ARTIFICIAL INTELLIGENCE (AI) ON THE FUTURE OF TAX

ADMINISTRATION
Introduction
The emergence of new technologies, increased digitalisation and the unprecedented flow of
information is creating fundamental and rapid change across the economy. This is causing
many governments to re-examine how public service delivery is best carried out in the 21st
century. All tax authorities are re-thinking not only what they do but how they do it. This
involves looking at the position and role of tax administration in the wider “tax ecosystem”
and against the objectives of lowering costs, enhancing voluntary compliance and reducing
the burdens arising from paying taxes, thus helping to promote growth and investment .

Among the new technology, we have Artificial intelligence (AI).AI is a broad term that refers
to techniques making machines “intelligent”. AI research and application utilize automation
to enhance or replicate human intelligence to improve the analysis and decision-making
capabilities of machines. It allows complicated and time-consuming tasks to be completed
more effectively and efficiently.

AI initiatives in taxation
AI in the taxation area consists of a wide variety of narrowly and sharply defined tools for
particular tasks. Some of the applications of AI may not be well known or expected.

To understand how AI is applied in taxation, Cas Milner and Bjarne Berg (2017) believe that
AI’s role in taxation is like a software that can automatically adapt to the input of different
content and make judgments without specific instructions While AI robots acting as tax
accountants is currently believed to be unlikely, they can perform various roles, such as
assisting tax auditors in detecting errors, classifying accounts and transactions, assessing tax
audit risks, and increasingly propose favourable tax strategies within the framework of
complex global laws. There are many prospects for applying AI in taxation. It is instructive to
consider some of the most successful accounting firms. Below, we briefly discuss their
application and development in AI.

According to PwC’s official website, the combination of tax accounting and AI, integrating
the accounting system of the entire enterprise is similar to an ecosystem. The following
illustration depicts an automated system that provides an integrated tool that improves
efficiency, enhances data quality, and adapts to ever-changing risk environments through
constant information flow between three parts.

KPMG’s official website recently reported the launch of “Tax Service,” an

Intelligence tax product that assists Chinese companies to solve automated tax compliance
problems. In China’s tax system, policy changes are frequent and complex. If the tax
information or the review process is omitted without the help of AI, serious tax risks and
consequences could result, which can be avoided by using AI that can integrate and conform
to the current policy to speed up tax declaration process. In addition, KPMG’s “Tax Service”
product can perform automatic preparation of VAT and corporate income tax returns, as well
as local additional tax calculation tables, trend analysis, and the timely detection of potential
errors, risks, or abnormal conditions.

The future-state ecosystem.

