Some federal agencies are beginning to reap the rewards of their investments in artificial intelligence. The Department of Veterans Affairs is prototyping an AI technology that can spot kidney failure days before symptoms appear, allowing patients faster access to life-saving treatments. When Hurricane Florence struck America’s east coast in 2018, the Department of Defense’s new Joint Artificial Intelligence Center used another AI prototype to direct rescuers to victims.
The General Services Administration (GSA) is deploying AI-powered bots across its work. There’s a GSA bot to help new hires navigate the agency. Another bot engaged in financial record-keeping has saved GSA workers some 70,000 hours of labor. A GSA call-center bot has been used by 45,000 people, 75% of whom did not need further assistance.
The Department of Health & Human Services (HHS) has prototyped an AI that offers guidance on when to outsource work to government contractors. The technology, which cost HHS $300,000 to develop, identified $100 million in potential savings.
There are promising signs of inter-agency collaboration on AI, too. One group, the Robotic Process Automation Community of Practice (CoP), has more than 750 executives drawn from over 50 agencies. In January 2020, CoP, which seeks to spread best-practices on robotic process automation across the U.S. government, released its first RPA Program Playbook. “If the government deployed RPA at scale and achieved only 20 hours of workload elimination per employee [per year],” says Gerard Badorrek, the GSA’s Chief Financial Officer and CoP Chair, “the net capacity gained would be worth $3 billion.”
These examples illustrate the incredible compelling promise of AI to raise the productivity of federal workers. Rising investment in AI in the coming years will raise the potential gains higher still. To get the most out of AI, however, more must be done to empower workers and executives to use such technologies effectively.
Modeling AI’s effect on federal workers’ productivity
The following is a snapshot of the steps and assumptions that went into our model.
First, we sought to understand how AI technologies will affect specific work tasks and skills, through automation and augmentation. Here’s how we calculated the shift in labor demand associated with AI:
- We used datasets—from O*Net Database of the U.S. Department of Labor, the International Labour Organization (ILO), and the U.S. Office of Personnel Management—to calculate the total time worked by workers in each country and industry (based on the task frequency of work activities for occupations derived from these datasets).
- Our subject matter experts tagged tasks according to how AI would impact a given task.
- We computed the total potential time susceptible to automation and augmentation for different occupations, based on the frequency with which such occupations perform the analyzed work tasks.
- We measured time savings and productivity gains based on assumptions about investment levels in AI, using elasticity coefficients from regression analysis performed on a panel of 14,000 global companies.
- And we assumed that labor supply matches labor demand at the average unemployment rate of the past five years.
Second, we examined the supply of skills (whether AI can produce value-added growth depends on the capacity of labor supply to satisfy the new demand for skills). Here’s how we calculated labor supply in 2028:
- We used population projections from the United Nations (using its moderate growth scenario), for ages 15–64.
- We used labor participation rates from the ILO (the average of the last five years available).
- And we used unemployment rates from the ILO (average of the last five years available), to calculate potential employment.
Third, we modeled GDP growth for 2018–28 under different investment scenarios:
- We used baseline labor productivity growth by industry and country, sourced from Oxford Economics.
- We multiplied projected labor productivity by employment levels, in 2028, to obtain value-added growth in the baseline scenario.
Fourth, we calculated AI investment levels, at the country and industry levels, using data sourced from the IDC Worldwide Semiannual Artificial Intelligence Systems Spending Guide -2018H2 (published August 2019), with Accenture Research estimates extending the forecasts to 2028.
Fifth, to create the role clusters shown in Figure 3, we conducted the following analysis:
- We employed principal component factor analysis to analyze the skills, abilities, and work activities based on the above datasets. This generated six distinct factors for skills and abilities, and five distinct factors for work activities.
- The importance of each of these factors was used to tag occupations into six groups for skills/abilities and five groups for work activities.
- We cross-referenced the groups against one another to identify the clusters that grouped at least 35% of workers within a skill/ability group. The result was 10 such groups, which represent our 10 clusters.
- We used these 10 clusters to categorize the workforce composition of 14 G20 countries and the U.S. government. We did this by creating conversion tables that matched each occupation code to U.S. occupation codes.
- Under the assumption that the same occupation utilizes similar skills and performs similar tasks across countries, the categorization of occupations within clusters was then applied to other countries and to the U.S. government.