Revue d'économie industrielle

169 | 1er trimestre 2020

Industry 4.0: Current Issues and Future Challenges

Artificial Intelligence and the Future of Work

Intelligence Artificielle et Avenir du Travail

Salima Benhamou

Electronic version
URL: https://journals.openedition.org/rei/8727
DOI: 10.4000/rei.8727
ISSN: 1773-0198

Publisher
De Boeck Supérieur

Printed version
Date of publication: 1 September 2020
Number of pages: 57-88
ISBN: 978-2-8073-9397-4
ISSN: 0154-3229

Electronic reference
Salima Benhamou, “Artificial Intelligence and the Future of Work”, Revue d'économie industrielle [Online],
169 | 1er trimestre 2020, Online since 05 January 2023, connection on 06 January 2023. URL: http://
journals.openedition.org/rei/8727 ; DOI: https://doi.org/10.4000/rei.8727

All rights reserved

 ARTIFICIAL INTELLIGENCE
AND THE FUTURE OF WORK
Salima Benhamou, Labour-Employment-Skills Department France Stratégie

Keywords: Artificial Intelligence, future of work, learning organisa-

tions, healthcare sector, transport sector, banking sector

Mots-clés : Intelligence artificielle, avenir du travail, organisation

du travail apprenante, compétences, conditions de travail, secteur de la
santé, secteur des transports, secteur bancaire

1. INTRODUCTION
Artificial intelligence—meaning the group of technologies that carry out
computationally tasks traditionally assigned to human beings—is central
to current debates around the world on social and technological changes.
About ten years ago, Artificial Intelligence (AI) technologies started to
achieve remarkable progress in a surprising variety of applications. This is
due to three interconnected technological advances: increases in process-
ing power that allowed for larger models to be trained with machine learn-
ing algorithms; the availability of large amounts of annotated data for
purposes of training these large models; and progress in machine learn-
ing theory resulting in improvements in learning algorithms. 1 In the near
future, technological progress may well enable AI to carry out increasingly
complex tasks, coming ever closer to rivalling human cognitive capacities.

1 For a discussion of the factors driving recent advances in AI, see Chollet, 2018,
pp. 20-23.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 57
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

The machine’s victory in the game of Go, the first autonomous vehicles,
and the successful use of computer-aided medical diagnosis software are
all emblematic of the advances made thus far. There are an impressive
number of areas where AI can be applied, including healthcare, transport,
banking and insurance, retail, science. This broad scope of application has
led to the view that AI is a “general-purpose technology” that potentially
will disrupt all aspects of life, economy and society (Brynjolfsson, E. and
McAfee, A., 2014).

Some observers see AI as an economic opportunity due to the productivity

gains it may generate (lower costs as a result of automation of operations,
improvement of coordination processes, production flow optimisation,
etc.) and the new markets it may create. AI is also perceived as a social
opportunity thanks to the processing of big data generated by connected
systems which may well give rise to new professions (data scientists, AI
programmers, etc.) and improve working conditions by taking over repet-
itive routine tasks. Others, however, see AI as a threat to employment
and as a technology that will aggravate inequalities and social polarisa-
tion, with the almost certain disappearance of whole realms of activity in
many sectors, and various professions, some requiring few qualifications
but others that are highly skilled (lawyers, auditors, physicians, etc.).

But how real is the risk for substitution of human tasks with AI? A num-
ber of researchers have put forward the hypothesis of massive automation
of existing jobs by new digital technologies including artificial intelli-
gence. The famous study by Frey and Osborne (2013), from the University
of Oxford, predicted that 47% of total employment in the United States
was at a high risk (70% chance or higher) of vanishing over the next two
decades. Other studies adopting the same methodology predicted similarly
alarming impacts in other industrialized nations (Bowles, 2014). The stud-
ies adopting the methodology of Frey and Osborne (2013) are carried out
at the occupational level and assume that the same occupational category
will be impacted in the same way regardless of the size of the firm and
the country of location. However, Arntz et al., (2016), in a study carried out
at the task level and taking into account country-specific differences in
the job content of the same occupations, estimated that only between 10%
and 15% of occupations are at high risk of being automated. 2 Furthermore,

2 See Le Ru (2016) for a critical assessment of the literature.

58 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

it is important to appreciate that such studies only focus on the technical

potential for job elimination without taking account the fact that new jobs
will be created by technical change. 3

Given the uncertainty in recent studies concerning the impact of AI on

jobs and employment, we turn to recent history in order to imagine what
work will be like in tomorrow’s world and the consequences of the adop-
tion of artificial intelligence on employment. History shows us that tech-
nological advances have never led to abrupt changes leading to massive
destruction of jobs. On the contrary, their effects have been progressive
and have always resulted in the emergence of new forms of work. But are
we not confronted with an altogether new phenomenon? Artificial intel-
ligence may not only impact the volume of employment but also its con-
tent, as it is no longer a matter of automating tasks depending primarily
on physical strength, agility and speed, as was the case in previous indus-
trial revolutions, but rather one of automating cognitive tasks.

It is difficult to know what will be happening in fifteen or twenty years

from now as regards technological advances and their dissemination and
appropriation in the world of work. Technology and how much it is used
are certainly factors of change, but they are far from being the only deter-
minants of the transformation of work organisation and practices. Other
factors also contribute to “shaping” work, including the legal environ-
ment, the economic context (competition in particular) and the social
environment as regards the level of education, access to training, indi-
vidual aspirations or demography. If we want to project ourselves into
the future in order to identify the benefits and risks connected with AI,
we must also incorporate these contextual factors which combined with
future advances in AI may well transform work and employment.

The objective of this article is to clarify the debate on the impact of AI on

work by starting from a consideration of what we know today about AI,
both in terms of its limits and its potential. Defining artificial intelligence
is the first step in providing the keys to understanding future transfor-
mations (Section 2). The uncertainties regarding its longer-term potential
serve to “inflame” the ongoing debate, given that AI raises questions that

3 For the various limitations of the forward-looking literature seeking to predict the
future impact of AI on employment, see Muro et al., 2019, p. 20.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 59
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

go far beyond the world of work including fundamental questions con-

nected with ethics and social acceptability, personal data protection, AI
designers’ and users’ responsibility and transparency of human/machine
interactions. This article does not seek to supply exact answers on what
work will be like in tomorrow’s world. Its main goal to examine how AI
is being used today and, on that basis, to identify plausible possibilities
for its future use in particular fields of application and selected sectors of
activity. The sectoral approach is useful for identifying potential oppor-
tunities and risks in detail, including task evolution, learning dynamics,
increases in technical and social skills, changes in working conditions,
managerial practices and gains in or loss of autonomy (Section 3). The arti-
cle focuses on three sectors in which AI has already started to spread:
health, retail banking and transport. They are also regarded as sectors
with the potential for creating jobs and are therefore important focuses
of public policies. Finally, once the risks and benefits of AI on work have
been analyzed, the concluding section will consider what forms of work
organisation are best adapted to promoting human-machine complemen-
tarity, while enabling organisations to reconcile a high level of innovation
with a high level of quality of work and employment (section 4).

2. THE REALITY OF AI

2.1. The potential of AI

Artificial intelligence is a scientific discipline that is by no means new

with its foundations dating back to the beginnings of computer science in
the 1940s and 1950s, with numerous different methods whose purpose is
to reproduce cognitive functions by computers. The term “artificial intelli-
gence” itself was coined in 1956 by John McCarthy, one of founding fathers
of the field along with Allen Newell and Herbert Simon.

One branch of AI, known as machine learning or statistical learning, has

made spectacular progress over the past few years due to the remarkable
efficiency of multilayer deep neural networks in performing classification
tasks, of images in particular, following a learning phase based on a large
number of examples. Pattern recognition seemed to require human intel-
ligence, given the almost infinite dimensions of the problem to be solved

60 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

(the number of parameters characterising an image). Advances in neural

network design, however, show that this is not the case and this is what
sparked off the recent AI revolution that is raising new questions about
the transformations of work.

