Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
learning
More science
than fiction
About ACCA
ACCA (the Association of Chartered Certified Accountants)
is the global body for professional accountants, offering
business-relevant, first-choice qualifications to people of
application, ability and ambition around the world who seek
a rewarding career in accountancy, finance and management.
Alan Hatfield
Executive Director, Strategy and Development
4
Contents
Executive summary 6
Introduction 8
4. Ethical considerations 25
Conclusion 37
Appendix 1 38
Appendix 3 47
DISCLAIMER
Parts of this report make reference to machine learning products or other initiatives from
third parties. This is done for information purposes in response to requests for real-world
examples. The report does not constitute an endorsement of the particular products or
initiatives mentioned or a complete list thereof.
Executive
summary
On average for any given term: 62% of respondents had not heard of it,
or had heard the term but didn’t know what it was or had only a basic
understanding, 13% of respondents had a high or expert level of
understanding. This suggests a lot of potential for greater education and
awareness building among the accountancy community around the world.
Essentially, ML involves the machine, over time, being able to learn the
characteristics of data sets and identify the characteristics of individual
data points. In doing so, it ‘learns’ in the sense that the outcomes are not
explicitly programmed in advance. They are arrived at by the ML algorithm
as it is exposed to more data and determines correlations therein.
6
Machine Learning: More science than fiction | Executive summary
The report begins with an introduction responsibility for some form of disclosure
As with any technology, to the basics. This is because it is to highlight when a decision has been
with power comes important to have some appreciation of made by a ML algorithm.
what these applications are doing, to be
responsibility. And in the able to trust such systems and to The report considers a range of ethical
case of machine learning, understand how machine learning can be considerations relevant to professional
a step towards developing a greater level accountants, using for guidance, the
ethical considerations are fundamental principles established by the
of machine intelligence.
never far away. International Ethics Standards Board for
In this context, ‘intelligence’ refers to the Accountants (IESBA).
ability of the technology, in certain
circumstances, to make decisions or draw The ability of AI to take over jobs is a
inferences, without there being an narrative often recited in the media. And
instruction to treat a given dataset in a there is certainly some truth about the
fixed, predetermined way. But it does not ability of these technologies to do a
mean that the technology has suddenly variety of tasks more efficiently – indeed,
developed an independent as mentioned above, this report
consciousness – this is not about robots specifically explores some of these areas.
going on the rampage!
But even sophisticated technology such
The market is recognising the power of as AI appears to struggle with the full
ML with 2 in 5 respondents stating that contextual understanding and integrated
their organisations are engaged with this thinking of which humans are capable.
technology in some way. This includes Despite advancements in AI, it does not
those who stated that their organisations yet appear to be the case that human
are in full production mode dealing with oversight can be done away with
live data (6%), advanced testing with completely; or that the technology can
‘go-live’ within 3-6 months (3%), early take into account human factors, such as
stage preparation with go-live within when building client relationships or
12 months (8%) and in initial discussions leading successful teams.
exploring concepts/ideas (24%).
ACCA’s work on the emotional quotient
Applications for adoption range across (EQ) strongly demonstrated the need, in
diverse areas, including for example, a digital age, for competencies related to
invoice coding, fraud detection, corporate emotional intelligence (ACCA 2018). In
reporting, taxation and working capital fact as we look ahead, the Digital
management. The report explores various Quotient (DQ) and EQ are best seen
products and initiatives across these areas. combined for either to be really effective
for professional accountants.
These findings reinforce the need for the
accountancy profession to prioritise Even outside behavioural areas such as
building awareness and understanding in leadership, core technical activities
this area, as organisations will increasingly require judgement and interpretation that
need these skills. In fact the biggest draw on multiple considerations. ML can
barrier to adoption cited in the survey provide truly insightful information, using
was the lack of skilled staff to lead the sophisticated algorithms to analyse
adoption (52%). historical data sets. But in some
situations, a human may choose to take
As with any technology, with power note of this but for perfectly valid
comes responsibility. And in the case of reasons, make decisions based on
ML, ethical questions are never far away. additional/other factors, that do not
Professional accountants need to consider, follows patterns seen in the past.
and appropriately manage, potential
ethical compromises that may result from Looking ahead, professional accountants
decision making by an algorithm. have an opportunity to develop a core
understanding of emerging technologies,
Who has accountability in this situation? while continually building their
What is the risk of bias, given that ML interpretative, contextual and relationship-
algorithms will inevitably reflect any led skills. They can then truly benefit from
bias in the data sets that feed them? the ability of technologies such as ML to
About 8 in 10 respondents were of the support them in the intelligent analysis of
view that organisations have a vast amounts of data.
7
Introduction
Machine learning (ML) is part of an umbrella of terms used when there is a reference to artificial
intelligence (AI), the latter term having been coined as far back as 1956.
Most early AI work relied on a ‘decision So what has caused this? It is worth interrogating this observation.
tree’ approach to mapping options, for Data-driven insight is at the heart of the Broadly speaking, there are two levels of
example, in chess, mapping all possible ‘intelligence’ driving AI. And it is the AI – specific or weak and general. As it
opening moves and subsequent counter- exponential increase in the availability of currently exists, the term ‘AI’ refers to weak
moves. With even relatively simple data and unprecedented computing AI. This means the use of AI in solution-
problems, such as a retailer making power for processing this data that have specific applications, for example in
customer-specific recommendations, the jointly contributed to moving AI identifying patterns within a large volume
vast number of options in a decision tree increasingly from fiction to fact. of transactions. What is not currently
led to a combinational explosion that possible is artificial general intelligence –
could not be processed by even the most the sort of AI often depicted in films and
capable hardware. television, with robots displaying human-
This created a series of disappointments like intelligence and characteristics.
about AI, a so-called ‘AI winter’, where While there are some who believe this
computing capability lagged behind latter type of so-called ‘sentient’
theoretical approaches and fell understanding may one day be possible,
significantly short of hopes for the current technological reality appears to
creation of usable applications. In recent be far away from this. As many experts
years, however, AI has enjoyed renewed have noted1, high-performance adult-
interest. This is not science fiction; rather level intelligence for a single activity, such
it is now increasingly found in consumer as needed for playing chess, can be
technologies and business applications. easier to model than human mobility or
perception – even that of an infant.
1 Referred to often as Moravec’s Paradox, the discovery by artificial intelligence and robotics researchers Hans Moravec, Rodney Brooks and Marvin Minsky in the 1980’s that, contrary to traditional
assumptions, high-level reasoning requires very little computation, but low-level sensorimotor skills require enormous computational resources.
8
As a finance professional it is important • Andreas Georgiou, Sage • Ruth Preedy, PwC
to develop an appreciation of all this, • Dorothy Toh, King’s College London • Shamus Rae, KPMG
given that machine learning is being
increasingly used in accounting software • Lisa Webley, University of Birmingham • Stuart Cobbe, Brevis
and business process applications. • Maria Mora, Fujitsu • Thomas Toomse-Smith, Financial
This report aims to aid the process of Reporting Lab.
developing this understanding.
9
1. Machine learning
and accountancy
Double-entry accounting traces its roots to the medieval period, and from that time onwards it
has served as the worldwide basis for business record-keeping. The business processes by which
those records are created, and by which independent auditors evaluate the accuracy and
completeness of those records, have evolved over time.
