Content uploaded by Jo Ann Oravec
Author content
All content in this area was uploaded by Jo Ann Oravec on Dec 21, 2023
Content may be subject to copyright.
Modeling the AI-Driven Age of Abundance:
Applying the Human-to-AI Leverage Ratio (HAILR)
to Post-Labor Economics
Benny Traub, Inspira AI Corp.
Izzy Traub, Inspira AI Corp.
Dr. Jo Ann Oravec, University of Wisconsin at Whitewater
Dr. Phil Peper, Arizona State University
Dr. Paul W. Thurman, Columbia University
Abstract
This paper explores the transformative impact of AI on automating knowledge work leading to
the anticipated 'Age of Abundance' in a post-labor society where work is performed by
machines rather than human labor. Through a detailed model incorporating variables such as
cost of computing, AI model efficiency, and human-equivalent production output (derived from
the human-to-AI leverage ratio, or HAILR), we provide a nuanced albeit tentative analysis of
future productivity trends and economic realities.
The model, integrating conservative estimates like a 30% annual improvement in AI model
efficiency, projects a substantial increase in productivity; by 2044 it indicates that just four
hours of productive human labor could yield as much as 636 years of equivalent output. The
model is not intended as a precise prediction, rather a framework to allow scientists and
laypersons to visualize the inevitability of the coming Age of Abundance. The assumptions are
incidental. If work is automated at scale, one may reasonably change the assumptions in the
model and still arrive at the same conclusion: extreme abundance.
This research also critically examines the potential job displacement in knowledge and office
work sectors, suggesting a loss of 9 out of 10 jobs by 2044 due to AI automation. The model
also shows how the remaining workers will be empowered with their efforts “leveraged” by AI
technologies.
We highlight the economic and societal implications of these findings, including the need for
proactive public policy and corporate strategy to navigate the challenges and opportunities
presented by AI-driven transformations. The study underscores the criticality of grasping these
shifts in timely ways for future workforce planning and societal adaptation. Although the model
will certainly need to be revised to accommodate technological, political, and social changes,
we believe that its simplicity, flexibility, and clarity can earn it a significant role in policy
discourse.
1
Electronic copy available at: https://ssrn.com/abstract=4663704
Literature Review
The notion of ‘knowledge worker’ was developed by Peter Drucker in his 1959 book, The
Landmarks of Tomorrow
(1)
. Drucker was a pioneer in contrasting knowledge work with manual
labor in his managerial analyses. In our paper, the phrase “knowledge work” includes cognitive
work generally performed on a computer. This includes the efforts of programmers, scientists,
writers, and engineers who produce and handle information.
(2) Office work (from simple clerical
tasks to complex, multi-stage efforts) is especially amenable to AI automation. AI capabilities
are also making many knowledge work efforts involving high-level thinking (such as medical
and legal jobs) amenable to automation.
Many early efforts to analyze and understand the dimensions of knowledge work from
economic perspectives were inspired by the 1970s writings of Marc Porat
(3)
, following the lead
of Fritz Machlup
(4) in the 1960s. Mapping the impact of particular technologies such as AI on
the productivity of workers has been an activity of many researchers in the decades that
followed, as outlined in the se ctions to come. The modeling effort described in this paper is
intended to use straightforward terms and common-sense concepts in ways that make the
models usable in public policy deliberations as well as in community outreach or business
planning. Providing clear yet powerful data visualizations and conceptual tools in these forums
will focus these discussions and stimulate the production of useful insights.
Introduction
The idea that artificial intelligence will transform the working lives of people worldwide is not a
new one. AI practitioners and thought leaders such as Elon Musk
(5a) and Kai Fu Lee
(6) have
argued that AI will bring about an “Age of Abundance” that will dramatically decrea se the cost
of goods and services through efficiency and economy of automation. T he model in this paper
presents one perspective on how these dramatic cost reductions will take place. The
productive clout of individual workers will also expand dramatically as it is leveraged by AI
technologies. Just as the philosopher Archimedes reportedly said “Give me a firm place to
stand and a lever and I can move the Earth,” workers in the years to come will have their
efforts dramatically enhanced by the leverage of automation.
Automation through AI will supplement and then eventually replace many human workers
(7)
once the technology is adopted, particularly in knowledge work where jobs are conducted
solely on a computer. Discussed here is the critical question of just how much work will be
accomplished by AI systems within the next two decades, how many net jobs will be displaced
and what the impact will be upon the cost of goods and service. Jobs will go away, but exactly
how many? Predictions about the “disruption” that AI will cause have been expressed by
many corporate and public policy leaders, but specifics are often missing.