Application of AI In Tax Administration

 Tax agencies can use AI systems to communicate differently with different groups of
citizens for maximum impact. One analyst noted that if an individual “had a track
record of not paying early, but their financial data said they were doing pretty well”,
then they should be picked up by a specially designed model that says “X has a
capacity to pay but not a propensity to pay”. In that case, rather than “starting with a
weak letter, then moving a bit more firm and so on, [they] might jump to stage three
of the intervention process and go harder earlier”. Even low-risk scores can inform
action. In these cases, “if you have a low risk score, then this might just encourage us
to send an SMS, as we think you’re going to pay, but we could increase efficiency on
this group [of individuals]”. This avoids spamming customers who would not respond
to such approaches, and allows agencies to develop specific interventions tailored to a
particular group.
 AI has opened up opportunities for new, simple policy options which were not
possible before.
 Tax agencies can also use AI to save money by not having to take on seasonal labour
and temporary workers. Agencies often hire a large amount of seasonal labour to
clean or fix millions of forms which are not correctly processed through the systems.
One tax agency explained that while these workers tend to “look for errors and clear
those items”, this is possibly due to the increased use of both logic-based systems and
machine learning by agencies in automating this process. They are “increasingly
working out how to automate the fixing of these forms”, and have even realised that
the seasonal workforce “made a lot of errors, and [agencies] have managed to detect
and correct many of these with new automation processes”. The money tax authorities
 save in this way can then be invested in other tasks, such as making relations with
citizens smoother, easier, and more human-centric.
 AI also allows agencies to predict potential non-compliance to maximise revenue and
fight tax avoidance. Models are often trained with particular sectors and groups in
mind, and used across the organisation by auditors in their day-to-day roles. One
agency recently invested in a large external project to build an insolvency prediction
system for firms, illustrating the importance of better intelligence about business
health rather than just about illegal practices. Some agencies are also experimenting
with clustering methods, using AI systems to understand which taxpayers are most
similar to each other, and to segment them in new ways. Social network analysis is
helping to uncover complex VAT frauds that span multiple actors.
 AI can be used to augment this as datasets become larger and more complex.
 Data manipulation: Reading, understanding and manipulating data AI can enable tax
data strategy at a time when availability to more granular data is needed to respond to
tax complexity, governments’ increasing demands for transparency, and to manage
financial and reputational risk.
 Performing structured and unstructured tasks With AI, a digital workforce can be
constantly available to support Finance and Tax, improving efficiency and
effectiveness.
 Tax notice processing Tax notices are an unstructured data problem. PwC leverages
AI to extract key terms from this unstructured data in various federal, state, and global
tax notices to automate tracking and preparation of responses. Data annotation is the
first step to help the machine understand what key terms are to be extracted from tax
notices. The following AI capabilities apply:
• Natural language processing: Converting scanned tax notice images to text
• Machine Vision: Understanding the text and extracting specific data.
 Account classification Account mapping is a classification problem. PwC leverages
AI to predict trial balance and transaction account names, mapping to a PwC
taxonomy for a variety of purposes including the calculation of tax adjustments where
affected book income or expenses are spread across multiple accounts.
The following AI capability applies:
• Machine learning: Human in the loop classification algorithm.
 Tax compliance and reporting Tax compliance and reporting is a standardisation
problem. PwC leverages AI in combination with RPA to perform structured and
unstructured income tax compliance activities. Structured, routine, manual activities
include gathering trial balance data from source systems, preparing routine book/tax
adjustments, completing forms, and posting tax accounting entries. The following AI
capability applies:
• Robotic automation (RPA): Mimicking human behaviour in desktop or web based
applications to perform tax compliance and reporting activities.
 Tax chatbot Tax guidance is a question/answering problem. PwC leverages AI for
understanding the natural language of tax questions and leading the user to the
appropriate answer based upon the intent of the question. The chatbot uses PwC
curated knowledge bases and is able to respond quickly and efficiently to tax related
 questions. The following AI capability applies: • Deep question and answering: Deep
learning and knowledge bases to understand and respond to tax questions.
 Another benefits of AI is its ability to complete time consuming tasks, such as
research or information retrieval, more accurately and efficiently with minimal
manual input thereby allowing its users to focus on other activities, such as those
requiring complex analysis and judgment.

Challenges/limitations
 The tax world is characterized by a culture where knowledge is king. The profession
attracts people who like working with a myriad of rules that must be applied to a
unique set of facts, and are good at complex problem solving. Experimentation with
unfamiliar technology is not encouraged–especially where it is seen as a threat– or
where people misunderstand the operational risk of adopting it. Since AI is inherently
probabilistic, there are inevitably situations where the machine comes up with an
answer which a panel of authorities would judge to be “wrong”. But this is to ignore
the fact that many of these questions are currently being answered by nonexperts who
don’t always arrive at the right answer, either. So, any risk assessment needs to
consider whether AI decreases the risk of an incorrect conclusion, rather than the
absolute error rate of an AI application.
 The second barrier is a derivative of the first. In order for the machine to learn, it
requires curated content and it is a reality of some of the more sophisticated
classification issues that qualified specialists may not all classify things in the same
way. So, if a classifier is coded according to the logic of one specialist (or even a
panel), the results will not necessarily align to the recommendations of other equally
qualified specialists.
 Another barrier to entry is the attitude of both tax professionals themselves and their
clients. There remains a wariness of the new, especially as to the reliability of the
results, which means that both parties to an engagement want the reassurance that a
human subject matter specialist has reviewed and confirmed the results of the
machine’s work. This is compounded by the natural reaction of many tax
professionals, that the use of AI will erode their perceived added value while
heightening engagement risk. The reality is that by allowing the professional to focus
on the more intractable problems rather than the low-level tasks which machines can
do, there is scope for AI to accentuate the value that the individual contributes.
 There is also a limit on the types of tax problems which merit the investment in
developing an AI solution. Many issues arising in the tax advisory field are
characterised by a unique fact pattern, which makes the issue as a whole impossible to
solve efficiently using AI. Nonetheless, if it is broken down into its component parts,
it may be that some aspects fall into the high-volume, fuzzy matching, classifier
problem for which this technology is particularly well-suited. The key to seizing such
opportunities is becoming sufficiently confident in the accuracy of the machine’s
outputs that, professionals and clients are happy to have AI as a component of the
team.
 Organizations which decide to take the plunge with AI, need to recognize that training
machines is difficult and they are not as developed as humans. This means that in
order to teach the machine and build confidence in its accuracy, skill is required to
know what questions should be asked and what the optimal answer is to each
question. Practitioners who have the skills and experience to build the corpus, ask the
right questions and coach the machine in the correct answers, and have the time to
dedicate to this process in preference to high-paying advisory projects, are very thin
on the ground. This could be a major practical barrier to the adoption of AI in tax
 Another interesting challenge, the tax agency was unwilling to share model weights or
detailed explanations with auditors in case they gamed the system in question. This
could happen in two main ways, according to the agency. Firstly, the auditors might
use explanations of risk scores to find cases they could develop into a larger fraud, at
which point they could let it accumulate and catch it afterwards. Under proposals by
which auditors get a commission or in part earn their own wage, this creates a system
of perverse incentives. Secondly, and perhaps more importantly, there were concerns
that auditors would start making judgements based on the explanation rather than the
risk level. If risk scores looked at particular cities, for example, and marked them as
high risk, auditors might start searching for a disproportionate number of cases in
certain regions. This would be problematic, because the model might be flagging a
more nuanced correlation, so higher-level management might be unable to control
issues of skew, bias and discrimination within the AI systems they deploy and
manage.

