The analysis of job and task change was completed by breaking down the Australian economy into ~3000 occupations and ~2000 tasks.

The primary data source used in this analysis was the O*NET Database (US). The O*NET database contains survey ratings for a range of job-specific tasks for over 1000 occupations. The study used the 2018 and 2023 databases. Results were then converted to Australian occupations and timeshares. For lower-level occupations where data was not available (for example specialisations), results from higher level occupation groupings were used. There were 4 main transformations of the data.

The first step was to convert the frequency of performing a task to the amount of time spent on a task.

• The O*NET database contains a distribution of survey responses on the frequency of tasks performed by occupation on a scale of 1-7, representing ‘yearly or less’ to ‘hourly or more’.
• To convert each task by occupation into a single score, a weighted average of the survey responses was calculated and rounded to the nearest whole number.
• These scores were then converted into the implied number of times the task is completed within a 40-hour, 5 day work week. For example, ‘daily’ assumes the task is completed 5 times a week.
• Each O*NET task was mapped to an O*NET Detailed Work Activity (DWA) using the O*NET Database correspondence between tasks and detailed work activities. There are approximately 30 different detailed work activities for each occupation.
• For an occupation, a task can be assigned to more than one DWA. In these cases, the task frequency was allocated to each DWA it mapped to. DWAs can also map to multiple tasks. In this case, the frequencies were aggregated together.
• Once we have the frequency of performing tasks (DWAs) we need to estimate the time it takes to complete it. The time taken to complete a task can be formulated as the product between the frequency of performing it and the time it takes to complete it once. The sum of the time spent across all tasks for an occupation will equal the total time spent by a worker in their job. It is assumed that all workers surveyed work 40 hours a week, and that within a 3-digit ONET occupation group the time taken to perform a task once is the same.
• The above formulation gives us a system of linear equations across all occupations groups subject to the constraint that the time spent to a task once is non-negative. Least norm programming is employed to solve this system of equations and determine the time taken to perform a task once.
• The time taken to perform a task once is then multiplied by the frequency with which it is performed to determine the total time spent on a task for a given occupation. For occupations where sample sizes were low, averages from higher-level occupation groupings were used to calculate the time spent on a task. This data is used to calculate the timeshares for each task performed in an occupation.

The second step was to map ‘Standard Occupational Classification’ (SOC) to Australian and New Zealand Standard Classifications of Occupations (ANZSCO).

• Each ONET SOC was mapped to the single closest ANZSCO 6-digit occupation, based on Jobs and Skills Australia's Labour Market Insights correspondence table. These 6-digit occupations were then mapped into ANZSCO 4-digit occupations, by taking the median occupation by task change for each task within the occupation. This step has the added benefit that it can reduce the bias from the approximate methods employed in step 1.

The third analytical step was to calculate the change in the time spent on the same task between 2018 to 2023 for each occupation.

• The change in timeshares was calculated by taking the difference in timeshares between 2018 and 2023. Appropriate ceilings and floors were applied to account for small sample sizes and related outliers.

Finally, the projected change in timeshares from 2023 to 2035 was calculated.

• Timeshares for each task, in each occupation, were extrapolated linearly out to 2035 using the annual change in timeshares from the last 5 years.
• The timeshares were converted back into weekly hours by multiplying by 40.

To calculate wage growth, employment change and skills levels across regions, the team analysed Australian Census 2011, 2016 and 2021 data from the Australian Bureau of Statistics (ABS). The data was gathered using the ABS TableBuilder tool. Each metric was validated by comparing the distribution to similar ABS statistics.

Wage Growth

• The Total Personal Income (weekly) field in the census data was used to calculate average annual wage growth by occupation between 2011 and 2021.
• The data was filtered to only include only full-time workers. This is used to calculate the full-time equivalent weekly income per occupation and reduce variability caused by part-time workers.
• Census data reports the number of workers within defined income ranges. For each income range, the median income level was used. The mean income level per occupation was then calculated using a weighted average of each income group.
• The average annual wage growth per occupation was calculated between 2011 and 2021 using a 10 year compound annual growth rate.
• Workers within ABS Supplementary occupation codes are included in the national annual growth rate, but were removed from the final occupation outputs due to variability.
• The distribution of the final figures compared well with other ABS wage growth statistics.