Developing and integrating AI systems into the workplace will come at a cost, but the data
show how catastrophic it will be for corporations and governmental agencies that do not adopt
AI technology and how imperative it is to get started immediately. Specifically, the present
2
Electronic copy available at: https://ssrn.com/abstract=4663704
paper introduces a new concept termed the Human-to-AI Leverage Ratio, or “HAILR,” to
describe how much productive output AI automation can accomplish for each minute of human
input. Using HAILR, our model predicts AI output will be so large in the coming years that a
single minute of human labor could produce 2.7 years in human-equivalent output. Whether
this increase in output will provide for societal "abundance" or produce less optimal results
depends on how corporate leaders, public administrators, and community members engage
with each other in planning and disbursement efforts.
Much has been said about how AI tools will increase the speed and reduce the cost of
production, but little has been documented about human-performance and the role it plays in
the future productivity boom. We predict that HAILR will increase rapidly over the next twenty
years, and surprisingly, that the high rate of productivity will come to be even more dependent
on the efficiency of the humans who work to support the AI systems, as one minute of lost
human productivity will initially mean the loss of hundreds, and ultimately thousands of minutes
of productive output. Thus human capital will become far more important in the creation of
value and competitive advantage than it is today, which is consistent with the conclusions of
some managerial consulting firms such as McKinsey: “we’d argue the need for excellent
management will grow even greater.”
(8)
Figure 1 shows human labor decreasing as automation of work increases.
Figure 1: Percentage Human vs AI Output
Source: Inspira AI Corp. analysis
Three Categories of Work
Three categories of work are presented.
Human-centric work
This is work that does not include any automation. It is singularly human. The journey towards
automation has only just begun and we anticipate that by the end of 2024, the vast majority of
all work will still be performed by humans alone.
Human-assisted work
This is work where humans play a role, but much of the productive output is produced by AI
systems. In this category of work, humans may participate by planning or providing input,
3
Electronic copy available at: https://ssrn.com/abstract=4663704
giving AI ‘leadership’. Today this looks like creating prompts to generate articles or images,
then reviewing the work and fine tuning the output. However, in the future, the library of tasks
in this category will expand to many thousands of use cases.
As shown in Figure 2, this category of work will grow rapidly over the next decade but will
eventually shrink, as the next category of fully autonomous work takes over an increasing
share of the workload. The model in figure 2 illustrates 20% of the work currently performed by
humans on computers will be performed by this category of human-assisted work by 2044.
Figure 2: Human, Human-Assisted and Autonomous
Source: Inspira AI Corp. analysis
AI-Autonomous work
This is work initiated and performed solely by AI systems, without any human involvement at
all. If the server doesn’t go down, this work will never stop. Today the number of items on this
list is small, and mostly invisible to us humans. Think of the clock on your phone changing to
daylight savings time automatically. The list of fully autonomous tasks will grow rapidly over the
next two decades, taking more and more of the load from the above two categories. Our
calculations project that 70% of work currently performed by humans on computers will be
performed by this fully autonomous category by 2044.
Knowledge work will be automated faster than manual labor
The model in this paper focuses on knowledge work, specifically work that is currently
performed by humans on computers. Broadly this affects any job that involves manipulating or
moving information. This would include domains such as engineering, accounting,
mathematics, writing, design, clerical, sales, data entry, legal work and much, much more. This
includes jobs which require communication as existing AI technology can already demonstrate
human-level empathy and emotions during conversation
(9)
. For the purpose of this paper,
knowledge work does not include jobs that primarily require physical labor, such as
construction or house keeping. It is anticipated by the authors that knowledge work will be
automated at a much faster rate than work which requires manual labor.
While it is widely accepted that the amount of work that is automated will grow, the rate of
automation will vary widely for each job role
(10,11)
. Some tasks are more difficult and expensive
to automate. Replacing manual labor requires the development and production of mechanical
4
Electronic copy available at: https://ssrn.com/abstract=4663704
robots, which are expensive. The operation of robots requires computing hardware, software,
and ongoing maintenance, all of which adds to the cost of production. Conversely, the
automation of office work requires only computing hardware and software, without the
mechanics. Thus, it is far easier and less expensive for white-collar industries to automate than
it is for industries that rely upon human labor
(11, 12)
.
