Hold the “Horses of the Apocolypse”! A recent study from, Neil Thompson, Maja S. Svanberg and Wensu Li courtecy of MIT FutureTech, in partnership with Martin Fleming from The Productivity Institute, and Brian C. Goehring from IBM’s Institute for Business Value called “Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision?” has been released. It provides insightful analysis of the economic feasibility of automating tasks using AI, particularly focusing on computer vision (CV).
It reveals that the economic cost of AI deployment considered as a whole – training, infrastructure, etc. – is not always more cost efficient that human labour capital. Ultimately it attempts to highlight that most “AI will take my job” studies merely parallel what tasks exist in the labour markets versus what capabilities exist in AI platforms. Such studies do not make any additional evaluations of to the likelyhood or speed of these adoptions from the perspective of the deployment of capital.
The study implies that like a good cake, it will take time for the AI to bake into industry on a large scale. As I reviewed this study I came up with 5 fundamental takeaways
1. Understanding the Practical Limits of AI in Business
AI technology, including computer vision, is advancing significantly. But, its practical application in business is not solely determined by its technical capabilities. A crucial gap exists between tasks that are technically feasible to automate and those that are economically viable to automate. This distinction is paramount in understanding the real-world implications of AI on labor and productivity. The study reveals that although a significant portion of tasks might be technically automatable, only a fraction of these are economically attractive for businesses to automate. This finding challenges the common narrative of an imminent, widespread job displacement due to AI. This suggests, instead, a more gradual, nuanced integration of AI in the workplace.
This point is illustrated by the example of a small bakery. One task that bakers perform is to visually confirm the freshness of ingredients; a task that could be handled by CV. The bakery has five bakers, each earning $48,000 per year and freshness checks consume 6% of their daily task set. This represents a potential has potential labor savings of $14,000. However, this amount is significantly less than the cost of developing, deploying, and maintaining a computer vision system, making it uneconomical to replace human labor with CV in this scenario
2. Balancing Technical Feasibility and Economic Viability
The decision to automate a task using AI hinges not just on technical feasibility but significantly on cost-effectiveness. The paper’s analysis underscores the importance of considering both the upfront and ongoing costs associated with AI systems.
Factors like the cost of data, training, implementation, and maintenance play a pivotal role in determining whether automating a particular task makes economic sense. Interestingly, the study finds that only a small percentage of the tasks “classically” identified as being exposed to AI displacement, would actually be cost-effective to automate at todays costs. This highlights the need for businesses to conduct a comprehensive cost-benefit analysis before jumping on the AI bandwagon.
The estimate indicated that only 23% of the worker wages paid for vision tasks in U.S. businesses would be attractive to automate. Despite technical feasibility, the majority of tasks are not, yet, economically viable to automate from a holistic all cost perspective.
3. Economic Attractiveness of AI Varies Widely Across Tasks
The economic attractiveness of automating tasks with AI varies considerably across different tasks categories and industries. Tasks involving higher wages, less variety in tasks, and large-scale operations are more likely to be economically viable for AI automation. Conversely, tasks in lower-wage roles, involving diverse activities, or in smaller firms may not offer the same economic incentive for automation. The impact of AI on the job market will be uneven, potentially creating pockets of disruption but not an across-the-board displacement of workers.
4. Mitigating AI’s Impact Through Proactive Measures
This economic reality would promote a gradual pace of AI adoption, offering a window of opportunity for policymakers and educators. There is room for proactive measures, such as retraining programs and policy interventions, to prepare the workforce for an AI-augmented future. The study’s findings advocate for a strategic approach to AI integration in the labor market. Emphasis on equipping workers with new skills and adapting policy frameworks to support a smooth transition.
While 36% of jobs in U.S. non-farm businesses have at least one task that is exposed to computer vision, only 8% of these jobs have at least one task that is economically attractive to automate. This indicates a significant gap where policy and retraining could play a role in facilitating AI integration
5. Future of AI Adoption: Cost and Scale as Key Drivers
The future trajectory of AI adoption in businesses will likely be influenced by two primary factors. First, the cost of AI deployment itself. Secondly the scale at which these deployments are made. A significant reduction in AI system costs or adoption of AI-as-a-service models would accelerate AI integration in various tasks. This would alter the cost-benefit analysis for many businesses, potentially expanding the range of tasks that are economically viable to automate. As such, the landscape of AI in business is dynamic, with technological advancements and economic factors continually reshaping the boundaries of automation.
In a hypothetical bare-bones implementation scenario, where data and compute costs are assumed to be free and only minimal engineering effort is required, the amount of economically attractive firm-level automation only increases to 49% of jobs (0.79% of total compensation). Vast fragmentation of tasks across jobs can make even modest implementation budgets prohibitive.
Similiar to my article regarding Balance of Perspective this study showcases a better understanding of the “on-the-street” reality. An interesting study that attempts to objectively evaluate the hype surrounding current and emerging AI technologies against the realities of economic capital deployment, massive variety of worker tasks and the issues of large versus small scale business.
Colin Ryan
ChatGPT was used for editing of this document and images were created with MidJourney.