Despite this, an accountant from the late ML is capable of many amazing things staying current. The capabilities that
1500s and one from the late 1900s would but do accountants really have a need for machine learning offers could assist the
have had enough assumptions in common, any of those amazing things to do the job work of professional accountants in
linked to the double-entry approach, to well? On the whole, the answer appears various ways over time. One of the key
allow them to have a professional to be ‘yes’, and this is not just a matter of drivers of this is the proliferation of data.
conversation in a meaningful way.
problems, rather than actual solutions to a Note: remaining respondents said ‘Equal hype and reality’
specific set of business problems or needs.
10
Machine Learning: More science than fiction | 1. Machine learning and accountancy
It is estimated that around 90% of all the This rapid growth in the volume of
It is estimated that around digital data in the world has been created financial transactions, if not properly
90%
since 20162. And the rate at which new managed, could pose a threat to the work
data is being generated is not just of accountants. For auditors, this may
growing, but appears to be growing relate to the sample they need and its
exponentially, rather than in an ability to be representative of the
incremental or linear manner. population, enabling them to form
conclusions that can be generalised
It is fair to point out that not all this data is beyond the sample.
of all the digital data necessarily of interest to accountants. But
even looking at areas of more obvious As referred to by Forbes5 and others, the
in the world has been interest, such as financial transactions, the volume of transaction data is estimated
created since 2016 trend towards increasing amounts of data to grow significantly between now and
remains relevant for various reasons. 2025. So, there will be a need to deal with
• In much of the world, digital methods orders-of-magnitude more data, rather
are rapidly replacing cash as the than incremental increases, and a need to
preferred way of paying. In China, for understand the distribution and profile of
instance, mobile payments are rapidly this significantly enlarged pool of data.
reducing the relevance of carrying cash3. An implication of this will be pressure
• Internet of Things (IoT) devices, on current resources and the ability to
streaming services and transactionally scale-up procedures reliably to
priced cloud-based hardware and understand the population being
software solutions have led to the assessed, for example to deal with larger
growth of small-value, high-volume sample sizes. But in fact technology like
financial transactions. machine learning could go beyond that
• The success of financial inclusion with the possibility for reviewing entire
initiatives around the world has led to populations to assist the auditor to test
many more participants in the global for items that are outside the norm.
financial system. From 2011 to 2018, Such developments may make ML a
over 1.2bn people entered the financial matter of necessity rather than just
system for the first time, and each of competitive advantage; as the latter
them is a source of financial transactions will reduce anyway, when many in the
that did not previously exist4. market start to adopt it.
140
120
Zetabytes
100
80
60
40
20
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025
Source: IDC Global DataSphere, November 2018
2 https://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/#4d881f9a60ba
3 https://www.pymnts.com/cash/2019/mobile-payments-cash-china/
4 Global Findex database
5 https://www.forbes.com/sites/tomcoughlin/2018/11/27/175-zettabytes-by-2025/#1b6ead315459
11
2. Navigating
the terminology
AI is often used as an over-arching term for advanced computing capabilities, with machines
being able to ‘think’ for themselves. And as mentioned earlier, specific or weak AI is the current
reality, as opposed to artificial general intelligence. Such nuances can be useful to bear in mind,
when sifting through the range of terms involved.
The challenge is that there is no definitive forecasting future sales on the basis of their well as in ML literature. Nonetheless, data
industry standard or agreed definition of dependency on an underlying driver might analytics is generally seen as a task that is
exactly what each of these terms means. involve the use of a simple linear regression. controlled and led by explicit human
This can result in confusion and instructions. The more advanced use of
differences of opinion, making attempts The schematic in Figure 2.1 shows an these techniques (and others) on large
at definition a minefield. Nevertheless, it overlap between data analytics and data sets, which can eventually enable a
is helpful to have some view on these machine learning. This is to represent that machine to function, in some sense,
matters, particularly for those new to the there can be some overlapping of the without explicit instructions, for example to
field, and the schematic shown in Figure techniques used, for example regression draw inferences, is generally a characteristic
2.1 represents one attempt at this. exists both in data analytics literature as more closely associated with AI/ML.
12
Machine Learning: More science than fiction | 2. Navigating the terminology
Robotic process automation (RPA) has Deep learning (DL) and Natural Language
RPA is in fact a piece of been placed outside the AI circle in Processing (NLP) are generally thought of
programmed software Figure 2.1. This is because, despite the as being within the ML family. They can
word ‘robotics’, RPA does not refer to handle more complex data, including
that implements a robots in the sense of the human-looking unstructured data, such as images. This can
defined sequence of intelligent robots sometimes depicted in allow for greater complexity of patterns
activities – like a very the media. RPA is in fact a piece of that can support, for example, image
programmed software that implements a recognition or speech recognition. These
high-end Excel macro. defined sequence of activities – like a are briefly discussed later in this section.
very high-end Excel macro. There isn’t an
AI element in this and it is, at its heart, Finally, and more generally, a term that
process automation: in other words, can come up during references to AI is
taking a defined process and repeating it ‘cognitive technologies’. It is relatively
tirelessly, quickly and without errors. difficult to agree a definition for this term,
which can refer broadly to technologies
While this section discusses these terms that seek to replicate the way the human
as static entities, it is worth noting that brain processes/interprets information.
this can be simplistic because these
technologies are moving rapidly. One of the criticisms levelled at AI as a
Innovations across different technologies term is that it is frequently used to refer
do not happen in isolated groups. to technologies that are expected to
arrive 5–10 years in the future, and that
One area of emerging innovation is the they permanently remain 5–10 years in
combination of RPA and AI elements, the future! In reality, technologies are on
so-called intelligent process automation a continuum of evolution, where they
(IPA). This is increasingly being explored acquire more ‘intelligent’ characteristics
by various technology companies, eg over time as the technology evolves. And
Alibaba via its Aliyun Research Centre. often, once a capability is realised and
IPA is a form of standard robotic process becomes mainstream, the AI label gets
automation (RPA) in which the system can embedded into business-as-usual
learn over time from the data and technologies and processes.
processes on which it is working. With this
element, over time, IPA might provide Increasingly, ML techniques are being
opportunities for process improvement as buried deep in applications and websites,
much as process automation. replacing traditional software in ways that
may not be obvious or visible. An example
Coming back to Figure 2.1, ML is a is Uber’s pricing system. Where 10 years
sub-set of AI that is generally understood ago this would have been hard-coded
as making predictions or decisions from logic, a trained model now makes these
the analysis of a large historical dataset. decisions. It looks nothing like artificial
Essentially, it involves the machine, over general intelligence, but it performs a
time, being able to learn the specific task to great accuracy. Viewed
characteristics of data sets and to identify from the outside, the embedding of this
the characteristics of individual data AI software creates an increase in the
points. This allows it to identify operating effectiveness of the whole – a
relationships in complex and large data cost-saving development even if not a
sets that would be more time consuming radical change.
or more difficult for a human to see. An
ML system can be said to ‘learn’ in the A well-documented example of AI that has
sense that over time, as it is fed more become ‘normalised’ is optical character
data, it can improve its recognition of the recognition (OCR), ie the ability to extract
patterns therein, and apply this improved text from scanned copies and documents.
recognition to new data sets that it may The traditional method involved a
not have seen previously. Machine rules-based template that had to be set
learning is increasing in its relevance as a up in advance, with the system extracting
tool for business use and is discussed in and mapping the patterns to the text in
detail later in this section. line with that template. Templates easily
become complex, for example to cope
with data tables or even text in columns.