The Phases of Automation
Phase I - Automation Potential
Automation potential is the general maturity of technology and its capacity to automate work.
Chui et al.
(14) refer to this as the potential for technical automation - the capacity for automation
to complete work conducted by humans. Such potential is the availability of general technology
which acts as building blocks which can be customized for specific use cases, then integrated
into production environments. It is AI technology that is capable of automating tasks, but which
in and of itself, automates nothing. As examples, deep learning and large language models fall
into this category. According to Chui, M., Ellingrud, K., et al. of McKinsey, as of 2023, the
global Automation Potential is currently between 60 to 70% of all hours worked glob ally
(14a)
.
This number could reach 90% before 2035
(14b)
. In theory, at that time 90% of all tasks could be
au tomated, but the actual number depends on Phase II and III.
Figure 3 demonstrates how McKinsey’s data for Automation Potential reaches 90% before
2035 and 95% before 2040.
Figure 3 - Automation Potential: Maturity of Technology. Adapted from McKinsey
(13)
5
Electronic copy available at: https://ssrn.com/abstract=4663704
Even when using McKinsey’s Late Scenario, the world will be capable of automating 90% of
work before 2035, but capability does not guarantee adoption.
Phase II - Customization
This phase involves tailoring and configuring the technology to suit specific use cases,
environments, or operational requirements. During this phase, the general capabilities of the
technology are fine-tuned to address the unique challenges and needs of a particular
organization, industry or task. This may involve modifying the technology to integrate with
existing systems, adjusting it to comply with industry standards or regulations, and ensuring it
meets the specific performance criteria of the use case. Third parties may create customized
solutions, then make them commercially available to organizations, or companies may, as part
of Phase III adoption, initiate the development of custom solutions.
Phase III - Adoption
This phase involves the decision of business executives to adopt automation into their
production process. This may involve integrations with off-the-shelf solutions, and/or retaining
qualified people or firms to build custom solutions. The development and implementation of
strategies for the maintenance and updating of the AI systems follows. As opportunities or
serious issues with the systems are identified, the current systems may be optimized, or
scrapped and new systems begun (back to Phase I or II). Future AI technology may become
capable of self-optimization, so a return to previous phases may not obligate human input.
Chui et al.
(14) suggested that the pace of workforce transformation is likely to accelerate. While
the trajectory of the potential for automation does suggest acceleration, frictives exist that may
result in a more homogeneous pace of adoption. These include the relative cost of automating
certain tasks over others, the repetition rate of tasks, the value of completed tasks and the
motivation of stakeholders to invest in automation, not to mention unknowns such as
regulatory compliance requirements, only just emerging.
Priorities of Automation
We offer for consideration that frequent and high value tasks could bring an economic return
faster than their counterparts and thus, highly repetitive tasks are likely to be automated before
tasks with less frequent repetition, as are tasks that have high value (think medical diagnosis).
The forces of repetition and value are balanced by the difficulty and the cost of customization
and integration, as the engineering requirements to automate dissimilar tasks varies widely.
We anticipate the possibility that over time, with the assistance of AI, the pace of automating
tasks will increase and the cost of development will decrease for similar tasks. As easy and
high value tasks are likely to be automated first, that will leave the difficult and lower value
tasks for last, or at least until the cost of development and engineering for automating those
remaining tasks drops to a cost-efficient level.
6
Electronic copy available at: https://ssrn.com/abstract=4663704
Auto-Automation
A coming scenario on the horizon is when AI becomes capable of creating customizations and
integrations with little or no human engineering. Call it ‘auto-automation’. Consider a task that
is performed only once. It generally isn’t economically feasible to engineer the automation of
such a task, as the automation engineering could easily take longer than performing the task
itself. However; a more advanced AI with finetuned domain knowledge could potentially
develop a customized solution on demand. We posit auto-automation will be needed to push
automation upwards of 90% for all work. In some fields, auto-automation capabilities may
emerge more quickly than in others (fields in which the assumptions behind processes are
more transparent, for example). However, the trajectory indicates auto-automation capabilities
will eventually be widespread, leading to even more extensive and pervasive automation of
work.
