Day 1 Presentation - History of AI and Integration of Accountancy.pptx

DAY 1


1 / 38
volgende
Slide 1: Slide
newEditor

In deze les zitten 38 slides, met tekstslides.

Dit is niet oké

Onderdelen in deze les

DAY 1


The accountant 20 years ago: the backbone of business finance

AI's Effect on Accounting Today

AI has entered the scene, with most accounting professionals embracing it. AI capabilities are reaching new heights with no sign of stopping—it’s expected to save the financial services industry $1 trillion by 2030.

In 2024 alone, OpenAI released six new ChatGPT models to satisfy the growing demand for newer and better developments—with 65% of companies planning to invest in generative AI in the next five years globally.

For the accounting industry, the uptake of AI is mostly unavoidable, with 80% of accounting professionals reporting increased AI functionality in their existing software. Change is already here—and the reception is mostly positive. 85% are excited, or at least intrigued, by the prospect of AI.

Despite this, only 19% believe that their peers share their excitement. This suggests that the industry is thinking about AI, but it isn’t talking enough about it. Firm leaders have the greatest influence on these conversations. Owners, partners and directors continue to be the most optimistic about AI, seeing it as an ally and an opportunity for growth. On the other hand, individual contributors (accountants, bookkeepers, enrolled agents) and those in operations, technology, and administration roles are more likely to be skeptical or scared—although that number has declined since last year (-7% YoY). So, while some hesitation is natural—and necessary for a fully rounded discussion— attitudes towards AI in the accounting industry are increasingly more positive. Firm leaders especially are lighting the way forward.

Understanding the Bigger Picture

  • AI for accounting is no longer a novelty—it’s a necessity.

  • 83% of accounting firms are making the most of AI, offloading routine, low-level tasks onto automation, generative apps, and virtual assistants, so they can prioritize their clients, staff, and business strategy.

  • With more capacity to focus on work that drives growth and profitability, AI-positive firms are positioned to out-strategize and out-scale their traditional counterparts, giving them the edge in an increasingly competitive market. Some have even started to introduce new services (13%) and more are considering following suit (37%)— making it harder for AI-resistant firms to catch up in the long run.

  • Accounting professionals continue to see the benefits of AI:

    • AI isn’t science fiction anymore. It’s a business reality—and with most software vendors applying AI within their product, it will become part of an accounting firm’s tech stack sooner or later. Staying competitive means staying agile.

    • On average, accounting professionals save 56 minutes a day (a potential of 18 hours a month) with AI, with over a third (38%) of this time spent on communication tasks like drafting emails and managing correspondence. Role, geography, firm size, and services offered did not have a significant bearing on the amount of time an employee can save. But skill level did, with advanced users saving 71% more daily minutes than beginners

    • When firms invest in AI-focused education and training, they save a daily average of 66 minutes per employee—22% more than those that don’t. That’s a difference of 40 hours annually. Despite this, 57% of firms offer no AI training, while only 37% actively offer or are developing educational resources related to AI. Firms at ground zero—with no AI or AI training—are losing almost 7 weeks a year to their AI-progressive competitors.

    • By saving time, firm leaders are creating space for more meaningful work. In the future, accounting professionals may choose to, or be asked to, develop new advisory or technical skills to satisfy evolving client expectations, which means their capacity (and work-life balance) is just as important as their capability. And with small (26%), medium (54%), and large (56%) firms worried about recruiting and retaining employees, firm leaders can’t afford not to prioritize employee satisfaction.

    • As the accounting talent shortage continues to persist, with more accountants retiring than entering the profession, firm leaders continue to shoulder the pressures of attracting and retaining top accounting talent. AI and other emerging technologies can take some of this weight.

    • Young people want to work in progressive, tech-forward industries. Accounting can be that industry—it’s ripe with possibilities. Firm leaders should feel empowered by these opportunities, and learn from thought leaders in the accounting space who are experimenting with AI and sharing their findings.








THE NEED FOR RESKILLING IN AI

12



Artificial Intelligence consists of several areas, including machine learning, deep learning, and generative AI (each a subset of the previous). To truly unlock AI's value for the justice system, it is imperative to understand the details of each of these areas and how they work. The skills required to identify and effectively embrace new AI use cases are changing, especially for those leading AI transformation.

The legal industry will be heavily impacted by the use of AI (notably, across each of its different areas – and NOT only generative AI). 