13
Machine Learning: More science than fiction | 2. Navigating the terminology
The AI-driven leap here has been to This is also important in order to have
On average, for each of remove dependence on the rules-based an appreciation of how ML relates to,
the terms considered, template; in other words for the AI to or differs from, other terms often
create its own mapping between the mentioned in this area. In the survey,
about one-third of layout and the text or character to which ‘data analytics’ was the best understood
respondents either had it should be mapped. As this has become with only one-fifth of respondents stating
not heard of that term, or more common, however, it is generally they were not sure about how it differed
thought of as just ‘OCR’ and the AI- from ML (Figure 2.3).
had heard of it but did not enabled back-end is forgotten.
have an understanding In ML one is dealing with a powerful tool
Among the respondents to ACCA’s with tremendous potential. This is
beyond that.
survey, the understanding of certain because AI encompasses an enormous
terms was much greater than that of range of applications. These include
others. On average, for any given term, recommendation engines; fraud
one-third of respondents had either not identification; detecting and predicting
heard of it, or had heard of it but didn’t machine failure; optimising options-
know what it was (Figure 2.2). trading strategies; diagnosing health
conditions; speech recognition and
While professional accountants may not translation; enabling conversations with
need to develop ML algorithms chat-bots; image recognition and
themselves, this section will provide an classification; spam detection; predicting
introductory sense of how ML works in everything from how likely someone is to
the background. This matters because it click on an advertisement, to how many
influences trust – and the ability to have new patients a hospital will admit;
a view on whether one can trust the through to autonomous vehicles.
decisions of these systems and the
contexts in which they operate.
50%
AVG. 11% 20% 31% 25% 12% 1%
40%
30%
20%
10%
0
I’ve never heard of this I’ve only heard the term Basic level Medium level High level I’m an expert
of understanding of understanding of understanding
FIGURE 2.3: For each of the terms below, state if you are not sure how it differs from or relates to machine learning
40%
35%
34% 34%
30%
25%
26%
20%
20%
15%
10%
5%
0
Natural Language Processing Robotic Process Automation (RPA) Artificial Intelligence Data Analytics
14
Machine Learning: More science than fiction | 2. Navigating the terminology
15
Machine Learning: More science than fiction | 2. Navigating the terminology
6 It is important to know how to recognise excessively large additions to the data sets that do not add any incremental value and that result in ‘over-fitting’.
16
Machine Learning: More science than fiction | 2. Navigating the terminology
17
3. Applications of
machine learning
There are a variety of applications for ML and this section gives a flavour of some of these.
As might be expected, there is a spectrum of ways in which ML can be adopted.
FIGURE 3.2: How comfortable would you be with machine-learning-based decision making on the following specific tasks?
n NET: Comfortable n NET: Not comfortable
60%
50% 53%
40% 47%
43% 41% 40% 39%
30% 36%
31%
27% 27% 25%
20% 24%
19% 21%
10%
0
Classifications of transac- Accounting Decisions on Fraud detection Recruitment short-list, Medical/health related Financial decision
tions and/or assets and measurement audit testing ie deciding suitability decision, for diagnosis about you, for
liabilities for accounting to call for interview investment planning
and tax purposes
Note: 1–5 scale with higher number indicating greater comfort; NET Comfortable is sum of 4, 5; NET Not Comfortable is sum of 1,2
18
Machine Learning: More science than fiction | 3. Applications of machine learning
FIGURE 3.3: The main barriers to using machine learning in respondents’ organisations
60%
50%
52%
49%
40%
30%
20% 24%
21% 19% 17%
10% 14%
11% 4%
0
Lack of skilled Cost Poor quality No clear benefit Don’t know/ Insufficient Regulatory/legal Ethical Other
staff to lead implications of data from using no barriers volume of data requirements dilemmas
the adoption machine learning
19
Machine Learning: More science than fiction | 3. Applications of machine learning
recommendations, and more than 250m Both the Invoice Coding and Bank
In this risk assessment, bank reconciliation recommendations. Reconciliation models are based solely on
supervised learning Xero estimates that with 800,000 invoices the experience of the specific business,
filed each day in Xero this is a collective not on those from a wider pool of
algorithms can be used saving of 307 hours. entities. This naturally limits the degree of
to help identify specific ‘intelligence’ demonstrated, and prevents
On coding of invoices, the Xero software the software from applying pre-built
types or characteristics ‘learns’ how a business codes regular knowledge to new customers. The
that warrant greater items and auto-fills on the basis of this company recognised the challenges with
scrutiny; and improve ‘understanding’ of history, rather than the this, early on: ‘It’s true that there is
labour-intensive traditional use of default potential to learn from other organisations
targeting of the areas
codes. Using this approach, it correctly as well, but our early research has shown
of focus for the audit. codes 80% of transactions after just four that there is huge variation in practice and
examples. The company’s blog post encoding between different businesses
suggests that it is using a logistical – far greater than we expected’.
regression approach to get the best
prediction but, understandably, for This kind of standardisation is envisaged
competitive reasons details of the as a future enhancement as it can lead to
predictive algorithms are not available. further efficiency improvements in customer
activity, but highlights the challenge in
According to Kevin Fitzgerald, Asia creating an ‘intelligent’ coding bot.
Pacific Director for Xero:
IMPROVING FRAUD DETECTION
‘We see machine learning algorithms
One of the areas where ML can help is with
being helpful in providing intelligent
risk assessment. The reference here is to
support that can free up the time of
the ability to assess the likelihood of fraud,
professional accountants to focus on the
inaccuracy, misstatement, etc. based on a
financial and strategic agenda of their
mix of empirical data and professional
clients or their own organisations’.
judgement. In this risk assessment,
When initially implemented, these supervised learning algorithms can be
codings were provided as suggestions to used to help identify specific types or
the user, and required specific, albeit characteristics that warrant greater
easy, validation or correction if necessary. scrutiny; and improve targeting of the
Xero deliberately did this so that the areas of focus for the audit. In this
algorithm would learn user behaviour. context, the choice of an appropriate ML
The company has stated: ‘We’re watching method can be valuable for audit testing.
very closely the rate that customers Using ML as part of the audit process is
actively disagree with suggestions by in relatively early stages, and publicly
choosing something else, and the rate of available empirical data to support the
later recodes of suggested accounts. On assertions of improvement are being
recodes, the system absolutely learns steadily built over time. One example is
from those. It’s part of the basic idea – it a study commissioned by the Comptroller
only knows what it’s been taught. If it and Auditor General (CAG) of India (Yao
learns from correct accounts, the et al. 2018).
suggestions will be correct’. This goes
beyond a static rules-based approach to CAG is an independent constitutional
a true ML capability. body of India. It is an authority that audits
receipts and expenditure of all the
For bank reconciliations, the Xero ML organisations that are financed by the
software integrates with that of many government of India. One of the CAG’s
banks, which feed account transaction duties is to uncover organisations set up
records automatically into Xero. It then for fraudulent reasons. In fulfilment of this
matches bank transactions with payment duty, each year it selects a number of
and receipt records in Xero, with organisations to be audited. Some are
automated coding based on how similar selected via public complaint or direct
transactions have been previously coded. referral, while others are selected by
As with invoice-coding, the ML for bank monitoring news sources and business
reconciliation incorporates user results but, historically, a significant
modification to transaction matching to number are selected by random sample.
improve recommendations.