Methods
The model utilizes the metrics below. Certain assumptions made, such as the rate at which AI
automation will be adopted. It is important to note that changing the assumptions changes only
the timing and depth of the societal changes involved with the Age of Abundance. The general
outcome of monumental reduction in the cost of services remains extreme, changing by only a
matter of degree.
For example, the current assumptions in the model predict that one person will be able to
produce over 600 years of productive output in a single day of work. This is achieved due to
the massive leverage that automation will be capable of delivering. Computing hardware
continues to get faster and AI models continue to get more efficient. As more and more work is
fully automated, the output is less tightly constrained by human input.
One may change the model assumptions enough to reduce the projected output from 600
years worth of production down to 1 year of production, (for a single day worked by a human).
But regardless of the weights applied to the model, it mathematically demonstrates how the
cost of goods and services will plummet, resulting in an Age of Abundance. If we continue on
the trajectory towards full automation, the general directions and dimensions of the outcome
appear to be certain. Each variable of the model is described below:
1 - Cost of Computing
Cost of Computing refers to the cost of storage, bandwidth, networks and computational
calculations performed by CPUs, GPUs, memory and the like. We used an annual decay rate
of 15%. Our model begins in 2024 with a cost of $1 for 1 unit of computing, but the starting
place is irrelevant to the outcome, as a cost per unit of $100 produces the same output.
2 - Computing Units
As the cost of computing decreases, the number of Computing Units one can purchase for the
same amount of money increase. If the cost of computing decays by 15% per year, in twenty
7
Electronic copy available at: https://ssrn.com/abstract=4663704
years we can purchase 16.37 units for the same money as will purchase 1 unit of computing in
2024.
3 - AI Model Efficiency
AI Model Efficiency refers to the ability of a model to take advantage of computational capacity.
The efficiency of AI models is growing at a pace that can often be well characterized as
exponential, with some experts predicting a 100x increase in efficiency over the next 5 years.
This is due to a number of factors, including advances in hardware, software, maintenance
strategies, and algorithms.
For example, the development of new hardware technologies such as neuromorphic chips is
specifically designed to accelerate AI workloads. Additionally, software advances such as
compiler optimizations and new programming languages are making it easier to develop and
deploy efficient AI models. Improved strategies for system maintenance are eliminating many
problematic issues, keeping systems running. Finally, algorithmic advances such as model
pruning and quantization are helping to reduce the size and complexity of AI models without
sacrificing performance.
These factors are all contributing to an increase in the efficiency of AI models, which is
enabling new applications, and potentially accelerating the adoption of AI across industries.
It was reported that OpenAI’s GPT4 model was ten times more efficient than their previous
model, GPT3.5. Elon Musk has stated in an interview with Rishi Sunak that the capability of AI
is growing at the pace of five-fold or ten-fold per year.
(5b)
Statements by thought leaders indicate the trajectory of development, but are not predictions
of what will happen next, and certainly not over the next twenty years, so we opted for a
conservative estimate of model efficiency improvement at 30% each year.
The starting parameter value for Model Efficiency was 1, which gives us a relative
mathematical starting point for the schematic.
4 - HAILR Multiple
The Human to AI Leverage Ratio illustrates the productivity-leverage obtained through
automation, where one minute of human input produces multiple ‘Output-Minutes’ by AI
systems. The multiple is the reciprocal of HAILR.
Example;
● Human works 1 minute
● AI system outputs 100 minutes of human-equivalent productivity
● HAILR = 1 to 100
● The reciprocal in this example is 100, which is used as the starting multiple in the
model, demonstrating the leverage afforded by the AI system
8
Electronic copy available at: https://ssrn.com/abstract=4663704
The use of 100 as the model’s starting parameter satisfies the intention of the model, which is
to demonstrate that automation at scale results in abundance. The actual multiple in
production depends highly on the specific use case. E.g. In just one minute of prompt
engineering by a human, an LLM may produce several hours worth of writing, or several days
worth of artwork. The number of variables that influence the output minutes, and even the
human input, are diverse across use cases. Starting with a multiple of 100 demonstrates the
intention of the model, while still being observable by readers in their experience with available
automation tools.
We posit that HAILR will increase over time, as the cost of computing drops and the efficiency
of AI models increase.
5 - Proportion of Automation (PoA)
The Proportion of Automation is the proportion of human labor that will be automated, with the
PoA projected to increase over time.