There is a hype-driven rush from law firms and LegalTech vendors to push out AI-driven solutions and services – noting the business optimisation and cost efficiencies they can bring. There is the opportunity for AI to help address access to justice, but there are significant risks that only those with resources will be able to leverage the benefits of Gen AI and this will only serve to further exacerbate existing inequalities.

There is a need to understand the detail behind these areas of AI, empower themselves to unlock value from AI, and assess the correct posture in which AI is utilised.

For example, for all the excitement regarding AI, few have noted its strong reliance on well–governed and appropriate–quality data. Understanding the fundamental differences in technical approaches is crucial.

What is AI?

Understanding how AI solutions work...

13

TURING TEST

14


15

DEEP LEARNING



Risks and Considerations in AI adoption


The importance of ethical behavior in accounting

  • Ethical behavior is undeniably the cornerstone of trust and reliability in the tax and accounting industry. People must be able to trust their financial professionals implicitly; ethics are the foundation of transparency, accuracy, and accountability. Accountants play a critical role in upholding these principles, acting as stewards of their clients' financial information.

  • While AI can make tax experts’ workdays easier, these professionals must remain vigilant regarding potential risks and ethical implications associated with these technologies. They must actively engage in ongoing education and training to stay abreast of emerging ethical concerns and ensure that AI is used responsibly and ethically to avoid issues with the IRS and other tax authorities worldwide.


AI's Impact on Ethical Dilemmas

  • AI-driven tools offer unparalleled efficiency and precision in data analysis, financial forecasting, and compliance monitoring. Yet, over time, they may usher in ethical challenges, as well. As algorithms make increasingly complex decisions, concerns have already arisen regarding data privacy, algorithmic bias, and the erosion of human judgment. Accountants must navigate these ethical dilemmas with care, balancing the benefits of AI with the need for moral integrity.

  • For instance, one concern is the potential for AI algorithms to perpetuate or even exacerbate existing biases that exist in the data they analyze. AI systems may inadvertently discriminate against certain groups or produce misleading insights without careful oversight and proactive measures to address biases. Accountants need to be aware of this and double-check any work they do with the assistance of artificial intelligence. By integrating ethical considerations into their usage of AI tools, accountants can uphold their commitment to integrity and professionalism while taking advantage of the transformative power of AI.


  • We find that the main challenges in AI tool development are related to transparency and explainability, AI bias, data privacy, robustness and reliability, fear of auditor overreliance on AI, and the need for AI guidance. Study participants report that audit clients are not yet using complex AI in financial reporting, but that the interaction between AI-aided financial reporting by clients and auditing could become a novel challenge in the future. While new techniques to audit client AI are being developed, the lack of trust in AI tools discourages auditors from auditing through AI and instead they audit around it by reviewing the outputs the tools generate.

Ethics, accountability and data security

  • AI is a powerful assistant but not a replacement for human judgement. Ethical concerns arise when decision-making is automated without human oversight. Now is a great time to focus on upskilling your team’s soft skills such as effective communication.

  • AI should be seen as a supplement, not a decision-maker, in critical financial or advisory processes, therefore all AI-influenced decisions should be reviewed by qualified professionals.

  • Just as we do with all digital technology, data security must be a top priority for firms using AI.

  • Knowing the systems you are using is essential for this. For example, ring fencing your data in specific tools so it cannot learn from your data, or switching off the capabilities to learn in tools such as ChatGPT.

  • It is also recommended that you perform regular audit, implement access controls and educate your team regularly.

Client concerns about AI usage

  • Clients may feel uneasy about their data being used in AI-driven processes. To address these concerns ensure you are transparent with how you are using their data and how you aim to protect it.

  • Always give clients the choice. For example, if you are using a note taker. However, this could also be a good opportunity to discuss the main benefits of AI including how they may want to use it in their business.

  • It is also important that we remember clients will be coming across these tools in their personal and business life which may mean they expect their accountant to be using them.

  • Whilst AI offers immense potential to revolutionise accounting, its use must be underpinned by strong ethics, accountability and data security. By adopting a clear AI policy, choosing secure tools, addressing client concerns, and maintaining human oversight, firms can responsibly embrace AI and reap the rewards it presents.