Employment

• The standard ABS metric for employment was used to calculate the annual change in employment levels by occupation and region. This field includes the number of persons who were employed and worked full-time or part-time, and those who were employed but away from work.
• The data was filtered for both occupation and Greater Capital City Statistical Area (GCCSA) (Place of Usual Residence). Greater Capital City Statistical Areas were used to distinguish regional and rural employment trends from urban employment trends, while maintaining adequate sample sizes for statistically significant analysis. Using a larger regional area also has the added benefit of reducing possible variation between Place of Usual Residence and Place of Work data.
• The number of persons in each occupation, per Greater Capital City Statistical was calculated for 2016 and 2021.
• The annual change in employment was calculated using a 5-year compound annual growth rate. This was completed for each occupation and Greater Capital City Statistical Area. The employment growth for all occupations within each Greater Capital City Statistical Area as well as the national occupation employment growth rate were also calculated separately.
• Postcodes were mapped into each Greater Capital City Statistical Area region for the final tool. Where a postcode crosses a Greater Capital City Statistical Area boundary, the postcode was allocated to the region with the highest overlap.

Skill Level

• The ABS, Level of Highest Educational Attainment at a 3-digit level (HEAP) was used to calculate the skill level of each occupation in 2021. This field contains a list of common Australian qualifications and equivalents.
• Different point scores were assigned to each qualification, ranging from 0 (No educational attainment) to 17 (Higher Doctorate).
• The average education level by occupation was calculated by taking a weighted-average score for each occupation. This was used to formulate the rankings of occupations by average skill level.
• The median and mode qualification for each occupation was calculated directly from the distribution without the assignment of scores.

To understand the potential impact of Generative AI (GAI) on occupations, the primary component involved analysing O*NET task-level data to understand which tasks could be automated or augmented by the GAI. The timeshares for each task were then estimated and the results were converted to Australian occupations (similar approach to the one described in Analysis of Task Change).

The steps involved to determine whether a task could be automated or augmented by GAI were:

• The first step was to tag tasks in the O*NET database as either language based or non-language tasks. Language tasks’ include natural, mathematical, computational, and other ‘languages’. Non-language tasks are those that require some type of manual labour. Since, Large Language Models (LLM’s) have shown significant leaps in capability and performance, this analysis focuses specifically on these models. This analysis is based on the capabilities of GPT-4 (the current state of the art LLM). This means the analysis does not include the impact of image-generating models and other modalities.
• The language tasks identified were then assessed against three criteria:
  • whether the task requires human to human interaction (as opposed to human to computer interaction),
  • whether the task is non-routine and/or non well-defined,
  • whether the task requires human involvement enforced by law, ethics or social conventions.
• Each language task was then categorised as high potential for automation or augmentation based on the number of criteria it met:
  • High potential for automation: tasks do not meet any of the criteria, i.e., it involves human to computer interaction, is relatively routine and well defined, and there’s no human involvement enforced by law, ethics or social conventions
  • High potential for augmentation: tasks meet only one of the criteria, i.e. either human to human interaction required, or it is routinary/well structured, or human involvement enforced.
• All non-language tasks and language tasks that met at least two of the criteria were categorised as unlikely to be impacted by GAI.
• The assessment of tasks against the criteria and subsequent categorisation, was done using a combination of human and machine tagging. GPT-4 was prompted (few-shot) along with some example classifications and used to classify other tasks. A combination of the criteria scores (number of criteria met) from both approaches was used for the final categorisation of a task. Outputs from the GPT-4 tagging were validated by selecting a random sample of results for human review.

Accenture is a leading global professional services company that helps the world’s leading businesses, governments and other organizations build their digital core, optimize their operations, accelerate revenue growth and enhance citizen services—creating tangible value at speed and scale. We are a talent and innovation led company with 738,000 people serving clients in more than 120 countries. Technology is at the core of change today, and we are one of the world’s leaders in helping drive that change, with strong ecosystem relationships. We combine our strength in technology with unmatched industry experience, functional expertise and global delivery capability. We are uniquely able to deliver tangible outcomes because of our broad range of services, solutions and assets across Strategy & Consulting, Technology, Operations, Industry X and Accenture Song. These capabilities, together with our culture of shared success and commitment to creating 360° value, enable us to help our clients succeed and build trusted, lasting relationships. We measure our success by the 360° value we create for our clients, each other, our shareholders, partners and communities. Visit us at www.accenture.com.