Many previous sources forecast ultimate job disp lacement of net 40%-50%, notably the Oxford
Study
(12) as far back as 2013. We do not challenge these previous predictions, however, we
point out that they are generalized across all industries, whereas knowledge work is far easier
and less expensive to automate than manual labor. The reduced friction in the automation of
knowledge work, and in particular, work performed on a computer indicates that this will result
in faster and more comprehensive automation of knowledge work, and therefore the net job
loss will not be equally distributed across all industries, but in fact is weighted towards work
that does not require mechanical robots, at least over the next twenty years.
The journey of towards automation is in early stages, so we started with a total Proportion of
Automation of 1% in 2024. As milestones, we set our model to 30% by 2030, which is
consistent with McKinsey’s Ellingrud, K., et al.
(13) and to 90% by 2044 (see figure 4), which is
also within the range allowed by McKinsey, although on the more aggressive end of their
scale. We explain why below, under “Forced Adoption.”
9
Electronic copy available at: https://ssrn.com/abstract=4663704
Figure 4: % of Automation of Current Work Activities. Adapted from McKinsey
(14e)
Forced Adoption
Adoption of customized AI and automation solutions by organizations is underway. If it appears
there is a slow start, this may be partially due to general lack of awareness by some, of the
current capabilities of AI, at least with regards to use cases within organizations (automatic
writing of articles is one of the few widely known use cases). It may also partially be due to the
current lack of publicly advertised, off-the-shelf automations that might benefit an organization.
Chatter in the development world indicates that many customized solutions are under
development, but have not yet hit the market.
If organizations are hesitating or reluctant to aggressively automate their operations, this could
also be due to the fact that competitors have yet to begin producing goods and services at
vastly lower costs. Adoption is currently optional. But if the trajectory holds, this optional state
of affairs will not last, especially as the labor market acquires considerable skills and know-how
related to leveraging AI applications. There will come a tipping point in pricing where certain
companies utilize automation to the extent that their operational costs fall low enough to make
a noticeable difference in their pricing strategies.
The tipping point of pricing will be different for every industry, but if a competitor automates and
can thereby lower its price significantly yet still make the same profit, this could make
competition untenable. In this scenario, every company will be forced to adopt or die. Under
this scenario, adoption will not be optional. Indeed, early adoption by some could create a
waterfall effect where adoption becomes a forced race (even a stampede), and those who get
started early enough may get such a lead that they cannot be overtaken.
10
Electronic copy available at: https://ssrn.com/abstract=4663704
McKinsey’s assessment is that automating 90% of work may happen as soon as 2043 or as
late as 208 0 . The model in this paper leans towards the early date, as it is highly plausible that
forced adoption becomes the norm, leaving companies with no choice but to automate.
As yet we have few public examples of dramatic price decreases, but they appear to be on the
horizon. It is our hypothesis that competitive price cuts will upend operational priorities with a
sense of urgency that will dramatically speed up adoption, and that there may be significant
opportunity for market leadership for those companies who are first to automate at scale.
6 - Input Minutes
The parameter Input Minutes refers to the amount of time a human is providing input to AI
systems each day. This is the time that would have previously been worked by a human to
directly produce value. In this model, the human input minutes are instead used to steer and/or
provide feedback to AI systems, which can presumably produce more value in less time than it
would have taken the human.
While there are a great many definitions of productivity, and a great many ways of measuring
it, for the purpose of this paper, productive work is herein defined as work that directly or
indirectly produces the goods and services that might be consumed by others. It does not
include time wasted, nor time spent on maintaining, such as cleaning one’s desk. The values
provided are based on the assumption that a typical human productively creates value for only
about half of the time they are working in average 8-hour workday. There are currently no
generally accepted productivity statistics for work of this definition, so we applied a measure of
just over four hours per day to allow for a modicum of loss due to inefficiency or tasks which do
not contribute directly or indirectly to the production of value. It is important to note that
whether the actual number is 2 hours, 4 hours or even 8 hours, the outcome of abundance is
only by a matter of degree. Extreme levels of abundance are still inevitable.
The input minutes are calculated by multiplying the Rate of Automation by the assumed
number of minutes per day that a human typically works productively. As a starting place, we
used 262 productive minutes per day and a 1% Rate of Automation, resulting in an initial
parameter value of 2.62 minutes per day where humans are steering AI systems to produce
value.