Audit – achieving consistency in data formats across clients

  • Incomplete or poor quality data as well as algorithmic bias and the lack of scalable AI infrastructure are potential risks of using AI. They call for more research, particularly interview-based related challenges involving the low maturity of AI tools, poor client data quality, and lack fo trust in technology. Technology-related challenges identified by Seethamraju and Hecimovic also include the lack of third party Ai tool compatibility with either the existing audit support systems or client ERP systems mainly because of the resources intensive data cleansing effort required, tool customisation needed to adapt to certain clients' industries, and the refinement of tools needed because the tools are still under development. 

  • On the client side, they highlight barriers such as client data quality and integration, and the potential for bias that can be difficult to detect and remove. 

  • Organizationally, the authors document differences in resources between Big 4 and non-Big 4 firms. Big 4 firms have dedicated significant global investment in auditor upskilling in contrast with non-Big 4 firms that often rely on off-the-shelf tools and encounter the lack of IT auditors and those with skills to interpret AI tool results and face cost and training pressures. Challenges related to humans using AI tools are also identified. For example, Fedyk et al (2022) find that the main barrier to AI adoption is the lack of trained human capital. In addition to the need to upskill audit talent, trust in technology is found to be another drawback to AI implementation in audits because of uncertainty and the lack of transparency related to how the analyses are generated and because auditors generally do not take part in the development of the tools and therefore find it challenging to trust them (Seethamraju and Hecimovic 2022; Samiolo et al 2023).

Continuation

  • A few studies focus on the ethical implications of AI technology use in accounting and auditing. Through the lens of ethical frameworks, Munoko et al. (2020) perform bibliometric analysis of publications to explore practical ethical challenges that could be encountered in AI enabled audit. They address issues such as the lack of transparency, concerns related to data privacy, fairness of data use, and user autonomy and accountability (Munoko et al 2020). Furthermore, they raise questions about audit quality, professional skepticism and concerns over AI tools’ influence over their users, governance of AI, and auditor competence. In another study focused on ethical AI use in accounting and auditing, Lehner et al (2022) perform a theory-driven literature review to identify five “challenges to AI-based ethical decision-making in accounting” (p. 119): objectivity, privacy and data protection, transparency, accountability, and trustworthiness. They highlight the need for understanding how to competently use and govern AI while being aware of its limitations.

Challenges of AI use in Auditing

  • Our interview analysis indicates that there are many challenges that accounting firms encounter as they implement AI in audits. 100 % of participants mentioned trust as a key issue in AI adoption in audits. Trustworthy AI can be conceptualized as fair and impartial; robust and reliable; preserving privacy; safe and secure; responsible and accountable; and transparent and explainable (Deloitte 2023). We have organized the challenges discussed by the interviewees according to the most relevant themes and identified additional themes that are more unique to auditing such as overreliance on AI and the need for AI guidance in audits.


Transparency and Explainability

The terms transparency and explainability refer to understanding the model and reasons for its outputs and is a risk associated with AI tools that use machine learning (ICAEW 2023b). Transparency of AI systems relates to the “information about datasets, processes, uses, and outputs that are shared between stakeholders” (Ammanath 2022, 63). Explainability in AI means “it is possible to understand how an AI output was calculated” (p. 77). Moreover, explainability is related to interpretability which centers on understanding relationships in a system as well as understanding “the training, the data, and the learned components, such as weights and parameters” (p.77). Finally, intelligibility addresses “how well a model can be interpreted in human terms” (p. 78).

Challenges related to the transparency, explainability, interpretability, and intelligibility of AI tools were raised by many interviewees mainly because of the need to communicate how the tool performed to the regulators. P-13 states the following:

It’s hard to train that AI, it’s hard to document and support how that AI works. (P-13).

P-7 further elaborates on the concern related to model explainability in the following:

  • Firm standard audit tools have to go through a very different process or actually the PCAOB has to, and the AICPA may be like, we have to build these in a way that is very open and transparent to the regulators and we have to go through that process very specifically to make sure that they understand what these tools do. There’s still a lot of gray area in AI and regulators are saying it’s tough because AI output is pretty opaque. We don’t know how it got to that answer. So, that’s why we layer in the Responsible AI Framework to say, “[…] We don’t know exactly how it made the decision, but it’s not biased. It’s following the standards and guidelines” and we’ve had a lot of good progress with regulatory bodies, governmental bodies to say, this is the best we can have right now, and the efficiency and the value of the AI is worth it. (P-7)

Some are developing approaches to improve model explainability (Zhang, Cho, Vasarhelyi 2022), although this capability requires considerable data science expertise. For example, P-8 suggests:

