be that of a bank teller. As your technologists predicted, a group of tasks is optimally automated by substituting ATMs (process automation) for humans. However, such tasks are only a subset of the reinvented bank teller “job.” For many work elements, human workers remain the optimum solution. For still other tasks, the reinvented work will combine humans and automation.

Let’s now look in depth at how automation has and will further extend into the financial services sector (see figure 3-1).

We explored how ATMs affected work in bank branches. When they first arrived in the 1970s, ATMs were expected to spell doom for bank tellers by taking over some of their routine and repetitive tasks. Indeed, in the United States, the average number of tellers fell from twenty per branch in 1988 to thirteen in 2004 to currently fewer than five in some branches. While this decline reduced the cost of running a bank branch, it allowed banks to open more branches in response to customer demand. The number of urban bank branches rose by 43 percent over the same period, so the total number of tellers actually increased. As we explained in chapter 1, rather than destroying jobs, ATMs changed bank employees’ work mix away from repetitive tasks and toward things like sales and customer service that machines could not do. As we move forward in time, we see automation shifting from a focus on transforming core transactional processes toward the application of intelligence to higher-valued-added activities like trading and analysis. In the future, even more nonrepetitive, highly cognitive work will either substitute for or augment human capability as AI moves from a focus on the “known unknown” (e.g., I know I don’t know what the optimal asset mix should be for someone of my age and risk tolerance) to the “unknown unknown” (e.g., I don’t know that I will need to reallocate my portfolio in the event of an unrelated market event, but AI will anticipate that and execute the transactions needed to achieve the right mix of assets).

As we expand our horizons beyond financial services, what various types of automation are available and how will they develop into the future? (See the sidebar “Why Is AI a Big Deal Now? Convergence.”)

Robotic Process Automation (RPA)

RPA is the simplest and most mature category. RPA automates high-volume, low-complexity, and routine tasks. For example, it has long been used to automate “swivel-chair” tasks that used to require a person to swivel from one data source to another to transfer or connect data from disparate systems. A common application involves transferring data between software systems or using simple rules to find information in emails or spreadsheets and entering it into business systems like enterprise resource planning (ERP) or customer relationship management (CRM). These tasks are often too simple for a complicated IT solution. Instead, simple process robotics can automate them quickly and cheaply, without requiring the management and training of labor. Xchanging, a UK-based insurance claim services company used twenty-seven Blue Prism robots to automate fourteen core processes, performing 120,000 RPA transactions per month and reducing the cost per process by 30 percent.² (See the sidebar “The Three Rs of RPA.”)

A typical RPA algorithm would look something like the following:

Cognitive Automation

The current headlines about work automation reflect cognitive automation, which replaces humans doing nonroutine complex tasks, literally automating human cognition. Cognitive automation uses tools like pattern recognition and language understanding. The retailer Amazon pursued strategic goals that include improving the quality and reducing the cost of customer service in physical stores. That resulted in the Amazon Go retail store in Seattle, which has no cashiers or checkout lanes. Customers shop and go, as sensors and algorithms automatically charge their Amazon account. These strategic goals and operational designs rest on reinventing jobs. Automation does the tasks of scanning purchases and processing payment. This doesn’t mean the end of store associates, but their work changes. Humans still do tasks like advising in-store customers about product features. Cognitive automation in the form of machine learning, using scalable cloud computing resources, has produced systems that can recognize patterns and understand meaning in big data in a human-like way. This recognition intelligence is a combination of artificial intelligence, specifically machine learning, and sensors. It is at the heart of automating tasks like voice and image recognition, voice conversion to text, and natural language understanding.