AI brings together a range of fields including logical reasoning, knowl-

edge representation, and natural language perception and processing. Its
main applications at present are connected with advances in machine-
learning techniques, deep learning in particular, which usually requires
the availability of big data. A first type of application consists of radically
simplifying human/machine interaction. Voice recognition and synthesis
and natural language processing (NLP), whether to engage in simple con-
versations between people and machines or for automatic translation, are
just a few examples of this initial group of generic applications usable in a
whole range of activity fields. A second type of application is recognition
of specific patterns in big data, resulting from multiplication of sensors
or organised collection. Examples include image and video analysis, facial
recognition and detection of breakdown precursor signals. These two
major use categories, both of which are closely connected with the degree
to which the activities concerned are digitised, are already possible. 4

Without going into the details of these technologies, suffice it to say that
they reproduce existing classifications or achieve well-defined objectives
such as winning a game. Even though the exact mechanisms that result
in such efficiency are not yet fully understood from a theoretical point of
view, the technology is nonetheless deterministic and controlled. It is the
AI programmer that chooses the software architecture he or she wishes to
use (type of neural network, number of layers, etc.), the learning method
(initialisation algorithm and updating of weights for each neuron) and
the training data to be used. We are therefore a long way off from an
autonomous system operating independently of its designer.

Such systems can possess a real capacity for learning in the sense of the
automatic exploration of a solution space much deal larger than the algo-
rithm’s designer would have been able to imagine. This is how AlphaGo
software, which was initially trained to play Go on the basis of millions of
recorded matches, was subsequently able improve its strategies by playing

4 For an overview of machine learning applications, see Alpaydin, 2016, Ch. 1.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 61
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

against itself until it ultimately succeeded in winning against a human

being. This example does not serve to demonstrate any intelligence or con-
sciousness on the machine’s part, but rather an ability to solve compli-
cated problems characterised by a high configuration space.

2.2. The limits of AI

Although AI has made considerable progress in recent years, there are

several limitations to this technology that prevent it from competing
with humans on complex tasks or activities. These limitations are mainly
related to the access and quality of big data and the inability to under-
stand and explain complex mechanisms that are not based on determin-
istic laws.

First, with respect to its use of big data, AI requires colossal computing
capacity to train algorithms for the exploitation of deep learning. For
example, according to Jangquing Jia, director of engineering for Facebook’s
AI platform, “to train one typical ImageNet model takes about one exa-
flop of computing” 5. Achieving human brain capability would require sev-
eral orders of magnitude of increase in computational power that is out of
reach today, which is a severe limitation of AI.

Secondly, training algorithms requires data properly annotated by

humans, which may require considerable effort, particularly for training
large neural networks. The work consists in putting the data in a form
appropriate for training the algorithms by cleaning them, annotating
them and converting them into a usable format for the users. For example,
ImageNet, an image database, required nine years of work and its contrib-
utors manually annotated more than 14 million images 6. The data must
also be sufficiently representative of the problem to be solved.

Furthermore, the quality of AI systems also depends on the training base

on which the algorithm was built. If the training data contain biases in
terms of such factors as gender or location the algorithm will naturally

5 See: https://www.techrepublic.com/article/four-ways-machine-learning-is-evolving-
according-to-facebooks-ai-engineering-chief.
6 See: https://www.kdnuggets.com /2018/05/data-labeling-machine-learning.html.

62 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

reproduce these biases in its recommendations. The biases may come from
the very modelling of the system architecture and in particular from the
categorization of the possible decisions and the nature of the performance
criterion or the family of decision rules accessible by the algorithm. The
data used to train the algorithms may also be non-representativeness of
the population concerned or simply reflect the structural biases of soci-
ety. These biases can then lead to inequities in the treatment of individu-
als and raise the question of the value to be given to the choices made by
the algorithms. For example, a recent study has shown that a risk predic-
tion algorithm used to define the amount of financial aid for health care
gave a lower score to blacks than to whites (Obermeyer and al., 2019). After
eliminating the discriminatory feature of the algorithms, the percentage
of black patients considered to be sick rose from 17.7% to 46.5%. This bias
was due to the fact that the algorithm predicts health care costs rather
than illness. Inequality in access to care means that black patients spend
on average less money on medical treatment than white patients. Thus,
while health care costs can be an effective indicator of population health,
significant racial biases can occur. This risk can also arise when AI is used
to quickly screen and select job applicants based on their characteristics
or career paths, as was the case with the system used by Amazon. 7 Thus,
ensuring the absence of bias in the data can require considerable human
and financial effort, both in terms of data collection and the subsequent
“on-the-ground” verification testing necessary for both learning and final
performance verification.

Finally, AI is capable of performing not only simple tasks but also com-
plicated tasks as long as these tasks are based on predetermined rules or
standards that are highly standardized in a mass of codifiable data. It is
therefore difficult for AI-based systems to deviate from standards or to
think for themselves. The learning dynamic is based on a routine mecha-
nism and limited to a very specific context. This limitation makes it dif-
ficult for AI to solve complex problems such as managing unpredictable
human behaviours, understanding people in all their complexity, show-
ing empathy, or performing several complex tasks at the same time. It is
no coincidence that the most important successes of AI operate mainly on
images that are very standardized in terms of digital content. In fact, it

7 See: https://www.theguardian.com/technology/2018/oct/10/amazon-hiring-ai-gender-
bias-recruiting-engine.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 63
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

would be more appropriate to talk about artificial learning than artifi-

cial intelligence. Finally, even if the current output of AI is based on the
availability of a large number of events (often several thousand) and on
significant computing power for learning, the results are not very gener-
alizable to other situations. Advances are still a long way from portending
the advent of so called “strong” AI, which would actually be comparable
to human intelligence in its ability to understand context and make use
of “common sense” in its ongoing capacity to learn. Such an achievement
seems well out of reach at the present time, as the researcher Yann LeCun 8
has emphasised, even if the generic character of the technologies devel-
oped so far is enough to give us a foretaste of future impacts on all sectors
of the economy.

Most of the literature in economics, including the substantial body of

research following the methodology of the seminal study by Frey and
Osborne (2013), has focused on how AI will substitute for different tasks
and in some cases for entire occupations. In what follows, we show based
on empirical evidence at the level of organisations and sectors, that the
effects are more complex and that the creation of new tasks and compe-
tences is as important as the elimination of old. This also varies across
sectors. This article, therefore, aims to make explicit the missing “comple-
mentarity” dimension in the analysis of the impact of AI on work trans-
formation.

To realistically illustrate the current and potential impacts of the diffu-

sion of AI, this section focuses on the diffusion of existing AI technologies
or the potential for their diffusion in the near future with a high degree of
certainty. It leaves unaddressed the question of the possible impact of rad-
ical advances in technology, including the ability to achieve “strong AI”.
In order to provide concrete examples of the possibilities being opened up
by artificial intelligence and its impacts on work, the analysis of three
sectors of activity—transport, banking and health—all of which provide
useful examples of the trends underway. Together, the three sectors cover

8 “As long as the problem of unsupervised learning has not been solved, we will not
have a truly intelligent machine. This is a fundamental scientific and mathemat-
ical question, not a question of technology. Solving this problem may take many
years, even several decades. Truth to tell, we know nothing about it”, Yann LeCun,
Computer Science and Digital Sciences Chair, 2016-2017, class at the Collège de France.

64 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

a range of realities in terms of types of production in both industry and

services as well as types of actors coming from both the private and public
sectors and from various institutional and regulatory contexts.