  • I Can dig a little bit deeper and say, “What’s this thing actually doing?” Let’s not dive into the code just yet. What can we do to make this more interpretable to us and there’s a variety of different ways you can do that. Different techniques that are out there, LIME, Shapley, if you’ve heard of these different things that try to get to explainability. […] different forms of regression to get to feature importance. (P-8)

Others are using the methodology developed for clients’ use of AI, which centers on clients’ ability to explain their AI to the auditors, as the blueprint for implementing AI in audits. P-3 shares the following:

  • How do you provide the trust in AI that everybody needs to say the information I am getting is correct. […] You can’t give somebody information that they make a decision on and it’s not the right information, but it’s the same question that our clients are having. If they are putting artificial intelligence into their financial reporting processes, their controls can’t get them there. They need to have a way of telling us how they know it’s right and so we’re working on methodology for the clients’ use of AI, but it becomes a mirror image. It’s going to be applied to our own AI tools. (P-3)

Communicating how AI develops its outputs will be a challenging area for firms to work through with their clients and regulators. While different techniques are starting to be used to explain this information, lack of knowledge and the complexity of these models may hinder some from relying on the technology for financial reporting areas. As a result, in cases where machine output is not sufficiently transparent, auditors might default to judgments based on their existing knowledge of the client (Samiolo 2023).




Fairness and impartiality

AI bias refers to the phenomenon of an algorithm producing results that are systemically prejudiced (e.g., by favoring one group of outcomes over another) due to erroneous assumptions in the training process. AI, if not sufficiently controlled, can then spread the bias found in the data used in its training (Deloitte 2023). P-17 shares the concern over AI bias in the following:

  • The one thing personally I worry about with AI is data bias. How do we make sure we’re feeding it really a representative data population, so it’s not giving us a biased answer because I gave it biased data. (P-17)

Of note, AI bias in auditing is not necessarily related to demographic or socioeconomic characteristics of individuals like it is in other fields, but rather it refers to incorporating sufficient information in data sets and proper training of algorithms to adhere to the correct regulations. P-20 shares:

  • We are training a system, and we are inputting the regulations (i.e., US standards, international standards, PCAOB). Perhaps we only upload three of four regulations so as a result, the model is biased against the fourth regulation […] When we train –did we upload the right information—did we include the right regulations, did we include a complete data set? (P-20)

  • To address algorithmic bias, the firms are developing tools to audit AI. For example, P-7 states:

  • We are building tools to audit AI. We have been engaged, maybe not for audit clients, but to audit the bias of AI models. (P-7)

  • However, there are challenges with using AI tools for detecting AI bias in auditing, although such tools are being developed in computer science and used in other fields (Lee, Resnick, and Barton, 2019; Roselli, Matthews, and Talagala, 2019). P-20 states:

  • The PCAOB standard on auditing says we can’t assume or indirectly test something. We need to have direct testing of a control. Inferring it works (through the machine learning algorithm) versus having to directly test the control – will the PCAOB adjust their rules for this? (P-20)

  • Because much of the AI bias stems from the underlying data, several interviewees discuss a few data-related challenges. One of those roadblocks is accessing data from client ERP systems. P-7 states:

  • It’s too risky right now and again the systems that are working across these ERP systems are pretty clunky as we know it, because we have to get the data out of them. AI needs good data to be effective and we’re just not there yet. (P-7)

  • P-20 suggests focusing on data architecture because if data is not architected consistently and with proper cybersecurity, then errors will be introduced into AI.

  • Furthermore, P-12 believes that using client data alone might not be enough as it is unlikely that it will contain the needed degree of outliers and variability. One potential solution is to supplement real client data with synthetic data but ways of implementing this is an ongoing challenge:

  • Even if you have access to hundreds of client data sets, you may still not see all the things you need to see to prove your tool is robust enough to be able to detect everything that the regulatory bodies expect to do. […] In the future the remaining challenge is how we can augment the real-world data with potentially synthetic data […]. (P-12)

  • The profession has some significant ongoing challenges related to AI bias, both actual and perceived. While no clear solutions have been identified, firms will need to find ways to address finding representative data populations that are supported by good quality data.





Privacy

  • Privacy in AI refers to “guarding sensitive information, gaining consent for use of that data, ensuring models are resilient and do not leak or divulge protected data, using models in a way that respects privacy, and meeting the emerging laws and regulations around the world that do not just encourage but indeed mandate privacy” (Ammanath 2022, 128).