20
Machine Learning: More science than fiction | 3. Applications of machine learning
The rationale for this weighting was that a One of the interesting features of the
false positive merely triggered a human CAPS model is that it optimises results not
investigation, which would presumably only in relation to likelihood of detection
reveal that a firm was indeed genuine, but also to the return on investment (ROI)
while a false negative allowed fraud to of the program itself. In other words,
continue undetected. information about staff availability and
the cost of a compliance investigation are
In aggregate, the most accurate inputs into the training set, and the
algorithms were able to identify suspicious predictive outputs are not only the
firms correctly 93% of the time. The likelihood of fraud but also the projected
reported results were quite detailed, but in expected value of any potential recovery.
summary, of the 10 different ML methods
tried in the study, no one method proved In 2017, Intel published a white paper
to be the most accurate across all that summarises the findings across the
transaction types and industry groups (Yao five years that CAPS has been running in
et al. 2018). Therefore, understanding production. There are some noteworthy
what algorithm to use and why is findings. As a control, Intel continued to
extremely important. These findings perform some compliance audits by
demonstrate not only the potential value random selection. The dollar value of
that ML techniques can add to the audit recoveries remained the same over the
process, but also the importance of five-year period; in other words, they scaled
having a sufficient understanding of ML with the capacity of the audit team and not
techniques to be able to select the most with Intel’s revenue growth. On the other
appropriate methods for specific instances. hand, in 2012, when the study started, the
dollar value of recoveries from CAPS-
While the above example relates to triggered audits was nine times that from
government and is relatively recent, there randomly selected audits. Over the five-year
are earlier examples of private companies period, the supervised algorithm continued
experimenting with ML. Intel, for self-training and, in 2017, CAPS-triggered
example, established ‘Intel Inside’, a recoveries grew up to 19 times those
cooperative marketing campaign in which generated from randomly selected audits7.
7 https://www.intel.ai/nervana/wp-content/uploads/sites/53/2018/05/AI-Optimizes-Intels-Business-Processes.pdf
21
Machine Learning: More science than fiction | 3. Applications of machine learning
8 https://www.ibm.com/watson/stories/kpmg/
9 https://www.cbsnews.com/news/irs-cant-do-the-math/
22
Machine Learning: More science than fiction | 3. Applications of machine learning
23
Machine Learning: More science than fiction | 3. Applications of machine learning
10 https://kirasystems.com/resources/case-studies/deloitte/
11 https://www.ey.com/en_gl/better-begins-with-you/how-an-ai-application-can-help-auditors-detect-fraud
12 https://home.kpmg/xx/en/home/insights/2017/09/strategic-profitability-insights.html
13 https://www.pwc.com/gx/en/about/stories-from-across-the-world/harnessing-the-power-of-ai-to-transform-the-detection-of-fraud-and-error.html
24
4. Ethical
considerations
Ethical behaviour is a necessary attribute for everyone in society, in both their personal and
professional dealings. But for the profession this element is additionally hard-coded into the
very definition of what it means to be a professional accountant. And within organisations, it is
a key requirement that the finance function provide constructive challenge to ensure that
business decisions are grounded in sound ethical principles.
The IESBA (International Ethics Standards The ethical challenges posed by ML are DEALING WITH BIAS
Board for Accountants) Code sets out five explored in this section by focusing on
This is arguably the most frequently
fundamental principles of ethics for five areas. For each area, a scenario is
discussed source of ethical challenge. At
professional accountants, which establish examined where the IESBA fundamental
its root is the fact that ML algorithms, both
the standard of behaviour expected of a principles could be compromised. In
supervised and unsupervised, may need
professional accountant (see Appendix 1). most scenarios most of or all the
to be properly interpreted in order to
So when considering the potential of ML, principles may be at risk but, to draw out
avoid confusing correlation with causation.
professional accountants need to think specific points, only one or two
not only of the potential benefits – as compromised principles may be A case in point is algorithms that assess
demonstrated by the preceding section highlighted. For those interested more recidivism risk. These algorithms
on use cases – but also the ability to broadly in digital ethics, beyond ML construct a profile of convicted
create long-term sustainable advantages. specifically, ACCA’s report on Ethics and defendants and provide a score that is
This latter aspect depends in no small trust in a digital age also addresses said to represent the likelihood that one
way on ensuring that ethical relevant considerations (ACCA 2017). will be a repeat offender. As with medical
considerations are given sufficient diagnosis solutions, these are decision-
emphasis when exploring ML adoption. support tools. Therefore, the sentencing
decision still remains with the judge. But
Trust can take years to build and an instant
the increasing reliance on scores that
to be destroyed. Clearly ethical behaviour
these algorithms generate may create
is a non-negotiable requirement for its
pressure on judges, who may be
own sake. Nonetheless, it is also clear
perceived as ‘soft on crime’ if they
that breaching best practice in this area
impose a lesser sentence than is
can inflict real damage on the brand/
indicated by such an algorithm.
reputation and intangible value of an
organisation. In today’s social media-driven In theory, these algorithms are free of
world, bad news circulates quickly, and racial bias, as the defendant’s race would
not paying attention to ethical behaviour not be included in their training set. But
as new technologies are adopted can these training sets are based on historical
expose organisations to significant risk, data, and this data is informed by which
both financial and reputational.
25
Machine Learning: More science than fiction | 4. Ethical considerations
communities and groups have been more whether they have been biased in favour
Data is the single most involved with law enforcement in the of assuming the outcomes are valid
important and non- past. In turn, such communities may be merely because they are supported by
the least likely to have jobs, access to an ML algorithm.
negotiable requirement higher-quality education, health care, and
for powering the use other such variables where racial bias may On a different note, the principle of
have been empirically proven to exist. The professional behaviour requires
of ML. In order to take compliance with relevant laws and
result is that despite having no inherent
advantage of data in racial bias themselves, these algorithms regulations. If there is evidence of
a sustainable way, an can make even more systematically systematic bias, then the organisation may
biased decisions than the humans they be in breach of certain laws. For example,
organisation needs a
have been designed to support: they regulations such as the European
coherent data strategy. ‘learn’ racial bias from the data. Directive [2004/113/EC] modified in 2012
are targeted to disallow gender bias.
The issue behind such bias can extend
even before initial convictions are made, Professional accountants may face
and not just for repeat offenders. Here internal pressure to ignore the issue,
the algorithms, with their base of such as if it is possible to argue a lack of
historical data, may unwittingly end up evidence and that in a statistical approach
answering the wrong question – not the the answers will be correct over a period
likelihood of being guilty but the of time. Accountants may need to play a
likelihood of being arrested. role in, for example, guiding colleagues
to reassess the model with a different
Scenario emphasis on the gender variable. It will
An ML model for improving the be important to maintain clear trails of
prediction of loan default was trained on communication to management, with
all the historic data available on documentation of details, responses
applications, approvals and defaults. The received and, if appropriate, escalation
model was tested against a sample of to relevant authorities. Also key to this
historic data and shown to have high is a basic appreciation of the inputs and
accuracy in predicting default. A review outputs associated with the model and a
by an underwriter of a sample of view on the metrics and key performance
applications and decisions was indicators. This may be required when
conducted before sign-off for live use. gathering feedback and monitoring for
issues, such as customer complaints,
Several months into the process, a clear as a leading indicator of problems.
pattern emerged that women were A questioning approach, rooted in
significantly over-represented among those professional scepticism, and a growth
whose loan applications were rejected. The mindset willing to grapple with new
underwriter investigated further and found challenges, will both be important to avoid
a number that should have been approved. being overawed or afraid to dig deeper.