These applications reflect automation that is taught rules and procedures by humans, but in newer, deep learning, the machines teach themselves. This automation is applied to increasingly more diverse, abstract, and advanced tasks. The Google DeepMind team created a computer called AlphaGo that famously defeated master players at the complicated game of Go. To train AlphaGo, DeepMind fed the system thousands of games that amateur and professional human Go players had played. AlphaGo used the games to develop winning strategies and identify good and bad moves. More recently, the same DeepMind team created AlphaGo Zero, a computer that played only by itself (millions of times), at first making moves at random until it recognized strategies. AlphaGo Zero got its name because it had zero help from humans beyond starting it up. AlphaGo Zero defeated not only human players, but ultimately even its predecessor, AlphaGo.³

Cognitive automation is typically used in three ways. First, to transform business processes, such as car insurers that use an app with image recognition and cognitive analysis capability to process photos of a damaged car, assess the damage, estimate the size of the claim, and send its recommendation to a human assessor for final approval, creating a simpler, faster, and cheaper claims process. This reinvents the former job of human field inspectors to a job of remote, high-level approvers and assessors. Such technology allows traditional jobs to be deconstructed and to augment or replace routine human activities with automation, resulting in the work being reinvented for greater efficiency, effectiveness, and impact.

Second, cognitive automation can develop new products and services. The same automation that reinvented the claims process enables a new service offering to car insurance clients, with features such as a chatbot that provides on-demand advice about repairs and payments to policy owners, right on their phone. Now, the jobs of customer service associates can and must be reinvented.

Third, cognitive automation can gain new insights with big data. In the auto insurance example, cognitive automation can analyze thousands of claims to identify locations most prone to accidents and compute customer premiums that adjust for driving in high-risk versus low-risk locations. Now, jobs such as data scientists and analysts can and must be reinvented.

You can see how convergence creates exponential automation opportunities. RPA is often a precursor to using AI, where RPA produces the necessary, clean, high-volume data needed to drive effective cognitive automation. Consider the previous RPA illustration with cognitive automation inserted:

This same convergence applies to work, creating similarly exponential opportunities and requirements to reinvent the work and its organization. Ultimately, optimizing work automation is an opportunity to consider an entire ecosystem of jobs and their relationships, deconstruct them, assess the ROIP of the work, apply RPA and cognitive automation, and then reinvent them. (See the sidebar “Uptake Keeps Trains Running Using Cognitive Automation.”)

Social or Collaborative Robotics

You may think of robots as machines bolted to the floor of an assembly line, performing one repetitive task. That’s still true, but increasingly giving way to social robotics. The word “social” refers to robots that move around and interact with people, using sensors, AI, and mechanical machinery. A subset of social robotics is “collaborative” robotics (cobots). Cobots are machines that actually sense the human worker and actively adjust to physically work with the human.

Baxter is a cobot that performs a wide range of assembly-line tasks, including things like line loading, machine tending, packaging, and material handling. Strategically, organizations acquire and deploy Baxter cobots to achieve these strategic goals:

• Safety: Baxter operates safely near humans, without needing a cage, saving money and floor space.
• Trainability: Baxter learns by watching the movements of human workers, reducing or eliminating the time and cost of traditional programming.
• Redeployability and flexibility: Baxter can execute a range of tasks and, because it is trainable, can be repurposed quickly to other tasks.
• Easy integration: Baxter connects with other automation on the line, often without any third-party integration programming or design.
• Compatibility: Baxter’s arms move like human arms, so assembly lines built for humans don’t have to be reconfigured to work with it.

Baxter isn’t the only social robot design. Social robots increasingly come in the form of drones that fly or swim, anthropoid robots that walk, and swarm robots that roll. Traditional robots were mostly limited to very routine and repetitive tasks, but social robots now automate both routine and nonroutine tasks. Freed from the assembly line, such robots can collaborate with humans in ways that were unthinkable before.

Swarming robots are reinventing warehouse operations and shipping at Deutsche Post AG’s DHL facility in Memphis, Tennessee, for the provider Quiet Logistics as it fulfills online orders for retailers like Bonobos and Inditex SA Zara.⁴ The strategic goals behind these applications of cobots is to reduce the cost of million-dollar fixed conveyor belts and warehouse transport systems. Cobots cost much less—$30,000 to $40,000.