3. THE IMPACT OF AI ON WORK: A SECTORAL

APPROACH
The three sectors analysed are generally considered to be ones that are
highly exposed to AI. Transport has attracted attention due to the devel-
opment of the autonomous vehicle, which has been much publicised since
Google, followed by Waymo, Uber, Tesla, General Motors, Navya, and other
corporations announced their intention or started to test vehicles. The
already advanced degree of digitisation in the banking sector as well as
the intangible nature of the “matter” it handles involving the exchange
of information on transactions, has lent itself perfectly to exploitation by
AI. The health sector is one that everybody is concerned with and there is
great interest in the scope for AI to be used as a support in processing data
connected to highly complex life science-based mechanisms. These three
sectors are regularly referred to in France’s national AI strategy (Villani,
2018) and in the AI strategies of others advanced countries (OECD, 2019). In
addition, in terms of entrepreneurship, they are often leaders in the rank-
ing of sectors in terms of investment in start-ups (OECD, 2019).

The discussion focuses in on whether the use of AI in certain fields of

activity could replace human work or, on the contrary, will prove to be
complementary to human work. For each sector, the analysis identifies
first the main areas of AI application and then focuses in on how AI sub-
stitutes and complements for skills of different occupations or professions.
In relation to this it assesses how the transformation of work affects skills
and working conditions. The analysis of these three sectors is based in
part on the results of case studies and surveys conducted in the framework
of hearings with AI experts and professionals (company directors, manag-
ers, engineers, doctors, start-ups etc.) for France Stratégie, and described
in more detail in Benhamou et al. (2018). 9

9 See: https://www.strategie.gouv.fr/publications/intelligence-artificielle-travail.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 65
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

3.1. Work Transformations in the Health Sector

Health is often presented as one of the sectors where artificial intelligence

could produce major transformations in the workplace. Medico-technical
robotics is already important in biology, pharmacology and surgery. Some
observers go so far as to envision a “medicine without a doctor” that could
fundamentally transform medical practice and the organization of our
health care system (Vallancien, 2015).

3.1.1. Several fields of application are concerned by AI

Nearly all fields of artificial intelligence that have been developed, includ-
ing image and video recognition, natural language processing, automatic
learning, and robotics find applications in health care. This is true for
diagnosis and therapeutic recommendations, surgery with smart robots,
personalized follow-up, the medical-social field, rehabilitation but also
prevention and clinical research.

For example, in the field of medical diagnostics, there are many AI tools
on the healthcare market and the scope of their applications is impressive.
They can be found in medical specialties such as oncology, which covers
all medical specialties, studies, diagnosis and treatment of cancer, car-
diology, ophthalmology, radiology and the detection of neurological dis-
orders (e.g. Alzheimer and Parkinson). Whatever the field, the principle
is always the same. Algorithms fed and trained by massive data (medi-
cal image recognition, medical research results, etc.) are self-programmed
to detect pathologies. The emblematic tool is the Watson computer soft-
ware from the IBM industrial group, which was introduced to the health-
care market in 2005. Watson has been used in particular at the Memorial
Sloan Kettering Cancer Centre, an American institute specializing in med-
ical research and cancer treatment, to assist in diagnosis and therapeu-
tic proposals. This type of software, presented as an “intelligent” medical
decision support tool, synthesizes a mass of information from millions of
medical reports, patient records, clinical tests and knowledge (updates)
from medical research. Some software may soon be able to diagnose can-
cer as well as, or even better than, specialists. According to a recent study
(Wang et al., 2016), artificial intelligence has been able to automatically
detect breast cancer with a 92% success rate, almost equivalent to that of

66 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

specialists (96%). When the doctor’s analyses and the diagnostic methods
from the automated software are combined, the success rate is 99.5%, with
a greatly reduced risk of error.

Other medical specialties use image recognition to diagnose, for exam-

ple, patients with eye diseases. This is the case of the robot developed by
Google’s DeepMind Health (AI) division, which by looking at thousands
of photos of retinas, has been able to establish a more reliable diagnosis
than that of a human ophthalmologist (Knight, 2016). Other Ultromics sys-
tems use AI to diagnose heart disease 10 or neurological disorders such as
Parkinson’s disease. 11 In the field of surgery, new generations of surgical
robots are emerging that are moving towards greater autonomy from the
surgeon (Azad et al., 2016).

Although there are numerous examples of AI applications in the health

field, the promises remain unclear at this stage. There is no health system
or health organization in the world that has been completely transformed
and actual health applications are very limited. No large-scale deployment
closely or remotely related to AI exists except for a few very isolated cases
such as IBM’s Watson, Google’s Deepmind or smart robots. There is also
little scientific evidence today on the effectiveness of AI in the different
fields of application mentioned above. The most highly recognized interna-
tional academic journals in the medical field have published few articles
on the evaluation of AI. 12 There is also very little medico-economic evalu-
ation of AI applications that measures their economic and social returns.
The information we have today on the impact of AI on diagnosis is mostly
limited to subfields of a very specific discipline, such as oncology.

10 See: http://www.ultromics.com/technology/.
11 See: IBM Research (5 January 2017): http://www.research.ibm.com/5-in-5/mental-
health/.
12 See, notably, Journal of the America Medical Association, Delanceit, New England Journal of
Medicine and Annales of Internal Medicine.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 67
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

3.1.2. Medicine is characterized by a high level

of complexity that AI only imperfectly integrates

The hypothesis of a total substitution between health professionals and

the machine is at present more fiction than science. Medicine is character-
ized by a high level of complexity that AI can only imperfectly integrate.
Indeed, the quality of a patient’s care (disease detection, therapeutic pro-
posal, patient follow-up, etc.) is a highly complex process which is directly
linked to the existence of a strongly “mediated” relationship between the
medical team and the patient. In this field, the existence of a standard
patient model is generally not sufficient to develop and implement a man-
agement strategy that is totally adapted to the individual patient. It is also
known that AI feeds on a large mass of data via algorithms that aim to
establish correlations to explain phenomena as a basis for deriving clini-
cal recommendations. The robustness of correlations between several phe-
nomena depends in AI on the mass of data collected. The more important
the data, the more robust the correlations are. It is therefore the big data
that allows AI to function and compete with humans through its ability to
process data from a mass of continuously updated information. However,
correlation does not necessarily mean causality. The causal mechanisms
that “explain” the occurrence of a disease and its evolution are often more
complex than “mechanical” correlations. The causes may be multiple,
some of which are difficult to codify, such as those related to the patient’s
socio-demographic environment or the feeling of symptoms, or even their
total absence as has been the case with Covid-19. Moreover, there may
be a very high variance between patients. The complexity of the health-
care professional’s work is to take into account all these specificities from
detection to the therapeutic proposal to be delivered to patients. In the
clinical field, correlations are not sufficient. Even evidence-based medi-
cine, which should be based on the most up-to-date evaluation results in
the medical field, can only provide knowledge and clinical recommenda-
tions based on an “average” modelled patient. However, the “real” patient
who is diagnosed during a consultation with a doctor does not necessar-
ily correspond to the “average” patient who has been modelled. This is
also true for therapeutic proposals. The doctor (specialist or general prac-
titioner) has to take into account the specific characteristics of the patient
to define the best treatment, or even to “negotiate” with the patient for his
or her adherence to the treatment. Research has shown, that the patient’s

68 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

commitment to the treatment prescribed for a disease influences the heal-

ing process (Institute of Medecine, 2013). However, only dialogue and
trust between a doctor and a patient make it possible to negotiate the best
“cost-benefit” trade-offs and to ensure the involvement of a family carer
(Institute of Medecine, 2013).

At the current stage, AI developments are often concentrated on a single

pathology—for example Watson on cancer—as are the clinical guidelines
that are the foundation of evidence-based medicine. But what happens
when a patient presents several pathologies, which occurs with increas-
ing frequency. For the moment, according to the Director of Foresight and
Research of the French Hospital Federation, Antoine Malone, this central
question remains unanswered, while the performance of health systems
in the future will largely depend on the ability to “manage” polypahtolog-
ical patients over time 13.