7 - Output
Output refers to how many units of human-equivalent output is produced by automation. This
is the product of the HAILR multiplier by the Input Minutes . At the initial parameter values, a
HAILR multiplier of 100 produces 262 Human Equivalent Output minutes. This is based on
2.62 minutes of human input. The output column of the model provides the Human Equivalent
Output in weeks, months or years to better make sense of the data as time goes by and as
production scales. Note that we assumed a work-week of 40 hours when calculating output in
weeks, and a year of work at 50 weeks to allow for two weeks of holiday.
11
Electronic copy available at: https://ssrn.com/abstract=4663704
The Model
This model is intended to reveal the inevitability of the reduction in cost of goods and services,
as human work is automated. While the authors consider the projected dates plausible, the
model is not intended as a prediction of precise dates. We invite researchers to manipulate the
variables and add their own refinements, with confidence that similarly remarkable outcomes
will result.
Year
Computing
cost
Computing
units
AI Model
efficiency
HAILR
Multiple
PoA
Input
Minutes
Output
minutes
Output
years
2024
1.00
1.00
1.00
100
1%
2.62
262
0.002
2025
0.85
1.15
1.30
150
3%
7.8
1,173
0.010
2026
0.72
1.32
1.69
224
7%
18
4,093
0.036
2027
0.61
1.52
2.20
334
12%
31
10,489
0.091
2028
0.52
1.75
2.86
500
18%
47
23,522
0.20
2029
0.44
2.01
3.71
747
25%
65
48,841
0.42
2030
0.38
2.31
4.83
1116
30%
78
87,621
0.76
2031
0.32
2.66
6.27
1669
35%
92
152,825
1.3
2032
0.27
3.06
8.16
2495
40%
105
261,112
2.3
2033
0.232
3.52
10.60
3731
45%
118
439,158
3.8
2034
0.197
4.05
13.79
5577
50%
131
729,491
6.3
2035
0.167
4.65
17.92
8338
55%
144
1,199,647
10
2036
0.142
5.35
23.30
12465
59%
154
1,923,907
17
2037
0.121
6.15
30.29
18635
63%
165
3,071,240
27
2038
0.103
7.08
39.37
27860
67%
175
4,883,028
42
2039
0.087
8.14
51.19
41650
71%
186
7,735,956
67
2040
0.074
9.36
66.54
62267
75%
196
12,216,817
106
2041
0.063
10.76
86.50
93089
79%
207
19,238,230
167
2042
0.054
12.38
112.46
139169
83%
217
30,217,414
262
2043
0.046
14.23
146.19
208057
87%
228
47,352,145
411
2044
0.039
16.37
190.05
311045
90%
235
73,232,541
636
Primary Findings
By 2044…
● One minute of human labor could produce 311,045 minutes (2.7 work-years) of
productive output.
● One person, with four hours of productive work, could produce 636 years of productive
output.
● Nine out of ten knowledge workers will be permanently displaced from employment.
12
Electronic copy available at: https://ssrn.com/abstract=4663704
These projections provide some sense of how disruptive these AI-driven changes in production
may be. Rather than demeaning human knowledge workers, the results show how productive
the remaining workers will be with the assistance of AI automation. Their efforts will be
leveraged with the capabilities of AI technologies.
Discussion
This paper has focused on the transfer of knowledge work from humans to machines. If, as we
suggest, most knowledge workers will eventually lose jobs without new jobs to replace them,
then additional research must be done to find those people meaningful purpose and livelihood.
The new human-AI collaborations may indeed be satisfying and enriching to the knowledge
workers who obtain them
(15)
.
It has been suggested UBI (universal basic income) be given to displaced workers, but this
may be problematic. Companies will not willingly give up profits to feed people who are not
contributing, thus taxation must be enforced. However, this has not consistently worked in the
past with large corporations so alternative ideas must be generated. For example, perhaps the
government could allocate land in rural areas to people who would like to return to a
self-sustaining, husbandry way of life. It would be an ironic outcome if the high tech movement
resulted in many of us living a more low tech lifestyle!
We acknowledge that in addition to automating tasks, AI may also be applied to aiding humans
in being more productive by helping them focus longer on higher value tasks, thus increasing
the time worked each day on tasks that produce real value. It is anticipated that were this to
happen, that it may give companies a significant competitive advantage, as each incremental
minute rescued could mean weeks, months or even years of additional productive output. For
as long as automated activities require the input of a human, the variable of rescued minutes
could bring about dramatically improved efficiency of output. On the level of human emotion,
having one’s time be productive to its maximum level could result in increased job satisfaction
and even self-actualization, thus both the employer and employee will likely benefit from
aggressive adoption of solutions that increase human productivity, not necessarily enjoined to
the automation process, but in and of itself.