Farmers and Allstate Insurance had the strategic goal to use automation to speed up their response to Hurricane Harvey victims.⁵ They used drones to reinvent the work of claims analysis and payment. Collaborative drones worked beside human claims adjusters to assess property damage. The drones accessed areas that humans could never reach or that were very dangerous. The drones gathered data and took pictures of the damage and sent it to a database. Claims adjusters no longer did the dangerous work of accessing damage sites and gathering data. Instead, they analyzed the database produced by their drone collaborators and made faster claims decisions. Farmers reports that the reinvented job combining a drone and a human claims adjuster can process three houses in an hour. Previously, human adjusters took a full day to process the same three houses. (See the sidebar “How Automation Is Evolving.”)

Reviewing Patient Information

RPA can integrate the diverse array of information about the patient that resides in different information systems. It integrates the patient’s biomarkers, medical history, lifestyle, previous treatments, and so on, creating a comprehensive view of the patient that was previously impossible. New patient information is integrated as soon as it is generated, transforming static information into a dynamic evolving snapshot of the patient.

Diagnosing Cancer

There is no intelligence associated with RPA. Adding cognitive automation with natural language processing, the automated system can now read this evolving data. It can compare each patient to thousands of other patients and score the patient’s cancer risk.

IBM Watson for Oncology (WFO) is a cognitive automation platform that achieved 90 percent success in diagnosing lung cancer.⁷ Human oncologists average about 50 percent success. Watson ingests over 600,000 items of medical evidence, reads over 2 million pages from medical journals, and searches up to 1.5 million patient records. Its knowledge far exceeds anything humanly possible. Memorial Sloan Kettering Cancer Center estimates that for human doctors, trial-based evidence makes up only 20 percent of the knowledge they use to diagnose patients and choose treatments. A human doctor would need to spend at least 160 hours a week reading journals, just to be aware of new medical knowledge as it’s published. WFO can much more quickly and accurately assimilate the vast amounts of new evidence added to the global cancer database and update its algorithms accordingly.

Evaluating and Choosing Treatments

Oncologists should evaluate and choose cancer treatments using the most recent practice and evidence-based guidelines. Can automation match humans in choosing treatments? WFO demonstrated very similar recommendations to a panel of oncologists in a double-blind study of lung, breast, and colorectal cancer.⁸ How? WFO extracts and assesses large amounts of structured and unstructured data from medical records, using natural language processing and machine learning to evaluate and choose among cancer treatment options. Approximately 90 percent of WFO’s recommendations agreed with those of a tumor board of fifteen oncologists. At first, it took the oncologists an average of twenty minutes to capture and analyze the data and produce recommendations, although they improved to twelve minutes with practice. WFO took forty seconds.

As WFO improves its capabilities, the job of diagnostic oncologist is reinvented. WFO analyzes typical and common cases, allowing human oncologists to focus on the unusual or difficult cases.

Executing the Selected Treatment or Surgery

The earlier vignette about Intuitive Surgical’s da Vinci Xi showed how cutting-edge collaborative robotics can actually enhance the act of surgery. Still, humans do most surgery, and most of the advances have been focused on making surgery as minimally invasive as possible. When performing robotic surgery with the da Vinci Xi—the world’s most advanced surgical robot—the surgeon controls miniaturized instruments that are mounted on three separate robotic arms, allowing the surgeon maximum range of motion and precision. The da Vinci’s fourth arm contains a magnified, high-definition 3-D camera that guides the surgeon during the procedure. In other words, the machine has no intelligence beyond that of its operator. It does not meet the previously defined criteria of social robotics because it has no AI or sensors.