In short, the quality of a diagnosis depends not only on the volume of

information available but on the quality of interpretation of complex
mechanisms that are not based on fixed natural laws and therefore are
not deterministic. Big data works well on simple and mechanical explan-
atory phenomena. But humans are constantly evolving with their envi-
ronment and this dynamic process means that in the field of medicine
and patient care, Artificial Intelligence will not be able to replace health
professionals. The engine of a car or an airplane is as complicated as its
design, and one can predictably know how an engine reacts in a given con-
text. This is not the case for a human being who has the ability to adapt
to changing environments and can respond in various ways to unpredict-
able events. The mere availability of information is not enough to influ-
ence behaviour. Obesity, for example, has become a major public health
issue in many Western countries despite the access to preventive health
applications via a smartphone, (Malone et al., 2020). Another difference
concerns social acceptability and responsibility. Unlike the use of AI for
predictive models in marketing and advertising, if there is a mistake in
medical diagnosis and treatment there may be human injury or death,
which raises major issues of social accountability.

13 Interview with Antoine Malone for France Strategie, 2018.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 69
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

The discourse to the effect that AI will lead to the disappearance of the
human practitioners in the field of health is not credible when one takes
into account all the complexity that the field of medicine covers. However,
this does not mean that AI is a “non-subject” in the field of health, nor
does it mean that certain professions—some of which are highly quali-
fied (e.g. radiologists)—are not threatened. We can say, however, that in
the coming decades AI will not replace the work of doctors and nurses and
there is little risk that that they will become mere algorithm executors.

What could be AI’s role if we are moving, as seems likely, towards a sys-
tem dedicated to the management/prevention of polypathologies with a
very strong psychological and behavioural component. What the system
will aim for will be long-term behavioural changes and this type of action
requires close and “human” interaction with patients (Alderwick et al.,
2015). What AI could bring is a better management of patients and iden-
tification of costs and adapted structures for global management (Malone
et al., 2020). Let us take the example of Kayser Permanent Washington
(2018) in the USA, recognized as an efficient health system in achieving
the triple aim of assuring quality of patient care, improvements of the
population’s health and these at the lowest possible cost (Foley et al., 2015).
These systems invest heavily in AI-based tools. But they also recruit a lot
of doctors, nurses, orderlies and even social workers to improve the man-
agement and prevention of complex pathologies, such as obesity or diabe-
tes because these pathologies are of socio-economic origin. They are also
developing local hospital structures within regions to be closer to patients
in order to influence their behaviour and minimize unnecessary inter-
ventions. So, it is likely that we will still have doctors and nurses’ assis-
tants in the future because the tasks that make up their jobs will not all
be open to automation. Indeed, it is more likely that there will be an even
greater need for healthcare professionals with the deployment of Artificial
Intelligence, if, thanks to big data and information gathering techniques,
prevention strategies adapted to each specific population can be improved
(Malone, 2018 and Malone et al., 2020).

70 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

3.1.3. AI will transform the way we work and interact

with patients

This section focuses on work transformation in a concrete way by identi-

fying the benefits but also the potential risks for medical workers them-
selves in several identified fields of application. One benefit is that AI has
the capacity to process millions of items of up-to-date data, thus generat-
ing gains in completeness, timeliness and the development of several pos-
sible hypotheses for disease detection. The physician will then have better
access to the information necessary for his decision-making including a
complete histology of the patient providing access to the best therapeutic
protocols based on evidence-based medicine.

If such “intelligent” expert systems gain the confidence of physicians and

the general public, the impact on the physician’s profession can be multi-
faceted. Physicians will be able to benefit from assistance in the manage-
ment of complex case diagnoses, with greater security in decision-making.
It will also lead to a strengthening of the doctor-patient relationship and
dialogue in order to provide the information necessary to understand the
diagnoses. The use of artificial intelligence will continuously increase the
level of “technical” competence of the physician as he or she is able to fully
exploit the most up-to-date clinical knowledge and medical practices. This
could result in a cognitive enrichment of the physician’s work, with less
time spent interpreting “routine” data that intelligent systems can han-
dle leaving the most complex expert cases to the physician. An increase in
skills will be all the more necessary as the doctor will also need to be able
to “challenge” the software and to explain the diagnosis and therapeutic
management responsibly.

These advantages, if they were to become widespread in the daily prac-

tice of physicians, would make it possible to combine human intelligence
and artificial intelligence in a complementary way in the service of better
decision-making, combining speed of decision-making and better optimi-
zation of patient care expenses. This is also the case in the field of surgery
with Star (for Smart Tissue Autonomous Robot), a robot that does not com-
pletely replace specialized surgeons but provides them with a tool capa-
ble of greater precision in performing certain procedures such as suturing
(Shademan et al., 2016).

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 71
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

3.1.4. The evolution of professions

Medical imaging professionals are particularly affected by the foresee-

able widespread use of automated image reading. If AI makes it possi-
ble to automate a part of conventional radiology for a specific field and
clinical situations, this will reduce the radiologist’s activity, despite the
growth in needs due to the aging of the population and the pervasive-
ness of chronic diseases. Ultimately, the question may arise as to the need
for the intervention of the radiologist in the establishment of the diag-
nosis. This automated step could be carried out by radio manipulators
trained in the development of diagnoses. It could also be carried out by a
non-radiologist equipped with tools for interpreting medical images. Such
developments presuppose an adaptation of the current regulatory frame-
work. Radiologists in such a scenario would reserve their time for the
interpretation of complex cases. In interventional imaging, on the other
hand, the need for their expertise will increase in almost all medical spe-
cialties. The profession of radiologist could evolve towards increased spe-
cialisation in interventional radiology for diagnostic purposes (punctures,
biopsies, etc.) for complex cases or for therapeutic purposes guided by med-
ical imaging. This evolution towards interventional radiology has already
been recognized by the profession, according to the National Federation of
Medical Radiologists 14.

In cardiology, new electrocardiogram (ECG) interpretation services are

being developed that rely on software rather than on the expertise of the
cardiologists. The software is capable of detecting rare or silent cardiac
abnormalities such as mitral leakage or cardiac arrhythmias which a spe-
cialist sometimes finds difficult to detect. This can provide the doctor,
whatever his or her speciality, with a gain in the quality of the diagnosis.
Such a service may transform the practice of the ECG by making its use
more frequent for more medical specialties (emergency physicians, gen-
eral practitioners, geriatricians, etc.) but perhaps also by opening it up
to non-physicians (nurses, firemen, etc.). It should free up the cardiolo-
gist’s time, if only through its ability to manage simple cases that will
be treated upstream without being referred to the physician. This could

14 Interview with Jean-Michel Masson, President of the Fédération nationale des

médecins radiologues (FNMR) as reported in Benhamou and Janin (2018).

72 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

refocus the physician’s activity on the most complex cases and lead to ben-
efits from continuous learning thanks to the software program’s capacity
to exploit the most up-to-date clinical knowledge and medical practices.
AI could accentuate the trend towards over-specialisation of professions
(coronographies, arithmology, cardio-paediatrics, etc.) by the increasing
digitalisation of certain medical devices. In cardiology, as in all fields, the
extent of the impact will depend on the quality of the tool. At the same
time, AI makes it possible to develop another type of service which con-
sists in interpreting ECG data collected over a long period (several days),
whereas these data were previously rarely collected. For the cardiologist,
it opens up the prospect of a new type of follow-up for his or her patients 15.