Limitations
There may be unforeseeable barriers and frictions to furthering automation potential, or
adoption of automation even if full automation potential is achieved
(16, 17)
. Barriers could prevent
the world from achieving automation at the levels shown in the model and frictives could slow
the trajectory towards mass automation. As just one example, some kinds of complex
non-repetitive work will be automated only when more self-managing AI exists, which we have
referred to in this paper as ‘auto-automation’. Today it is just too expensive to engineer
automations for most workflows that are less repetitive. If AI does not develop to the extent
that automation of low-repetition tasks can be economically automated, that would be a barrier.
13
Electronic copy available at: https://ssrn.com/abstract=4663704
Other barriers and obstacles could include government-imposed or union-negotiated
restrictions on the use of AI.
A limitation of the model itself is the proportion of automation. While automation of 90% of work
suggests the loss of 90% of jobs, in fact it could be far more severe. Consider that if 90% of
human labor is replaced, but the AI-enabled output produces years worth of equivalent work in
a matter of hours, then it is not necessary for 10% of the people to remain employed. Just 1%
of the current workforce, or even a small fraction of that, could conceivably meet all global
demand.
Another limitation to consider regarding the model is that rather than laying off 90% or more of
the workforce, employers may keep more people employed, but reduce their required hours or
give them sinecure responsibilities. Thus fewer people become unemployed. However, if the
stigma of being “unemployed” is removed, changes in society that lead to more meaningful
and fulfilling lives for everyone can be wrought
(18)
.
The definition of “knowledge work” is a limitation, in that there is no single definition and the
scope of what can be characterized within its boundaries is expanding. But generally, cognitive
work that is performed on a computer is ripe for automation, because anything that we can
perform on a computer, can theoretically be replicated by a computer.
Conclusions
The world is heading towards a seismic shift that will have unprecedented impact on the way
we live, not just work. Modeling this shift provides corporate leaders, policy makers, and
community participants with insights as to the economic and social changes the shift will make
on society. HAILR (Human to AI Leverage Ratio) enables modelers to explore the impacts of
AI on production with some level of specificity, unlike many of the narrative-style projections.
The accuracy of the model depends on factors that have not yet clearly emerged, such as the
scope of the regulation of AI by nations. Model building efforts such as the one in this paper
can equip communities, states, and nations to understand trends in productivity and ensure
that the increased output produces true "abundance."
Further Research
The HAILR multiple presented has been generalized across all knowledge work. It is
anticipated that some tasks and even entire job roles will be entirely automated, with zero
human involvement. Further research could explore granular calculations for specific
industries, job classes and tasks.
The measurement of true human productivity deserves additional research, as currently there
is no single, widely accepted method for creating a generalized metric that might be equally
applied across job functions. Sickles & Zelenyuk
(19) made a great effort, but their methodology
is apparently not convenient enough to use in production to the extent that it is widely used to
14
Electronic copy available at: https://ssrn.com/abstract=4663704
compare the productivity of jobs where the duties are vastly different, such as salespeople vs.
engineers. More development could result in a workable solution for production environments,
such that all employees could equitably be compared with a single metric.
This model is a starting point, to provide scientists with some basic, yet plausible math that
demonstrates how high levels of automation will result in an Age of Abundance. This isn’t just
a talking point, it is real. Yet we acknowledge the model could be improved by considering
additional variables. For example, humans require breaks from work because of fatigue and
illness so a model with this additional variable, and others, may improve the precision. In
addition, the ratio of less repetitive tasks to highly repetitive tasks might be considered, which
could give greater insight into the frictions that could slow the rate of transformation.
As AI may also be applied to aiding humans in being more productive by helping them focus
longer on higher value tasks, additional research should be conducted into the precise
application of AI to this problem.
Auto-automation is a topic for additional research. What can AI currently self-automate? How
far out on the horizon is auto-automation at scale? A more precise prediction regarding the
year-over-year Rate of Automation could benefit from this data.