AI is being incorporated into surgical procedures with technology like the Smart Tissue Autonomous Robot (STAR).⁹ It uses its own vision, tools, and intelligence to execute surgical procedures. STAR actually exceeded the performance of human surgeons. Researchers programmed STAR to do intestinal anastomosis, in which a surgically cut portion of intestine is stitched back together. In only 40 percent of the trials did human surgeons need to intervene with guidance.

This is a good example of optimizing the human-automation combination. The researchers concluded that the 40 percent of trials requiring human assistance offer clues to designing a new job that involves shared human-machine collaboration in operating rooms. The job of surgeons would now be to supervise procedures and hand over the correct tasks to the robot. In the new job, the automation executes and learns more routine or tedious tasks, leaving the human to focus on the more complex and unusual ones. Automation augments human capability.

Choosing nonsurgical or postoperative procedures is equally as important as choosing and executing surgical procedures. What role can automation play in these procedures? Patients with similar cancers can respond completely differently to the same treatment. These differences can often be predicted based on patient genetics. More precise and personalized treatment requires identifying which genetic factors predict remission or resistance. Does automation have a role here? A team of researchers fed AI the genetic data of cells and tumor tissue from breast cancer patients. The AI algorithms predicted that 84 percent of the patients would go into remission using the drug Paclitaxel. The genetic signature of the drug gemcitabine was able to predict remission using preserved tumor tissue with 62 percent to 71 percent accuracy.¹⁰

Coordinating Treatment with the Team

The work of coordinating care between the various members of the oncology team is vitally important to the care of the patient. While IT can help ensure consistent access to information and facilitate interaction between the oncologist and the other roles responsible for treating a patient and RPA can replace many activities associated with integrating data from multiple systems, the true value in this particular activity is the result of personal interactions between the various participants. In other words, the work of coordination shifts from one of gathering, reviewing, ingesting, and discussing data to one of insight-driven collaboration where the work is about brainstorming, exploring, and asking “what if?” Cognitive automation can augment such collaboration by applying intelligence to the data about the patient and helping each stakeholder better understand the unique implications of their work for each other and the patient. AI algorithms are able to run countless simulations from different combinations of activities to project and predict various outcomes that may result. This intelligence can then help shift and optimize the behaviors of the various team members toward those that actually can yield a better outcome for the patient.

Conducting Postsurgery Monitoring, Care, and Counseling

A great deal of postsurgical care requires tasks using the sort of empathy and emotion that no machine can do, but cognitive automation still plays a significant role. Cognitive automation, supported by RPA-created data, gathers and analyzes patient data. Caregivers can use these insights to know how different treatments work on patients with certain genetic makeups. The caregiving staff and physicians can now deliver personalized care that increases rates of patient recovery and reduces complications. Automating the routine tasks of an oncologist’s job reinvents it to focus on the empathetic and emotional tasks that humans do best, and are vitally important in patient recovery. Humans assisted by the AI insights also are more precise in prescribing drug treatments.

Table 3-3 summarizes our description. Notice how several types of automation converge on the work of cancer treatment. Automation reinvents individual jobs, but also creates opportunities to systematically reinvent the relationships among jobs. Automation both replaces and augments the tasks of oncology and cancer treatment. The goal shifts from executing a process or subprocess to addressing the original strategic goals.

As the half-life of skills continues to shrink, the growing premium on reskilling is causing many organizations to rethink the risks associated with full-time employment in order to reduce the risk of obsolescence. The different variations of work-task automation like the ones here can deliver viable solutions to all of the concerns we’ve discussed. Selecting the right technology for automating work tasks and improving performance is therefore critical for business, as is the alignment of the selected technology with a comprehensive future of work strategy. Recognizing how technology and AI can transform the performance and value equation provides a significant competitive advantage. Successful leaders will be able to translate the evolving pivot points in their business models into specific implications for work, looking beyond jobs, and to understand the transformative role AI can play in redefining the performance curve for the work of the future.

We will now combine deconstruction, ROIP, and automation into a playbook for optimizing work.