The suspicion is that the model was biased
against female applicants because it was STRATEGIC VIEW OF DATA
based on several decades of historic data
and this training set had a lower proportion Data is the single most important and
of sole applicant females. So the model non-negotiable requirement for powering
was biased to reject more loan applications the use of ML. In order to take advantage
from this group. of data in a sustainable way, an organisation
needs a coherent data strategy. In
For the accountant the fundamental practice, this means several things.
principle of objectivity could be
compromised in relation to issues of bias. The first is just the collection of a
The reference here is the avoidance of sufficient amount of data. Any meaningful
compromise of professional or business insight with low likelihood of bias
judgements because of bias, conflict of depends on having enough data across
interest or undue influence of others. all the categories/types that may need to
The accountant may have to consider be considered. The amount of data and
26
Machine Learning: More science than fiction | 4. Ethical considerations
range across categories is what enables The third step is continuous maintenance
All the hard-won benefits the algorithm to identify relationships of data quality. This is extremely important
of an otherwise coherent that might not be obvious to a human. because algorithms do not have a
standalone intelligence of their own: it is a
data strategy can be lost For professional accountants, ‘data’ can case of ‘garbage in–garbage out’. So if the
through breaches that mean more than just financial data; it quality of data feeding the model is poor,
might require customer data held in other the insights will be incorrect. A strategic
undermine the security of departments outside finance, as well as approach to data quality means having
data in the organisation. external data. This data-gathering robust processes to enable frequent data
process can require coordination across cleansing, not just big one-off data
multiple departments. The importance clean-ups involving external consultants.
(and possibly difficulty) of this
coordination should not underestimated. Once these three steps are in place to
As the survey shows, there can be many create the basic enabling environment for
different views on who in the organisation data, the fourth leg of the data strategy is
should be responsible for owning such to manage threats to this set-up. This
activities (Figure 4.1). relates to data security and the risks of
data theft and cyber-security more
Once a lot of data is accessible to the generally. All the hard-won benefits of an
organisation, the second step is to be otherwise coherent data strategy can be
able to store it in a way that uses a clear lost through breaches that undermine the
architecture and tagging. Otherwise security of data in the organisation.
there will be a lot of data, but it will be
impossible to find the right data sets As more data is acquired, especially if it is
needed for specific algorithms: a bit like personal customer data, the datasets will
having the internet but without a good become a target for hackers. Personal
search engine to find what one needs data has an increasing value to fraudsters,
when the requirement arises. This may money launderers and organised crime.
become particularly important in large This fuels increased attempts by hackers
complex/global organisations with to try and steal data or for employees to
long-standing silos of data across different be enticed into extracting data for the
systems. Complete restructuring of the purpose of sale to external agents.
entire data eco-system may not always be
Data strategy is not directly part of any ML
realistic, but it may be necessary to
use-case, but is an absolute requirement
accept that work is needed (even if over
even before beginning to use ML. There
time, incrementally) in this area if ML is to
may well be specialists who execute
produce the promised benefits.
specific parts of the strategy, eg creating
the architecture. But the overarching data
FIGURE 4.1: Who in the organisation should own the process and data gathering
associated with the use of machine learning in the organisation?
n Finance/accountants, 40%
5% n IT, 14%
12% n Strategy team, 15%
n Business unit, 13%
40% n Don’t know, 12%
13% n Another team, 5%
15%
14%
27
Machine Learning: More science than fiction | 4. Ethical considerations
28
Machine Learning: More science than fiction | 4. Ethical considerations
14 https://uk.reuters.com/article/uk-lufthansa-fares/lufthansa-to-query-german-cartel-office-over-price-analysis-idUKKCN1IV1HO
15 https://liveandletsfly.boardingarea.com/2018/01/04/lufthansa-price-fixing/
29
Machine Learning: More science than fiction | 4. Ethical considerations
30
Machine Learning: More science than fiction | 4. Ethical considerations
FIGURE 4.2: Would you be comfortable trusting a black box machine learning
algorithm provided by a well-known supplier?
n Yes, 59%
n No, 41%
41%
59%
FIGURE 4.3: Relative to a human making the same decision, what level of explainability would you expect from an algorithm?
60%
50%
51%
40%
30%
20%
21%
10% 16%
12%
0
More than for a human Same as for a human Less than for a human Don’t know
31
Machine Learning: More science than fiction | 4. Ethical considerations
Even once competence requirements are The publicity around AI can make it seem
In the race to obtain this satisfied, professional accountants could as if many companies are adopting AI
capability, businesses be at risk of compromise on the and ML into their strategic plans and that
fundamental principle of integrity. This is every software company either has or is
need to assess the quality because ethical behaviour requires being developing AI/ML capability. This creates
of the applications straightforward and honest in all an expectation from investors that all
available in the (crowded) professional and business relationships. companies should be using AI/ML and an
And there may be a legitimate question anticipation of enhanced value creation.
AI/ML marketplace. to be asked as to whether the company is One study identified that the term
being straightforward and honest with its ‘artificial intelligence’, was getting almost
customers. Some might argue that it is four times as many mentions on earnings
not the company’s job to be a ‘parent’ or calls as Big Data when looked at over a
‘moral police’ to its customers; the latter ten year period16.
make purchase decisions as adults. But
questions of integrity may arise as regards In the race to obtain this capability,
the extent of personal data gathered and businesses need to assess the quality of
its use in targeting specific vulnerabilities the applications available in the (crowded)
of the customers in question. AI/ML marketplace. They also need to
assess whether many of the claims being
Ultimately, the answers will depend on made by software vendors and consultants
the specifics of the case, but in dealing are backed by adequate evidence. Fear of
with these questions professional falling behind should not drive investment
accountants will need to ensure that the decisions without proper evaluation.
principle of integrity guides their decision
making, in the face of clear commercial While some applications will prove their
pressures associated with a certain course worth, there could also be instances of
of action. This will be particularly the case companies that claim to have an AI
in situations which satisfy the minimum product, while actually using humans to
threshold of legality, but not the higher simulate AI. This may be done to fulfil
bar of ethical behaviour. product demand while, in parallel, the
equivalent software capability is being
LOOKING BEYOND THE ‘HYPE’ developed. In other instances, humans
may be used to ‘train’ and the product is
AI has become a ‘buzzword’ in recent real but may not be fully developed.
years, and ML, as part of AI, has often
attracted similar attention. Undoubtedly, Respondents in the survey were
there are real benefits from using ML, but supportive of the need for third-party
unrealistic expectations, and the vested assurance. This can be a useful way of
interests of those selling this technology getting an independent view of assertions
mean that there is also a real risk of the made in general terms as well
misrepresentation of what is on offer. performance of the ML application in
specific instances (Figure 4.4).