Concluding remarks
Tax agencies have been at the forefront of deploying analytics in government, and AI
is no exception. Agencies have benefited from interventions, such as automating the
tasks for which they formerly used seasonal labour, to pursuing cases where they
would be more likely to recover tax revenue, to making it easier for taxpayers to be
compliant through predictive communications projects. Challenges remain, though.
Transparency is a double-edged sword – making a system too open may promote
gaming by individuals seeking to evade the tax system. It may also reveal
uncomfortable truths about areas or sectors that are neglected or disproportionately
targeted by auditors.[5] Furthermore, analysing uncertain and often expensive AI
design initiatives to understand whether they will provide sufficient benefit relative to
required investment is also an area in need of further methodological development. A
complex AI system has high risks but potentially high rewards – it might have huge
benefits, or might not work at all. Yet these issues are not insurmountable, and new
balances, methodologies and trade-offs will continue to be created and navigated.
Looking ahead, AI systems are set to continue to make tax easier to pay for citizens
and fairer in its collection and enforcement for all.
References
http://raw.rutgers.edu/docs/wcars/41wcars2/vinod.pdf
https://www.scirp.org/journal/paperinformation.aspx?paperid=86895
https://wts.com/global/publishing-article/wts_ai_study~publishing-
article?language=enhttps://wts.com/global/publishing-article/ai_tax-
opportunities~publishing-article?language=en
https://resources.centreforpublicimpact.org/production/2019/01/CPI-AI-Case-Study-
Taxation.pdf
https://becominghuman.ai/the-future-of-tax-technology-automation-5b8b05d2a7ce
https://readwrite.com/2019/03/22/ai-is-changing-how-we-pay-taxes-and-how-theyre-
collected-too/
https://papers.ssrn.com/sol3/Data_Integrity_Notice.cfm?abid=3244670
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3323867
https://www.pwc.com/gx/en/services/tax/publications/tax-function-of-the-future.html
https://igt.gov.au/publications/reports-of-reviews/future-of-tax-profession/chapter-2-
technology-impacting-on-the-tax-system/
https://www.pwc.nl/nl/assets/documents/the-data-intelligent-tax-administration-
whitepaper.pdf
https://www2.deloitte.com/content/dam/Deloitte/global/Documents/Tax/dttl-tax-
artificial-intelligence-in-tax.pdf
file://ctx-upm-01/FDR-DTD$/T33401718/Downloads/dttl-tax-artificial-intelligence-
in-tax.pdf
https://www.tpa-global.com/nieuws/2019-09-16/tax-authorities-are-powering-
processes-with-artificial-intelligence-ai
https://www.ey.com/en_gl/tax/how-ai-and-robotics-may-change-tax-job-duties
file://ctx-upm-01/FDR-
DTD$/T33401718/Downloads/Discussion_on_the_Development_of_Artificial_Intell.
pdf
https://www.reuters.com/sponsored/article/AI-and-the-future-of-the-tax-function