3.1.5. An intensification of work related to cognitive

exhaustion

AI has contradictory effects depending on its use in the same way as other
technologies do. If the gains made thanks to AI (speed of data processing)
result in an increase in the same proportions in the time devoted to cog-
nitive tasks (expertise, decision-making, solving complex problems, etc.),
this will leave little time for the doctor’s brain to “breathe” through alter-
nating between “stakeholder” and “little or no stakeholder” activities. If
the time “freed up” is only allocated to this type of task, we can thus iden-
tify a potential risk linked to what neuroscientists describe as professional
cognitive exhaustion to which healthcare professionals, and in particular
the youngest and least experienced ones, would be exposed.
a. Improvement of working conditions in the “care” professions

The care and monitoring of elderly patients suffering from one or more
chronic pathologies, particularly Alzheimer’s, is an extremely burden-
some task for the nursing staff. The management of these behavioural
disorders leads the staff to devote a great deal of time to them and makes
their working conditions particularly difficult (stress, feelings of profes-
sional inefficiency, burnout, etc.). These conditions often lead to the dis-
couragement of the health care team, even those that are experienced.
This is a field where intelligent social robots could intervene. Modern
robotics incorporating AI, whether in the management of sensors or in the

15 Interview with Yann Fleureau, CEO of Cardiologs for France Stratégie, 2018.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 73
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

programming of movements, is being developed in order to provide domes-

tic assistance for elderly or frail people. In Japan, robots are already being
experimented with to assist people in their daily activities, helping them
to move around, to move from chair to bed and vice versa. Residential
institutions for elderly dependents could be equipped with this type of
robot in a few years. There are already robots that provide solutions to
the problems of caring for elderly patients with cognitive and behavioural
disorders (Alzheimer’s, autism, etc.). These robots are equipped with arti-
ficial intelligence through numerous sensors and microphones that ena-
ble them both to interact with the caregivers who handle them and to
respond to the requests of elderly patients. Their physical presence and
their ability to interact socially through speech, facial expression and ges-
tures makes them ideal for working with people who have difficulty com-
municating verbally. Several international literature reviews in the field
identify enough acceptability of these robots, especially animal robots,
with show there are positive effects on the well-being of patients. The
Paro robot, which reacts to its name, compliments and touch, is the most
widely used robot in geriatric wards around the world. A study has shown
that the integration of these machines has reduced the physical and men-
tal strain on nurses and care assistants 16.

3.2. Work Transformations in transport sector

The major innovation brought about by the development of AI in the field

of transport will undoubtedly be the autonomous vehicle, even though the
timeline for its deployment remains uncertain. Of course, it all depends on
the degree of autonomy we are talking about, as automation of driving is
divided up into six levels from the international classification of the SAE
(Society of Automative Engineers). This classification defines what human
drivers and/or autonomous systems can and cannot do. It starts from level
0 which corresponds to no automation because all driving is done by the
driver up to level 5 which corresponds to the fully autonomous driving
of the vehicle in any situation (dense urban traffic, country roads, wind-
ing roads, etc.) which completely dispenses with a human driver. Level 5

16 See: http://www.mutualite-loire.com/index.php/nos-actualites/799-l-etude-inedite-
sur-les-usages-du-robot-paro-pour-des-residents-atteints-de-la-maladie-d-alzheimer-
ou-apparentee-en-ehpad-mutualiste.

74 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

has not yet been announced by any manufacturer, even though experi-
ments on open road traffic without a driver have been carried out since
2017 and in particular by two of the main manufacturers of autonomous
vehicles, Waymo, the Google subsidiary, and the French company Navya.
But according to John Krafcik, boss of Waymo, it will take decades before
level 5 is reached. Autonomous vehicles may still need a driver 17.

The technological maturity of the different levels and their subsequent

diffusion will therefore be decisive in the transformation of the trans-
port sector and the transformations of work and the level of employment
in this sector. Within 5 or 10 years, the degree of vehicle autonomy could
reach level 4, which corresponds to total autonomy but in very specific
contexts and where the surroundings are perfectly simplified and marked
out (moving and parking in a car park, driving on motorways). The dis-
cussion here, therefore, is limited to the impact of the diffusion of the
level 4 autonomous vehicle in a horizon of 10 years and only considers the
impact that artificial intelligence may have on road and rail transport,
which will be the segments most affected by the development of autono-
mous vehicles.

3.2.1. Application fields in transport sector

Other than the autonomous vehicle, AI applications in the sector mainly

concern the development of predictive equipment maintenance, logistics
and flow optimization. The use of industrial sensors is already widespread
to measure machinery wear points and equip production-chain control
points. Reduction in the cost of such sensors enables the collection of big
data. Artificial intelligence can process such data on a greater scale than
human processing can manage, so enabling the addition of more control
points and refining the diagnoses resulting from the analysis of such
data. This being so, companies can have smart diagnostic tools available
that facilitate maintenance operations and develop indicators prior to the
appearance of anomalies, which opens the way to predictive rather than
preventive maintenance (McKinsey Global Institute, 2015). Maintenance
and control operations are only carried out when required, before any

17 https://www.cnet.com/news/alphabet-google-waymo-ceo-john-krafcik-autonomous-
cars-wont-ever-be-able-to-drive-in-all-conditions/

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 75
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

anomaly occurs that might hold up a production chain, or before a piece

of equipment wears out. Predictive maintenance is of major interest to all
(rail and road) network and vehicle (aircraft, trains, heavy goods vehi-
cles, etc.) operators, as it optimises operations, limits downtime due to
maintenance and reduces servicing costs. According to the Mobility and
Transport Department of the General Commissariat for the Environment
and Sustainable Development, maintenance services may also be able to
anticipate and even avoid peaks of activity 18.

Artificial intelligence also enables optimisation of logistics in the event of

crises. When an incident occurs, mainline and metro train operations can
be seriously disrupted. When preventive maintenance has been unable to
avoid a crisis, its resolution may still be speeded up by artificial intelli-
gence. These days, crisis-scenario responses are standardised, with infor-
mation processing and coordination of required action as the two main
stumbling-blocks: two obstacles that AI can help remove. AI may well be
able to bring more specific and appropriate responses to crises by taking
more information into account. For example, in the event of a breakdown
on a metro line it will be able to take into consideration the number of pas-
sengers, which determine optimal speed for relieving congestion on the
line, the availability of replacement trains and the workforce required to
put them into service and the available alternative routes. Such optimisa-
tion of logistics and flows is only possible if AI is provided with real-time
data on a wide range of parameters, with all the risks of blockage that the
diversity of actors involved may generate 19.

It would seem possible for the applications described above to reach a level
of technological maturity enabling their deployment within the next
five to ten years. However, such maturity must also be able to respond to
various parameters affecting the dissemination of artificial intelligence
including the availability of massive data for the large-scale operation
of autonomous vehicles and respect for the privacy of individual owners
with the development of connected vehicles 20.

18 Interview with Marie-Anne Bacot for France Stratégie, 2018.

19 Interview with Hoang Bui, Head of Bureau of Transport Equipment, Mechanics and
Production Machinery, General Directorate of Enterprises for France Strategie,2018.
20 The CNIL has published a sector-specific reference framework enabling manufac-
turers to comply with the European regulation on data protection. This reference

76 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

In any future deployment of artificial intelligence in transport, issues of

data collection and exploitation will essentially be raised between com-
panies, and consequently will not concern the question of respect for pri-
vacy but rather that of sharing value. For optimal use of AI’s possibilities,
data on vehicle navigation and maintenance and on infrastructures will
have to be shared between several types of actors including, railway (SNCF
network and RATP in France) and road (Vinci, Bouygues, etc.) infrastruc-
ture managers and vehicle operators (Ouigo, RATP, road hauliers, etc.).
Economic and technical conditions with regard to harmonisation, quality,
interoperability, real time, etc. will therefore have to be clearly defined.

a. Impacts on jobs in the transport industry

The impact of AI on the transport professions will depend on the pros-

pects for the deployment of autonomous vehicles, which will themselves
vary according to the transport activity. As far as road freight transport is
concerned, the development of autonomous vehicles could threaten lorry
drivers.

Several reasons can be put forward. In Europe, the labour force in this
sector represents between 35% and 45% of total costs (International
Transport Forum, 2017), which can be a strong incentive for companies to
invest in technologies that promise savings. In addition, the sector is sub-
ject to strong international competition, which is an additional factor in
the penetration of innovations. The advent of level 4 autonomous vehi-
cles would allow automated driving in convoys on motorways, which is
an environment particularly suited to this type of vehicle. Also, on major
roads where many lorries travel, the formation of convoys makes it possi-
ble to reduce fuel costs while at the same time increasing safety through
the interconnection of vehicles 21. This automated traffic in convoys would
initially make it possible to increase truck driving times by modifying
the regulations on rest periods for drivers who would no longer be in a
driving situation throughout the journey. Eventually, the presence of a
driver could even be required only at the head of the convoy. A human

framework provides in particular for a scenario where “data collected in the vehi-
cle is transmitted to the outside to trigger an automatic action in the vehicle”. See:
CNIL (2017).
21 See: www.eutruckplatooning.com/About/default.aspx.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 77
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

presence would still be necessary to take over tasks that are not yet auto-
mated (e.g. refuelling).