  • Data privacy becomes particularly important at the AI tool training stage. P-3 discusses a concern related to the ethical use of client data in accordance with data privacy considerations when training a tool based on a vast number of contracts such that a client could become identifiable:

  • Another issue that we’re struggling with is […] that we would never monetize our client’s information. Let’s say you were building a document reader that pulls out all of the lease terms of a huge amount of contracts. If you train it on one client’s data, it’s going to learn some things and then it’s going to go apply that to another client’s data and all of a sudden − a knowledgeable auditor could read between the lines and start seeing the other client − understanding the other client’s data. […] General counsel will not yet say we’re in a position to learn across clients. (P-3)

  • P-18 states that clients are concerned about providing the auditors access to vast quantities of their data in fear of privacy and being audited around the clock. P-18 shares the following:

  • Even now the clients are hesitant to extract data, provide data, because there’s always a hesitancy, okay, “Well we don’t want to give you all our data, are you guys going to be auditing us 24/7 now?” […] The sizes of the data is still a problem, the security of the data is still a problem. […] We get the entire inventory movement for the entire year and then we can just mine that, so we can tell you how you’re recording your revenue from start to finish and when you’re recording it, when you’re collecting the cash. We can already do some of these things. If we have more data, we can do more. (P-18)

  • To address client data privacy concerns, P-20 suggests only using publicly available client and industry data to build models. P-8 describes another approach to using client data to train their AI tool while ensuring their privacy. Specifically, P-8 suggests that even though they are not permitted to use client data to develop benchmarks for clients in the same industry, they are able to extract certain pieces of client data such as control language and train their AI tool on the different ways that information can be presented. The tool uses that learning only to adjust the weights of the algorithm. P-8 outlines their approach to using client data as follows:

  • It took us a long time to get everybody comfortable with the concept of using client data to build, train and produce models. […] Some of our clients overtly expressly prohibit us from using their data and we abide by that, full stop. […] The way that we train and utilize client data is in a way such that we’re not permitted to disclose client information. We’re permitted to use it by the professional standards, but we’re not permitted to disclose it. […] By no means can [we] benchmark a client against somebody else. I would love to be able to say, “[…] We have a client in oil and gas within your equivalent of order to cash cycle. We’ve identified 10 individuals in your process flow that have a certain role in that process where your peers average three people, which means that perhaps you’re a little bit more highly manual than your peers, maybe you could use an ERP upgrade, some automation, some different back-end connection upgrades.” We can’t do that. […] “Let’s keep this to things that are going to help serve more universal means, like information extraction, like being able to extract those entities from that control language that I mentioned, that we use data from a variety of different clients intentionally, so that we can train it on different ways that different information can be stated.” That’s how we get the power of it, but we’re not necessarily regurgitating anything that the client provided to us. It’s really just being used to alter the weights of the algorithm. (P-8)

  • Producing strong AI models requires large volumes of data to train the models. The more data, the more accurate results the model will produce. It is critical for firms to protect client data and build trust with clients. It will be important for firms to be transparent around how the data will and won’t be used and demonstrate strong governance in this area to convince clients to provide access to more data.


Robustness and reliability

  • An AI model is likely to have an issue with robustness and reliability when a model performs well in training but fails to operate well in a dynamic environment after it is deployed (Ammanath 2022). Several interviewees addressed a concern regarding the ability of auditors to rely on tools that are changing after deployment. For example, P-3 shares this challenge in the following:

  • I Can tell you that the AI in it is very concerning to us. When we put a tool out, we go through what we call global certification. That’s the quality of the fit for purpose, the functioning, the data that was used, the methodology that goes with it. It’s very hard to certify something if it changes the next day or a week later. We’ve been very hesitant about how quickly we can move with AI. (P-3)

  • P-13 highlights the ongoing debate that surrounds the issue of allowing the tool to learn from real data once the AI tool is in production where it is potentially becoming more robust and reliable while considering the lack of ways to actually measure the accuracy of that tool. P-13 states the following:

  • It’s tough from a technology person’s perspective because that’s where the real learning happens. You can feed it as much dummy data or training data as you want, but in production you’re then increasing the volume of data 100X, right? That’s where the real benefit would come. We have those internal debates all the time. In fact, I just had one the other day of a tool that wants to use AI and push it to practice to get that feedback from the practice to say, oh, this recommendation was right and this one was wrong. It’s like, well, I don’t know if we can do that yet because we don’t have a mechanism yet to measure the accuracy or effectiveness very well because we don’t know what it’s doing yet. It’s a challenge for sure, especially in our regulated environment. (P-13)

  • Even though the use of continuous dynamic learning might be considered in early AI adoption discussions, we predict that audit firms will not use it due to the regulated nature of the industry. Furthermore, in rare cases where dynamic learning could be tested or used, there would be ways to measure the accuracy of predictions from machine learning models using MLOps systems (Kreuzberger, Kühl, and Hirschl 2023).