FIGURE 4.4: Importance of improving trust in machine learning by having third-party assurance of the elements of the system
n NET: Important n NET: Not important
70%
60% 65%
62%
59%
50%
40%
30%
20%
10% 16%
13% 13%
0
Input data Machine learning process Output
Note: 1–5 scale; NET Important is sum of 4, 5; NET Not important is sum of 1,2
16 https://www.analyticsindiamag.com/artificial-intelligence-increasingly-mentioned-earning-calls/
32
Machine Learning: More science than fiction | 4. Ethical considerations
33
Machine Learning: More science than fiction | 4. Ethical considerations
Disclosure and the level of transparency The IESBA framework is there to guide
More broadly, defending can have implications for the wider public professional accountants on an approach
the public interest interest. In this regard, respondents in the based on involving management/internal
survey generally tended to favour escalation, with further action via an
requires an ability disclosing where ML was being used in appropriate authority outside the entity if
to avoid seeing the the background, though there were some internal consultation fails to result in an
fundamental principles as nuances, as demonstrated in Figure 4.5. appropriate response. The point here is
that the fundamental principles are not to
a basic minimum that is For professional accountants, the be applied via a check-box approach. In
required for compliance. fundamental principles established by the example above, NOCLAR might
IESBA provide guidance when faced with mean, for example, that confidentiality
ethical questions. These principles must principles must be set aside because
be interpreted for the situation at hand – there is a wider case for sharing sensitive
something essential when dealing with data with relevant external authorities, in
new technologies and previously unseen the public interest.
scenarios. For instance, in certain cases,
there may be a need to consider More broadly, defending the public
involving external stakeholders. In this interest requires an ability to avoid seeing
scenario there may be a relevant the fundamental principles as a basic
application of the NOCLAR (Non- minimum that is required for compliance.
compliance with Laws and Regulations) The fundamental principles must
requirements set out by IESBA. permeate all aspects of how professional
accountants perform their roles.
FIGURE 4.5: Do you think that organisations should highlight when a decision has
been made by a machine learning algorithm?
n Always, 38%
1%
n In general yes, unless there is a
17% specific reason to NOT disclose
use of machine learning (eg
commercial, legal, regulatory), 44%
n Never, 1%
44%
34
5. Skills in a
machine-learning
environment
Professional accountants need an appreciation of how ML can affect their organisations. Use of
ML could affect the way they track and influence how the organisation creates value, require
changes to risk and control mechanisms, or perhaps create ethical considerations.
The survey results reflected a degree an enabler and is part of a wider pool of seem widely accepted that the element of
of scepticism about allowing algorithms factors, including people and processes. human oversight can be done away with
to take the lead in areas requiring ACCA’s skills framework (Figure 5.2), which completely or that the technology can
complex judgement and interpretation. guides its qualification, expresses this take into account human factors, whether
Depending on the type of decision idea with the Digital Quotient (DQ) being in building client relationships or leading
needed, only between one-tenth and one of a series of professional quotients successful teams.
one-third of respondents thought that a that interlink to create a ‘future-proofed’
machine algorithm could lead or be given professional accountant for a digital age. ACCA’s work on the EQ strongly
full reliance when applying such demonstrated the need for emotional
judgement (Figure 5.1). Technology, even the most sophisticated intelligence-related competencies in a
technology such as AI, still currently digital age (ACCA 2018). In fact, DQ and
Technology is clearly a significant factor in struggles to replicate the full contextual EQ are best seen as skill sets needed in
influencing the future direction of the understanding and integrated thinking of combination for either to be really effective
accountancy profession. But technology is which humans are capable. It does not for professional accountants in a digital age.
FIGURE 5.1: What is the role of ML in applying judgement appropriately in complex scenarios?
n NET: Machine learning leads/full reliance n NET: Machine learning has no/some role
80%
70% 74%
60%
62%
50%
52%
40% 47%
40% 40%
30% 36% 35%
29%
20% 24%
10% 16%
10%
0
Measurement of Presentation and Recognition and Selection and application Business judgement Ethical judgement
accounting estimates disclosure derecognition of assets of accounting policies eg future prospects eg scenarios that
and liabilities for the business are technically legal,
but not clear if ethical
Note: 1–5 scale; NET leads/full reliance is sum of 4, 5; NET no/some role is sum of 1,2
35
Machine Learning: More science than fiction | 4. Ethical considerations
Even beyond areas such as leadership, This is not to say that AI can be ignored. It is
Professional accountants core technical activities can require certainly the case that some new skills could
need to remain engaged, judgement and interpretation that draw become more prevalent as we look ahead
on multiple abilities, and where future to a world with increased use of AI, for
as AI and its component choices may not be fully informed by example the data scientist. As accounting
parts evolve. While some patterns seen in the past. methods come to rely increasingly upon
may want to delve deeper ML, there will be a need to get the most
The role of an accountant involves tasks in out of ML techniques, and to be able to
into the mechanics many familiar areas such as audit, corporate bridge the worlds of accountancy and data
behind AI, this may reporting, or taxation. But in the effective and statistics led insight. Data scientists can
execution of these tasks, the professional advise in areas such as evaluating which
be less appropriate for accountant is expected to have an ML tool to use in which circumstance, and
other roles. opinion, and to be accountable for that whether there could be ethical issues
opinion and liable for its consequences. inherent in the ML application.
The legal system, societal and cultural Professional accountants need to remain
values are founded on the principle that engaged, as AI and its component parts
only individual people or corporate entities evolve. While some may want to delve
can be held accountable. And there does deeper into the mechanics behind AI, this
not seem, at this stage, to be any broad may be less appropriate for other roles. At
consensus that this principle and a minimum, all finance professionals should
associated legally backed accountability maintain awareness of how AI is evolving
can be outsourced to AI algorithms. and be alert to how the developing
Almost all ML tools, even purely capability could overlap with their role.
unsupervised learning algorithms, This could range from how their own role
depend on the domain knowledge and can incorporate AI to be more efficient and
expertise of the people who develop serve clients or employers better, through
them. The task of ‘binning’, or turning to identifying how AI is being adopted
raw data into discrete measures that ML and used by these clients and employers.
algorithms can use to derive inferences This is also why this report has introduced
from their training sets, is still as much some specifics from the market, by way of
an art as a science, and typically requires information in the section on applications.
the collaboration of highly trained Appendix 3 suggests some avenues for
technologists, data scientists, and maintaining awareness of evolving
industry experts to be successful. developments in this area.
36
Conclusion
37
Appendix 1
38
Appendix 2
Country snapshots
1. UK 40
2. China 41
3. Malaysia 42
4. Singapore 43
6. Ireland 45
7. Pakistan 46
39
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
UK
UK 1: AI as hype or reality UK 2: Who in the organisation should own the process and data
n Now n 3 years’ time XX% Global gathering associated with the use of ML?