Local drivers could be used to take the trucks to the highway or provide
local services. A new logistics system for transporting the trucks could
then be put in place as is already the case with the rail transport. Trucks
would be driven by drivers to an interface area at the entrance to the
highways before joining a self-contained convoy, and picked up at the exit
for delivery at the final point. Thus, the reduction in the need for long-
distance drivers would be accompanied by an increase in the demand for
local drivers, who would benefit from better working conditions (shorter
journeys in a restricted geographical area). If the number of lorry drivers’
jobs were to be threatened, jobs as controllers could be created to supervise
vehicle fleets from a distance (Benhamou et al., 2018).

In France, as in Europe, transport of private individuals is mostly by per-

sonal vehicles. Autonomous vehicles are unlikely to have any major impact
on this segment as the advent of level-5 automatic driving is still diffi-
cult to predict. However, level 4 should already enable the development
of new public transport services that might replace a percentage of indi-
vidual journeys. Initial experiments underway focus on shuttle services
travelling routes in delimited areas. Navya, a world leader in this area,
has already deployed over 50 vehicles across the world on short-distance
routes (up to two kilometres). We may therefore imagine that the coming
years will see increasing numbers of autonomous shuttles providing new
public transport services on local routes with fewer potential passengers
and not covered by present-day services such as night services. The Rouen
Normandy Autonomous Lab experiment on mobility-on-demand services
on the open road is in this direction (Transdev, 2017).

Such autonomous shared shuttles ensuring local services could comple-

ment the existing public transport offer, competing with mass public
transport and the transport of individuals by taxi or chauffeur-driven
cars, in which case there may well be an impact on drivers’ jobs in these
areas. However, as long as level-5 autonomous vehicles only remains on
the drawing-board, taxis and chauffeur-driven cars will remain the lead-
ing means of transport for door-to-door journeys. Moreover, the circula-
tion capacities of autonomous vehicles that take to the road will not be able
to rival classical means of public transport on the most frequented routes:

78 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

that would lead to unacceptable over-congestion of highways. These devel-

opments will also be accompanied by the creation of jobs for supervision
of fleets, as well as customer-relations positions responsible for passenger
reception and safety.

New work organisation for maintenance staff could also emerge. The smart
maintenance tools that will form part of the standard equipment for new
vehicles and infrastructures could also be installed on existing vehicles.
People responsible for service and maintenance will consequently be faced
with a need for new competences, both as regards the tasks being carried
out and the tools being used.

Smart tools will provide help and even “instructions” in both the diag-
nosis and performance of maintenance tasks. It is hard to contradict a
machine on the origins of a breakdown, especially if it has not happened
yet as will be the case with predictive maintenance. Artificial intelligence
will not just indicate the component that needs repairing, it will also indi-
cate how the repair is to be carried out. To borrow an image from the
medical sector, it will provide both diagnosis and treatment, implying the
risk of staff losing an overall vision of how a vehicle functions and the
maintenance operations to be carried out. This could lead to deskilling of
maintenance work, with humans being responsible for their performance
without necessarily any deep understanding. At present, such a risk seems
limited by a general determination to preserve employees’ autonomy with
regard to a vehicle’s overall maintenance and not focus on specialising on
specific tasks which may later be automated. Increasing skills is there-
fore essential if this overall approach is to be maintained, despite vehicles
growing complexity and dissemination of new smart tools. Lastly, work
pace at a maintenance centre is likely to be affected by predictive mainte-
nance, which will enable better forecasting of workloads as well as limit-
ing activity peaks (Benhamou et al., 2018).

3.3. Impacts in the banking sector

The banking sector has been a pioneer in the adoption of IT tools to man-
age customer databases and to set up networks for online banking. The
banking sector was also one of the first to implement “expert systems”,
computer programs designed to process technical transactions. In fact, the

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 79
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

AI solutions employed in the banking sector cover a wide variety of func-

tions and technologies, which can be divided into four categories: applica-
tions oriented towards customer relations, back office operations, trading
and wealth management applications, and applications oriented towards
regulatory use. Thus, the strong presence of structured data and the dem-
aterialised nature of the vast majority of transactions make the banking
sector a fertile ground for the development of artificial intelligence solu-
tions (DGE, 2019).

With regard to customer relations, the most developed artificial intelli-

gence applications are to be found in the field of credit risk rating. Banks
have historically developed an ability to analyse the risk associated with
any loan applicant using statistical models. These models are now enriched
by additional data sources that may require artificial intelligence process-
ing. Similarly, AI is used by insurers to improve the granularity of their
offers and recommendations (IAIS, 2017).

But the main field of application and the one with the greatest potential
to transform work in the banking sector is that of conversational assis-
tants or chatbots (Athling, 2017). A number of “back office” operations can
be related to the financial activities of banks, including risk modelling
and optimising the use of capital. As for applications in the field of wealth
management, they are gradually focusing on the analysis of weak signals
that can provide useful information for investments. Finally, in the regu-
latory field, AI applications are linked to the detection of irregular trans-
actions and can also be used to optimise customer knowledge mechanisms,
for example by using image recognition to automatically extract useful
information from the scanned image of an identity document.

3.3.1. Evolution of the advisor’s profession

The changes brought about by AI could profoundly transform the profes-

sion of bank advisor. In its study for the Observatoire des métiers de la
banque, Athling (2017) highlights as the most impacted activities those
relating to compliance with regulatory, legal and tax changes specific to
the banking sector. These activities will be improved thanks to more rel-
evant monitoring and more advanced and personalised recommendation
tools, such as that provided by the legal search engine Doctrine.fr, which

80 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

will allow “on-demand” access to this information. The construction of

client profiles using AI tools will also enable advisers to process credit
applications more quickly, or to identify financial risks such as tax fraud
or money laundering more effectively.

Independently of the development of AI, the announcement of its advent

combined with new consumer expectations are pushing banks to trans-
form their activities to reorient themselves around a service available
24 hours a day, 7 days a week. They intend to respond to the dual promise
of instantaneous and quality service by combining an AI-based service for
the request sorting and management phase of the most frequently asked
questions with a remote human service, available at all times. A form of
low-cost service could also emerge where the consumer would only have
access to automated help, even if it means paying an additional fee to
interact with a human. At Orange Bank, the solution set up to filter cus-
tomer service requests has achieved a recognition and comprehension rate
of around 80%. 22 Artificial intelligence is not always able to provide a rel-
evant response, which means that customer service agents process about
one out of every two requests in the end. The agents mainly assigned to
the operation of the platform or to solving technical problems encountered
during its use could gradually see a double effect on their job: a reduction
in the number of dedicated employees and an increase in the complexity
of the tasks remaining to be processed.

The increasing efficiency of AI in answering questions related to the

online banking platform, which is already the priority means of interac-
tion between customers and their bank, will in fact free up time and facili-
tate the work of these agents by filtering the number of requests. The bank
could then choose to train these customer service agents to respond to
requests that traditionally fall under the responsibility of the bank advi-
sor. This development corresponds to the expectations of the customer,
who increasingly sees his adviser not as the one who shares responsibil-
ity for managing his portfolio, but as an assistant who must help him nav-
igate through the complexities of the banking system, making himself
available to unblock a situation.

22 Interview with Emeric Chaize, Chief Digital Officer of Orange Bank, for France
Strategie, 2018.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 81
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

This field is also seeing the emergence of new players who have one of
the key resources for setting up an AI-based system. Integrated services
groups specialising in customer relations, which have access to consider-
able amounts of data as part of customer service operations on behalf of
third-party companies, will be led to set up a replacement offer managed
solely by a robot—“bot-shoring”—which could drastically reduce their
costs.

