Blog: The State of Workplace Safety in the Age of Artificial Intelligence
The State of Workplace Safety in the Age of Artificial Intelligence
According to the National Safety Council (“NSC”), in 2017 the total cost of occupational injuries and deaths in the US was approximately $162B. Of this amount, $51B was attributed to wage and productivity losses, $52B in administrative costs, $46B in medical/property/fire costs, and $12B was related to employers’ uninsured cost (i.e., investigate and document incidents). The NSC also cites research which indicated that a $1 invested in injury prevention returns between $2 and $6 in benefits. And yet many safety professionals face significant challenges from senior leadership when arguing that additional training and equipment can reduce the number and cost of incidents.
In many companies, the current state of environmental, health, and safety (“EHS”) management consists of: (1) predefined steps in the form of job safety assessments which guide employees through the steps, hazards, and recommended actions associated with specific areas of concern; (2) safety and hazard observations that generate incident reports and indicators documenting behaviors and conditions that occur; and (3) employee training programs. These processes often generate retrospective findings that are captured in fragmented systems and unstructured documentation. As a result, its often difficult for EHS professionals and management to review, interpret, and act on insights to improve workplace health and safety.
Given the increasing complexity of business processes and need to increase productivity at an increasingly rapid pace, current approaches are unlikely to detect and proactively detect all incidents. Furthermore, traditional processes are often highly dependent on manual steps and hand-offs between a variety of constituents. Each breakdown increases the likelihood that incidents will occur. One important indicator is the identification of near-miss events. While accurate measures near misses is difficult to determine, conventional theories estimate that for every workplace fatality, there are at least 300 near misses which would mean there were approximately 1.5M near-misses in 2017 (based on Bureau of Labor fatality statistics).
Artificial Intelligence (“AI”) is ideally suited for performing routine tasks that involve performing highly repetitive tasks, analyzing large volumes of structured and unstructured data, and generating predictions that EHS executives and specialists can quickly act on. While a single AI capability such as deep learning or natural language processing can provide significant insights, an integrated architecture of complementary AI-enabled EHS management capabilities can have dramatic impact on improving workplace effectiveness.
To illustrate, consider the following components illustrated in the Exhibit and outlined below.
Each layer of the EHS management architecture is designed to address specific issues in the analytics life cycle. Together, they provide a holistic solution that can enhance organization performance and increase employee safety in ways that traditional methods and technologies could not.
AI Platform Architecture
Data engineering and data science activities require a technical environment with the capabilities that are needed to perform their specific functions. These capabilities often fall into three groups. First, the services layer contains the programs required for executing jobs, distributing the workload, scheduling jobs, monitoring status, and performing other activities. Second, the application layer standardizes the design practices for common AI applications, provides resilience, and supports scalability for AI applications. The third layer contains APIs and user interfaces to enhance usability by stakeholders who desire analytic capabilities with little or no code. While such environments can be on-premises or cloud-based, the latter version can provide significant advantages in terms of scalability, elasticity and cost.
Automated Data Acquisition
Collecting data is often one of the most problematic and time-consuming activities in analytics. For many organizations, EHS-related information is captured in a variety of forms including paper, speech, image, and machine-readable data. Automated data acquisition and machine-aided intelligent data processing is ideal for organizations who deal with advanced instrumentation, energy consumption monitoring, environment monitoring, diagnostic robotics, etc. Fortunately, new developments in technology (e.g., robotics process automation, chatbots, wearable sensors, equipment monitoring devices, drones, and distributed grid monitoring systems). automate this process while increasing speed, lowering cost, and improving human performance.
AI-Enabled Data Preparation
Most of the data that organizations collect is unstructured — data that doesn’t easily conform to an existing data model. As a result, the process of capturing, analyzing, blending, and interpreting data from unstructured sources can be time-consuming and expensive. Reductions in the cost of compute power and advancements in machine learning methods are now enabling organizations to more easily and economically convert unstructured data into a format suitable for analysis. Machine learning can be used to analyze images, audio files, sensor data, emails, social media, and text. As a result, by streamlining the translation, understanding, and preparation of data, analysts and decision makers are better prepared to understand what’s going on in their environment.
Data Scientist Workbench
When incidents occur, EHS managers often find subtle clues in job safety analyses, spot assessments, incident reports, near misses, employee complaints, and other data sources. These sources may contain dozens or even hundreds of attributes which would be impossible for humans to comprehend, let alone identify and pick up on the subtle patterns that may be most significant. Data scientists can choose from, and combine, a variety of machine learning techniques such as deep learning, reinforcement learning, and anomaly detection to expedite the prediction process and deliver meaningful actionable information to decision makers. For example, anomaly detection can be used with deep learning to prevent garbage in — garbage out modeling. Such algorithms can further enhance safety by automatically triggering alerts and other response actions in real-time under human supervision to identify incidents in the early stages and prevent further damage.
EHS Human-Computer Interface
While AI represents an opportunity to accelerate predictions and deliver actionable insights to decision makers (e.g., EHS executives, EHS specialists, front line workers, etc.), it’s important to minimize the complexity associated with user-facing applications and services. The variability of operating environments, job requirements, and user preferences brings challenges and questions that must be addressed in human-centered design. AI-enabled human-computer interfaces can come in a variety of forms including augmented/mixed reality devices, visualization, AI-driven recommendations, facilitated collective intelligence, computer-generated voice interactions, natural language generation, and other capabilities. Failure to consider such factors can inhibit user adoption or even worse, creates a level of complexity that limits a timely intervention and can result in increased risk to people, property, environment and business performance.
Considering the complexity of EHS environments, it’s highly unlikely that AI in its present state will replace human intervention. But by augmenting human performance, AI enables EHS professionals to move more quickly and with greater precision manage risk and control costs while promoting a safer environment for its employees and the public.