70%
n UK n Global
60% 70%
58%
50% 60%
53%
34% 50%
40%
40%
30% 40%
26% 40%
37%
20% 30%
13% 23%
10% 14% 20%
19%
0 10% 13% 14% 14% 15% 12% 13% 12% 5% 5%
All / Mostly hype Mostly / Entirely reality
0
Note: remaining respondents said ‘Equal hype and reality’ Finance/ IT Strategy Business Don’t know Another
accountants team unit team
40% 44%
38%
30%
UK 4: Role of ML in applying judgement appropriately in UK 5: Please tell us about the main barriers to using ML in the
complex scenarios? organisation more generally
NET: ML leads/full reliance n UK n Global n UK n Global
50% 70%
40% 60%
16% 17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
an gnit an
10% 14%
ilit f
lici f
sin ts
isc atio
iab ion o
po n o
s
ure
eth ut
12%
ies
es
ess
ate
for re p gem
11% 11%
est ent
l
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
40
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
China
CHINA 1: AI as hype or reality CHINA 2: Who in the organisation should own the process and
n Now n 3 years’ time XX% Global data gathering associated with the use of ML?
70%
n China n Global
60% 70%
58%
60%
50% 54%
40% 50%
34% 40%
30%
30% 26% 38% 40%
20% 30%
23%
13%
10% 20%
12% 21%
10% 14% 14% 15%
0 12% 11% 13% 12% 3% 5%
All / Mostly hype Mostly / Entirely reality
0
Note: remaining respondents said ‘Equal hype and reality’ Finance/ IT Strategy Business Don’t know Another
accountants team unit team
40%
42%
38%
30%
20% 24%
19% 20% 21%
10% 8% 6%
5% 4% 3%
10%
0
No plans for adoption Initial discussions and Don’t know Early stage preparation with Full production mode Advanced testing with
exploring concepts/ideas ‘go-live’ within 12 months dealing with live data ‘go-live’ within 3-6 months
CHINA 4: Role of ML in applying judgement appropriately in CHINA 5: Please tell us about the main barriers to using ML in
complex scenarios? the organisation more generally
NET: ML leads/full reliance n China n Global n China n Global
50% 70%
40% 60%
57%
30% 34% 36% 34% 35% 50% 52%
55%
29% 49%
26% 27%
20% 24% 40%
17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
an gnit an
10% 14%
ilit f
lici f
sin ts
isc atio
iab ion o
po n o
s
ure
eth ut
ies
es
ess
ate
for re p gem
5% 11%
est ent
10%
l
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
41
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
Malaysia
MALAYSIA 1: AI as hype or reality MALAYSIA 2: Who in the organisation should own the process
n Now n 3 years’ time XX% Global and data gathering associated with the use of ML?
70%
n Malaysia n Global
60% 58% 70%
59% 60%
50%
40% 50%
34% 26% 40%
30% 40%
29% 27%
20% 30% 34%
13%
10% 20%
11% 18%
10% 14% 14% 16% 15% 16%
0 13% 12% 5%
All / Mostly hype Mostly / Entirely reality 2%
0
Note: remaining respondents said ‘Equal hype and reality’ Finance/ IT Strategy Business Don’t know Another
accountants team unit team
40%
36% 38%
30%
28%
20% 24%
20% 21%
10% 8% 6%
10% 5% 3%
2%
0
No plans for adoption Initial discussions and Don’t know Early stage preparation with Full production mode Advanced testing with
exploring concepts/ideas ‘go-live’ within 12 months dealing with live data ‘go-live’ within 3-6 months
MALAYSIA 4: Role of ML in applying judgement appropriately MALAYSIA 5: Please tell us about the main barriers to using ML
in complex scenarios? in the organisation more generally
NET: ML leads/full reliance n Malaysia n Global n Malaysia n Global
50% 70%
69%
40% 60%
41% 40%
30% 36% 35% 35% 50% 54%
52%
29% 30% 49%
20% 24% 40%
21%
16% 17%
10% 30% 32%
10%
0 20% 24%
21% 21%
19% 19%
lea lega at ar t
17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
an gnit an
lici f
sin ts
isc atio
iab ion o
po n o
s
ure
eth ut
ies
es
ess
ate
for re p gem
11%
est ent
10%
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
42
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
Singapore
SINGAPORE 1: AI as hype or reality SINGAPORE 2: Who in the organisation should own the
n Now n 3 years’ time XX% Global process and data gathering associated with the use of ML?
70%
58% n Singapore n Global
60% 70%
63%
50% 60%
40% 50%
34% 46%
30% 40%
32% 26% 40%
20% 25% 30%
13%
10% 20% 22%
11%
10% 14% 16% 15%
0 11% 13% 12% 5%
2% 3%
All / Mostly hype Mostly / Entirely reality
0
Note: remaining respondents said ‘Equal hype and reality’ Finance/ IT Strategy Business Don’t know Another
accountants team unit team
40%
38%
30% 35%
27%
20% 24%
21%
10% 8% 6% 6% 6%
13% 13% 3%
0
No plans for adoption Initial discussions and Don’t know Early stage preparation with Full production mode Advanced testing with
exploring concepts/ideas ‘go-live’ within 12 months dealing with live data ‘go-live’ within 3-6 months
SINGAPORE 4: Role of ML in applying judgement appropriately SINGAPORE 5: Please tell us about the main barriers to using
in complex scenarios? ML in the organisation more generally
NET: ML leads/full reliance n Singapore n Global n Singapore n Global
50% 70%
40% 60%
40% 57%
36%
30% 33% 35% 50% 52%
30% 29% 48% 49%
27%
20% 24% 40%
17% 16% 16%
10% 30%
10% 27%
0 20% 24%
21%
19% 19%
lea lega at ar t
16% 17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
an gnit an
10% 6% 14%
ilit f
lici f
sin ts
isc atio
13%
iab ion o
po n o
s
ure
eth ut
ies
es
ess
5% 11%
ate
for re p gem
est ent
l
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
43
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
UAE 1: AI as hype or reality UAE 2: Who in the organisation should own the process and
n Now n 3 years’ time XX% Global data gathering associated with the use of ML?
70%
n UAE n Global
60% 70%
58%
50% 55% 60% 63%
34% 50%
40%
40%
30% 40%
40%
26%
20% 13% 30%
21%
18% 20%
10%
15% 7% 13% 7%
0 10% 12% 14% 10% 12% 5%
All / Mostly hype Mostly / Entirely reality 0%
0
Note: remaining respondents said ‘Equal hype and reality’ Finance/ IT Strategy Business Don’t know Another
accountants team unit team
40% 46%
38%
30%
27%
20% 24%
21%
10% 8% 6%
12% 5% 5% 5% 3%
0
No plans for adoption Initial discussions and Don’t know Early stage preparation with Full production mode Advanced testing with
exploring concepts/ideas ‘go-live’ within 12 months dealing with live data ‘go-live’ within 3-6 months
UAE 4: Role of ML in applying judgement appropriately in UAE 5: Please tell us about the main barriers to using ML in the
complex scenarios? organisation more generally
NET: ML leads/full reliance n UAE n Global n UAE n Global
50% 70%
17% 17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
15% 14%
an gnit an
10%
ilit f
lici f
sin ts
isc atio
iab ion o
po n o
s
ure
eth ut
12%
ies
es
ess
ate
for re p gem
11%
est ent
l
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
44
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
Ireland
IRELAND 1: AI as hype or reality IRELAND 2: Who in the organisation should own the process
n Now n 3 years’ time XX% Global and data gathering associated with the use of ML?