The availability of AI technologies that facilitate the banking advisor’s job

and reduce the volume of knowledge required by making them more avail-
able may also be an incentive for bank advisors to evolve towards greater
customer knowledge. Advisors could then take on more responsibility for
managing their clients, spending more time recommending investments
or sources of financing. In this scenario, social and decision-making skills
will be enhanced, and bank branches may be encouraged to focus on train-
ing in dialogue or negotiation skills.

Depending on the choices made by companies in the sector, artificial intel-

ligence may help to optimise the service and further the trend towards
dematerialisation, or it may reinforce the importance of the adviser by
giving him or her greater autonomy.

3.3.2. Transformation of support functions

The transformation of support functions within the banking sector is in

line with previous developments observed with the advent of the digi-
tal environment. With artificial intelligence, certain tasks, including the
most repetitive ones, are bound to disappear, particularly those related
to data collection, which will be optimised or accelerated. New working
modalities will emerge, where actors will have to learn how to interact
with the new AI-based system to help it progress (Benhamou et al., 2018).

As far as information systems are concerned, the arrival of methods

derived from artificial intelligence will not disrupt the organization. The
advances will be in line with the processes set up with “RPA” or Robotic
Process Automation—IT automation projects based on non-learning algo-
rithms, implemented since the 1990s and continuously developed. For other
activities, such as compliance activities, AI tools can lead to an upgrading

82 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

of transferable skills identified for example in the work of France Stratégie

where the bank’s employees demonstrate responsiveness, adaptability, as
well as office automation and IT skills (Lainé, 2018). Highlighting these
skills could increase their employability.

3.3.3. Evolution of skills specific to artificial intelligence

The needs specifically related to the IT development of AI in the banking

sector do not represent a particular stake in the total volume of employ-
ment in the sector. Nevertheless, they do require special adaptations that
deserve to be highlighted. The scarcity of skills pushes companies wish-
ing to implement solutions based on artificial intelligence to turn to spe-
cialized external organizations 23. Those who want to create these skills
internally will be forced to make major changes to their working envi-
ronment as AI projects require an extended phase of experimentation
and preparation which is similar to advanced research projects, requir-
ing scientific rigour and patience. Rather than focusing on internal skills
development, the major digital groups have chosen to open their research
departments to the outside world and to form partnerships with univer-
sities or research organisations. These transformations are a prerequisite
for the deployment of artificial intelligence.

4. CONCLUSIONS: LEARNING ORGANISATIONS

AND THE FUTURE OF AI
The objective of this article has been to use sectoral examples as a basis for
drawing more general lessons concerning the effects of AI on work. The
analysis has shown that not all tasks that make up the core jobs in the
three sectors analysed can be automated. Trades that draw their strength
from their human and social activities and mobilize skills that call for
creativity and complex problem solving will be preserved. This is the case
in particular for certain highly skilled occupations, such as doctors, or
low-skilled occupations such as care assistants or social workers. While

23 According to a survey carried out by EY in December 2017 among 200 AI profession-

als, 56% of whom estimated that a shortage of trained profiles was the main obstacle
to development of AI.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 83
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

in the transport sector the activity of driving may disappear in the long
run with the further development of the autonomous vehicle, more and
more supervisory tasks may also appear, thus transforming certain profes-
sions. This phenomenon is not new. Robotization in the automotive indus-
try is an old phenomenon which has led to the repositioning of workers
on supervisory tasks. This is also the case for the banking sector with the
digitisation of a large number of activities is leading to the evolution of
the professions providing advice to individuals. The article also points out
that AI has made it possible to carry out tasks that previously weren’t car-
ried out, either because they were too time consuming and tedious for
humans or because they were economically unprofitable. In the health
sector, for example, the analysis of electrocardiograms is a case in point
or in the banking sector the detection of anomalies in transactions using
AI-based devices.

In many cases AI-based devices appear to be used in ways that are comple-
mentary to the tasks performed by humans. Here, the human element in
the task is not eliminated and the person relies on a tool that helps him
or her: a tool to assist in diagnosis or therapeutic proposals in the medi-
cal field or a tool to support the customer advisor in the banking sector. At
the same time, alongside this process of complementing human skills, AI
can substitute for certain routine cognitive tasks. With technological pro-
gress, the scope for this sort of substitution is likely to increase, with AI
competing directly with a wider range of human cognitive abilities. The
skills associated with any task that follows pre-determined rules—simple
or complicated—risk being downgraded by future advances in AI.

For this reason, in order future-proof one’s position in tomorrow’s labour

market a major challenges will be to have the ability to learn continu-
ously and to develop new skills, especially those that are transversal to
the labour market (Benhamou, 2018). The sectoral examples point out in
particular that cross-cutting competencies—the ability to communicate
with others and influence decisions, the ability to transfer organizational
skills and know-how and the ability to manage hazards—will become
more important with AI.

In this light, learning forms of work organization that are based on a logic
of continuous learning could be particularly well suited to the challenges
posed by the integration of artificial intelligence. As the recent study by

84 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

Benhamou and Lorenz (2020a ; 2020b) shows, learning organizations are

based on the use of forms of work organisation that develop cross-cutting
competencies and support continuous employee learning. Adopting these
forms of work organization could be a relevant lever to ensure that AI is
used to enhance complementarity between machines and humans and not
simply to substitute for humans.

Another factor that argues in favour of the model of learning forms of work
organization concerns the innovation diffusion process itself. Indeed,
the principle of AI is to discover statistical regularities while remaining
“encapsulated” in an “expert” decision-making system that uses histori-
cal data. Paradoxically, this can favour a certain conservatism in human
decision-making. However, progress is not a matter of the past, but of cre-
ativity and risk-taking. This also speaks in favour of adopting a learning
organization design that encourages risk-taking and the development of
“systems thinking” to increase the organization’s ability to move beyond
“predetermined frameworks and norms” resulting from standardized pro-
duction processes. If, to the contrary, companies adopt more traditional
hierarchical forms of organisation there is an increased risk that AI will
be used mainly for purposes of substituting for human labour because the
production process in these organizational forms are based on a very high
degree of standardization and compliance with predetermined rules.

Thus, independently of the impact of AI on the level of employment, the

deployment of AI points to the need for a profound rethinking of work
organisation to support continuous learning capacities and the evolution
of skills. Much will depend on the competitive market strategy of the
enterprise and on the organisation of work that is adopted in its support.
While the use of learning forms of work organization contributes to the
competitive advantage of firms seeking competitiveness based on inno-
vation, the use of low-skilled employees with limited training and a low
capacity for learning contributes to the pursuit of competitive advantage
through cost-cutting. If AI is used in support of a logic of cost rationalisa-
tion, the virtues of training will be limited because human-machine com-
plementarity will not be sought.

The organisational challenges posed by the advent of AI are immense and

must be considered in a manner consistent with a nation’s education and
training system. In order to protect the labour market against the risks

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 85
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

of skills downgrading and obsolescence, it will be necessary to invest in

cross-cutting skills and to increase the capacity of individuals and firms
to continuously learn. While these forms of skills development can be sup-
ported through innovative policies at the level of the formal educational
and training system, they will also depend on the choices made by employ-
ers. Skills emerge in part from employees’ daily work experience and this
speaks to the need for complementary polices and frameworks designed to
promote the adoption of learning forms of work organization.