The authors acknowledge that the content in this paper will benefit from updates, which are not
possible to make after publishing. For this reason, updated versions may be found here:
https://inspira.ai/science/hailr-and-the-age-of-abundance/
WORKS CITED
1. Drucker, P. F. (1959). Landmarks of Tomorrow. New York: Harper & Row.
2. Reinhardt, W., Schmidt, B., Sloep, P., & Drachsler, H. (2011). Knowledge worker roles
and actions—results of two empirical studies. Knowledge and Process Management,
18(3), 150-174.
3. Porat, M. U., & Rubin, M. R. (1977). The Information Economy (No. 77). Department of
Commerce, Office of Telecommunications.
4. Machlup, F. (1962). The Production and Distribution of Knowledge in the United States
(Vol. 278). Princeton University Press.
5. Elon Musk
a. “we will be in an age of abundance”
https://www.telegraph.co.uk/politics/2023/11/02/rishi-sunak-latest-news-ai-summit
-elon-musk-donelan-live/
b. “It (AI) seems to be growing in capability by at least five-fold, perhaps ten-folder
per year”
https://www.youtube.com/watch?v=MLcGWOV7Ojw&list=WL&index=3&t=424s
6. Fu Lee, K. (2021). AI 2041: Ten Visions for Our Future. Crown Currency Press.
7. “Jobs are definitely going to go away, full stop”. -Sam Altman, CEO OpenAI
https://www.marketingaiinstitute.com/blog/sam-altman-atlantic
15
Electronic copy available at: https://ssrn.com/abstract=4663704
8. “We’d argue the need for excellent management will grow even greater”
https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insig
hts/the-organization-blog/middle-managers-hold-the-key-to-unlock-generative-ai
9. Traub, B., Thurman, P., Kay, C., & Oravec, J. (2023). Perception of Human
Characteristics in an AI Chatbot Acting as On-the-Job Coach. Available at SSRN
4647255.
10. Holzer, Harry J. (2022). Understanding the impact of automation on workers, jobs, and
wages. Brookings Institution. Retrieved from
https://policycommons.net/artifacts/4142261/understanding-the-impact-of-automation-o
n-workers-jobs-and-wages/4950046/ on 13 Dec 2023.
11. Casey, Kevin (2021). Automation by the numbers: 11 stats to know.
https://enterprisersproject.com/article/2021/11/automation-11-statistics-know
12. Frey, C. and Osborne, A. (2013) Oxford Study - THE FUTURE OF EMPLOYMENT:
HOW SUSCEPTIBLE ARE JOBS TO COMPUTERISATION?
13. Ellingrud, K., et al., (2023) Generative AI and the Future of Work in America. McKinsey
Center for Government. By 2030, activities that account for up to 30 percent of hours
currently worked across the US economy could be automated
https://www.mckinsey.com/mgi/our-research/generative-ai-and-the-future-of-work-in-am
erica
14. Chui, M., Hazan, E., Roberts, R., Singla, A., & Smaje, K. (2023). The economic
potential of generative AI. McKinsey Global Institute.
a. Current generative AI and other technologies have the potential to automate
work activities that absorb 60 to 70 percent of employees’ time today.
b. Technological adoption will be cheaper and easier to implement in knowledge
worker domains where the individuals simply work on computers.
c. The effects of AI on job automation could accelerate over time, but the ultimate
effects on workplace transformation depend on how widely technologies are
adopted and workers are trained to use the new tools.
d. The advent of generative AI has pulled forward the potential for technical
automation.
e. Global automation of time spent on current work activities
15. Oravec, J. A. (2023). The Shape of Workbreaks to Come: Reframing Cyberslacking
With Bossware and Artificial Intelligence. Journal of Organizational and End User
Computing (JOEUC), 35(3), 1-17.
16. Appenzeller, G., Bornstein, M., & Casado, M. (2023). Navigating the high cost of AI
compute. Andreesen Horowitz. https://a16z.com/navigating-the-high-cost-of-ai-compute/
17. Filippi, E., Bannò, M., Trento, S. (2023). Automation technologies and their impact on
employment: A review, synthesis and future research agenda.
https://www.sciencedirect.com/science/article/pii/S0040162523001336
18. Garcia-Lorenzo, L., Sell-Trujillo, L., & Donnelly, P. (2022). Responding to stigmatization:
How to resist and overcome the stigma of unemployment. Organization Studies, 43(10),
1629-1650.
19. Sickles, R. C., & Zelenyuk, V. (2019). Measurement of productivity and efficiency.
Cambridge University Press.
16
Electronic copy available at: https://ssrn.com/abstract=4663704