70%
n Ireland n Global
70%
60% 58%
50% 56% 60%
40%
38% 38%
30%
30%
20% 24%
21%
17% 7% 8% 7%
10% 6%
1% 3%
0
No plans for adoption Initial discussions and Don’t know Early stage preparation with Full production mode Advanced testing with
exploring concepts/ideas ‘go-live’ within 12 months dealing with live data ‘go-live’ within 3-6 months
IRELAND 4: Role of ML in applying judgement appropriately in IRELAND 5: Please tell us about the main barriers to using ML
complex scenarios? in the organisation more generally
NET: ML leads/full reliance n Ireland n Global n Ireland n Global
50% 70%
40% 60%
36% 38%
30% 35% 50% 52%
30% 29% 49%
20% 26% 24% 40% 43% 42%
17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
an gnit an
10% 14%
ilit f
lici f
sin ts
isc atio
iab ion o
po n o
s
ure
eth ut
12%
ies
es
ess
ate
for re p gem
11% 11%
est ent
l
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
45
Machine Learning: More science than fiction | Appendix 2 – Country snapshots
Pakistan
PAKISTAN 1: AI as hype or reality PAKISTAN 2: Who in the organisation should own the process
n Now n 3 years’ time XX% Global and data gathering associated with the use of ML?
70% 58% n Pakistan n Global
60% 65% 70%
50% 60%
40% 50%
34% 26%
30% 40%
31% 41% 40%
30%
20% 30%
13%
10% 20%
11% 18% 9% 9%
15% 15% 8%
0 10% 14% 13% 12% 5%
All / Mostly hype Mostly / Entirely reality
0
Note: remaining respondents said ‘Equal hype and reality’ Finance/ IT Strategy Business Don’t know Another
accountants team unit team
40%
38%
30% 33%
28%
20% 24%
21%
17% 8% 8% 9%
10% 6% 5% 3%
0
No plans for adoption Initial discussions and Don’t know Early stage preparation with Full production mode Advanced testing with
exploring concepts/ideas ‘go-live’ within 12 months dealing with live data ‘go-live’ within 3-6 months
PAKISTAN 4: Role of ML in applying judgement appropriately in PAKISTAN 5: Please tell us about the main barriers to using ML
complex scenarios? in the organisation more generally
NET: ML leads/full reliance n Pakistan n Global n Pakistan n Global
50% 70%
40% 60%
37% 36% 35% 35% 36%
30% 50% 53%
51% 52%
29% 29% 49%
20% 24% 40%
17% 16% 15%
10% 30%
10%
0 20% 23% 24%
21% 21% 21%
19%
lea lega at ar t
17% 17%
no ically os th men
d
ng tio and
bu pec t
the ros en
im of
los n
r if l, b e
an gnit an
lici f
sin ts
isc atio
iab ion o
po n o
s
ure
eth ut
12% 11%
ies
es
ess
ate
for re p gem
est ent
l
ica
ets co tion
ou ppli tion
ge
d d nt
ng m
0
d
an rese
nti ca
ass dere ogni
a lec
nti ure
fut s ju
Se
dl
hn ari
op staff
ng fit
f d nt
en al
s
ata
rs
ou eas
as
eg Ethic
eg es
ion
c
ML
ata
ts
uir /leg
rrie
usi ene
e o cie
Re
t io
mm
sin
fd
acc M
cat
he led
lum uffi
tc
ba
b
req ory
em
yo
Bu
ile
pli
acc
fro lear
d t skil
vo Ins
ad
/no
alit
lat
ld
im
gu
c
ica
lea of
ow
qu
st
No
Re
to ack
Eth
Co
kn
or
Po
n’t
L
Do
46
Appendix 3
A starting point for keeping abreast of basic developments is to look at how consumer
applications and capabilities incorporate AI techniques and functions. This can range from
smartphone ‘apps’, software, Web applications, consumer devices, IoT devices, to autonomous
cars and drones etc.
Technology publications are a good source of practical impact from AI. Forums such as AI4People provide valuable
details of these new capabilities. TED talks can be a useful insight to the direction of AI and how to address the social
way of gaining insight into a range of topics, including and ethical challenges.
technologies such as AI. A vast range of topics are covered
and available online at the TED website, all no longer than Case studies from software vendors and consultancies are a
18 minutes. (admittedly biased) source of information for how these
developments are being deployed but they do give
MeetUp groups are an easy way of gaining insight into examples of the benefits being achieved. Many software
technology (including AI) and of expanding one’s vendors will also provide training on AI capabilities.
understanding and connecting with people involved directly
in AI development and use. Many organisations hold Technology analysts (such as Gartner, Forrester, IDC) publish
‘hackathons’ for speedy creation and testing of ideas for reports and survey results, and run events on topics such as
using new technologies in an innovative way. These work AI and ML with a bias towards identifying and explaining the
best where technology and business professionals practical capabilities available from vendors.
collaborate, and are an ideal opportunity for finance
Organisations such as Digital Catapult, created to encourage
professionals to get involved, find out more, and show where
collaboration between business and technologists, are
they can add business value.
another source of information about applications, including
AI is an emotive topic that raises concerns over data access AI and ML, especially those from early-stage companies that
and sharing, as well as for jobs, employment and training. may not yet have broader visibility.
Accountants should stay aware of media coverage of AI and
There are independent course providers that provide
its impact on jobs; government policy on employment and
specific AI and ML education: eg EdX.
training; and research and surveys of consumer opinion.
Accountants need to be aware of the perceptions of social
17 [www.ted.com/talks]
18 [www.meetup.com]
19 http://www.eismd.eu/ai4people/
20 https://www.digicatapult.org.uk/
21 https://www.edx.org/course/machine-learning-columbiax-csmm-102x-4 https://www.edx.org/course/artificial-intelligence-ai-columbiax-csmm-101x-4
47
References
ACCA (2017), Professional accountants – the future: Ethics and trust Vaidyanathan, N. (2017), Divided we fall distributed we stand.
in a digital age <https://www.accaglobal.com/content/dam/ACCA_ The professional accountant’s guide to distributed ledgers and
Global/Technical/Future/pi-ethics-trust-digital-age.pdf>, accessed blockchain <https://www.accaglobal.com/content/dam/ACCA_Global/
20 March 2019. Technical/Future/Divided%20we%20fall%2C%20distributed%20we%20
stand%20-%20The%20professional%20accountant%E2%80%99s%20
ACCA (2018), Professional accountants: the future: Emotional quotient guide%20to%20distributed%20ledgers%20and%20blockchain.pdf>,
in a digital age <https://www.accaglobal.com/content/dam/ACCA_ accessed 20 March 2019.
Global/professional-insights/Emotional_quotient/pi-emotional-
quotient-digital-age.pdf>, accessed 18 March 2019. WEF (World Economic Forum), Marsh and McLenan Companies and
Zurich Insurance Group (2019), The Global Risks Report 2019, 14th
Datamaran (2018), Global Insights Report 2018: The Rise of ESG edition <http://www3.weforum.org/docs/WEF_Global_Risks_
Regulations <https://www.datamaran.com/wp-content/ Report_2019.pdf>, accessed 20 March 2019.
uploads/2018/10/GIR-2018.pdf>, accessed 20 March 2019.
Yao, M., Zhou, A. and Jia, M. (2018) Applied Artificial Intelligence:
A Handbook for Business Leaders. (USA: TopBots).
48
PI-MACHINE-LEARNING
ACCA The Adelphi 1/11 John Adam Street London WC2N 6AU United Kingdom / +44 (0)20 7059 5000 / www.accaglobal.com