REFERENCES
ACEMOGLU, D., and RESTREPO, P. (2017). Robots and jobs: Evidence from US labor mar-
kets. National Bureau of Economic Research Working Paper 23285, Cambridge
MA. Retrieved from https://www.nber.org/papers/w23285.pdf
ALPAYDIN, E. (2016). Machine learning: the new AI. MIT Press.
ALDERWICK, H., HAM, C. and BUCK, D. (2015). Population Health Systems, Going beyond into
integrated Care, February 2015. The King’s Fund.
ARNTZ, M., T. GREGORY and U. ZIERAHN (2016). “The Risk of Automation for Jobs
in OECD Countries: A Comparative Analysis”, OECD Social, Employment and
Migration Working Papers, No. 189, OECD Publishing, Paris.
ATHLING (2017). L’IA dans la banque : emploi et compétences, Observatoire des métiers
de la banque, décembre.
ASSOCIATION INTERNATIONALE DES CONTRÔLEURS D’ASSURANCE (IAIS) (2017).
FinTech Developments in the Insurance Industry, 21 février.
BENHAMOU, S. (2018). The world of work in 2030: Four scenarios in Work in the
Digital Age: Challenges of the Fourth Industrial Revolution, Ed. Max Neufeind,
Jacqueline O’Reilly, Florian Ranft, Rowman et Littlefield Intl., London-New York
BENHAMOU, S. and JANIN L. (2018). “Artificial Intelligence and Work”, Report France
Stratégie to the Minister of Labour and the Minister of State for the Digital
Sector.
BENHAMOU, S. and LORENZ, E. (2020a). Les organisations du travail apprenante,
Enjeux et Défis pour la France, Avril 2020. Rapport de France Stratégie.
BENHAMOU, S. and LORENZ, E. (2020b). Promoting Learning Work Organizations:
Issues and Challenges for France, April 2020. Note de synthèse, France Stratégie.
BOWLES, J. (2014). The Computerization of European Jobs, Bruegel, Brussels.
BRYNJOLFSSON, E. and MCAFEE, A. (2014). The second machine age: Work, progress, and pros-
perity in a time of brilliant technologies. WW Norton & Company.
CHOLLET, F. (2018). Deep Learning with Python, Manning Publications, Sister Island.
CNIL (2017). « Véhicules connectés : un pack de conformité pour une utilisation respon-
sable des données », 17 octobre.
DIRECTION GÉNÉRALE DES ENTREPRISES (DGE) (2019). Intelligence artificielle – État
de l’art et perspectives pour la France, études économiques, rapport final,
février 2019.

86 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0
 ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

FOLEY, T. and F., FERGUS (2015). The Potential of Learning Healthcare Systems, The
Learning Healthcare Systems Projects: http://www.learninghealthcareproject.org/,
The Health Foundation, New Castle University and Institute of Health and
Society.
FORD, M. (2015). Rise of the Robots: Technology and the Threat of a Jobless Future. New York:
Basic Books.
FREY, C.B., and OSBORNE, M. (2013). “The future of employment”, Oxford Martin
Programme on Technology and Employment, Oxford University.
GRABER, M.L., and MATHEW, A. (2007). “Performance of Web-Based Clinical Diagnostis
Support System for Internist”, J Gen Intern Med 2007; 23 (S1): 37-40: https://
www.researchgate.net/publication/5754108_Performance_of_a_Web-Based_
Clinical_Diagnosis_Support_System_for_Internists.
GRAETZ, Georg, and Guy MICHAELS. “Robots at work.” Review of Economics and
Statistics 100.5 (2018): 753-768.
HELITO, C. P., HELITO, P. V. P., COSTA, H. P., DEMANGE, M. K., and BORDALO-
RODRIGUES, M. (2017). Assessment of the anterolateral ligament of the knee
by magnetic resonance imaging in acute injuries of the anterior cruciate lig-
ament. Arthroscopy: The Journal of Arthroscopic and Related Surgery, 33(1), 140-146.
INTERNATIONAL TRANSPORT FORUM (2017). Managing the Transition to Driverless
Road Freight Transport, Paris.
KAISER PERMANENT (2018). Kaiser Permanente Washington, Population Health,
Program Description, October 2018: https://wa.kaiserpermanente.org/static/
pdf/public/about/population-health-2019.pdf.
KNIGHT, W. (2016). “An AI Ophthalmologist Shows How Machine Learning May
Transform Medicine”, MIT Technology Review.
LAINÉ, F. (2018). « Situations de travail, compétences transversales et mobilité entre les
métiers », Document de travail n° 2018-03 et Note de Synthèse, France Stratégie,
Paris.
LIANG, A., PIROTH, I., ROBINSON, H., MACDONALD, B., FISHER, M., NATER, U.M.,
SKOLUDA, N. and BROADBENT, E. (2017). A pilot randomized trial of a companion
robot for people with dementia living in the community. Journal of the American
Medical Directors Association, 18(10), pp. 871-878.
LE RU, N. (2016). « L’effet de l’automatisation sur l’emploi : ce qu’on sait et ce qu’on
ignore » La Note d’Analyse, no.49, France Stratégie, Paris.
MALONE, A. (2018). « Vers des systèmes d’information hautement performants, enjeux
cliniques, enjeux de gouvernance, enjeux systémiques », in Hervé, C. Stanton-
Jean, M., (eds.) Innovations en santé publique, des données personnelles aux données mas-
sives (Big Data), Dalloz, p. 157-171.
MALONE, A., POLITI, C. (2020). « Pour un système d’information et d’organisation popu-
lationnel », French Hospitals Federation.
MCKINSEY GLOBAL INSTITUTE (2015). “The Internet of Things: Mapping the Value
Beyond the Hype”, June 2015.
MURO, M., R., MAXIM and J., WHITON (2019). Metropolitan Policy Program, Brookings
Institute.
OBERMEYER, Z., B., POWERS, C., VOGELI and S., MULLAINATHAN (2019). “Dissecting
racial bias in an algorithm used to manage the health of populations”, Science.
Vol. 366, Issue 6464, pp. 447-453.

R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0 87
ARTIFICIAL INTELLIGENCE AND THE FUTURE OF WORK

OECD Report (2019). “Artificial Intelligence in Society”, June 2019.

SHADEMAN, A., DECKER, R.S., OPFERMANN, J.D., LEONARD, S., KRIEGER, A. and
KIM, P.C. (2016). “Supervised autonomous robotic soft tissue surgery”, Science
translational medicine, 8(337), pp. 337-64.
SILVER, D., HUBERT, T., SCHRITTWIESER, J., ANTONOGLOU, I., LAI, M., GUEZ, A.,
LANCTOT, M., SIFRE, L., KUMARAN, D., GRAEPEL, T., LILLICRAP, T.,
SIMONYAN, K., and HASSABIS, D. (2018). “Mastering chess and shōgi by self-play
with a general reinforcement learning algorithm”, Science, Dec 2018: 1140-1144.
SILVERSTEIN, M.J., RECHT, A., LAGIOS, M.D., BLEIWEISS, I.J., BLUMENCRANZ, P.W.,
GIZIENSKI, T. and KUSKE, R. (2009). “Image-detected breast cancer: state-of-
the-art diagnosis and treatment”. Journal of the American College of Surgeons, 209(4),
pp. 504-520.
INSTITUTE OF MEDICINE (2013). Best Care at Lower Cost: The Path to Continuously Learning
Health Care in America. Washington, DC: The National Academies Press. https://doi.
org/10.17226/13444.
VALLANCIEN, G. (2015). La médecine sans médecin ? Le numérique au service du malade, Paris,
Gallimard, Coll. Le Débat.
VILLANI, C., “AI for Humanity: French Strategy for AI” (2018). https://www.aifor-
humanity.fr/en/.
VISSER, M., VAN ROODEN, S. M., VERBAAN, D., MARINUS, J., STIGGELBOUT, A. M., and
VAN HILTEN, J. J. (2008). “A comprehensive model of health-related quality of
life in Parkinson’s disease”. Journal of neurology, 255(10), 1580-1587.
WANG, D., KHOSLA, A., GARGEYA, R., IRSHAD, H. and BECK, A.H. (2016). Deep Learning
for Identifying Metastatic Breast Cancer, Beth Israel Deaconess Medical Center
(BIDMC) and Harvard Medical School.
WORLD BANK (2016). “Digital Dividends”, World Development Report. Washington D.C.

88 R E V U E D ’ É C O N O M I E I N D U S T R I E L L E ➻ N ° 169 ➻ 1 E R T R I M E S T R E 2 0 2 0