  • According to P-8, there is a way for auditors to control the evolving nature of AI tools due to dynamic learning. P-8 informs that they allow a model to be retrained only during specific time periods that typically coincide with various platform updates. This approach prevents exposing auditors to the risk of a model reaching different conclusions at different times during an audit and the risk of a model drift which occurs when a model becomes less accurate over time. P-8 states the following:

  • The way we’ve deployed our tools, they don’t keep learning. They are, by design, not dynamically learning in a reinforcement learning loop. The way we’ve addressed it is by having a manual retraining process. We’ll look, we’ll assess to determine if there’s any drift in the way that our models are working based on the data that’s being applied. Again, a good example is information extraction routine. Is it still extracting, generally pretty reliably based on what people are inputting? But we don’t retrain it really midstream because we don’t want to create this paradigm of having two classes of audits during an audit year or even just a single engagement having a switch that they’re not aware of. And if they were to run the model at the end of the audit, they would reach a different conclusion. […] So we have fixed time periods in which we’re applying updates that align with the major updates of our platform in general so that it’s not just a surprise, we’re going to put this out there. […] Right now, nothing is dynamically learning. (P-8)

  • Although challenges exist associated with the dynamic learning aspect of the technology, it appears as though relying on training during specified periods may allow for more consistency and stability across clients. While there may be some sacrifice of improving the model with more recent data during these periods, consideration of the reliability, consistency and the ability to test the model may be more important.

  • 7.5. Auditor overreliance on AI

  • The challenges discussed above focus on the technological side of AI tool design and deployment. While these challenges are significant, it is equally important to address another theme that emerged from interview data analysis – challenges that the human auditors are likely to encounter when using AI tools. On the one hand, some interviewees (P-8, P-13) share that it will take time for auditors to gain comfort with using AI tools. On the other hand, interviewees are also concerned with potential auditor overreliance on AI tools (P-7, P-13).

  • P-13 suggests that in order for auditors and regulators to familiarize themselves with AI tools and test them, the tools will be released in “baby steps” using AI-generated clusters by industry sector or sub-sector. AI will be running in the background and auditors will not directly interact with AI (P-13). P-8 adds that when the early iterations of an AI model were deployed, auditors were assured that the model only provided suggestions that they could activate if they wanted to but were not required. The auditors were encouraged to use the tool to double check their work and for self-review, but it was not a formal audit procedure (P-8).

  • As familiarity with the tools grows among auditors, the firms are recognizing the human tendency to over rely on technology and are developing approaches to prevent auditors from doing so. For example, P-13 shares that the tool that evaluates lease contracts for non-standard terms requires the auditors to verify its accuracy by providing the screenshots of those contracts and asking auditors to check that the data were selected correctly. This is done because the firm cannot make a case for relying on the tool 100 %. P-7 informs that auditors are required to exercise their professional judgment in determining the extent to which they can rely on AI. P-7 states the following:

  • For any of our tools, we have required either training or guides and or in-app guidance that tells the auditor in not so many words: this is the degree to which you can rely on this and this is what you need to do supplementary to that to be comfortable that your work is true and you can rely on your work. So, we don’t have any tools yet where we just say, “Oh, poof, here’s the answer auditor, you’re done” − that doesn’t exist. (P-7)

  • P-13 calls for continued exercise of professional skepticism and critical thinking by auditors when interpreting the output of AI tools. For example, if an AI tool suggests that an auditor’s decisions are the same as the peers’, an auditor might wrongfully rely on this information when in fact something in that engagement makes it different from its peers. P-20 suggests focusing on educating auditors to understand where the data is coming from and how the data flows from beginning of the process to the end.

  • As with many technologies, concerns around overreliance on the tools exist for AI as well. Audit approaches are being adjusted to incorporate additional procedures to reduce the likelihood of this happening. In addition, auditors need to understand the tools are making suggestions, however the auditor should still continue to use their professional judgement and skepticism when using these tools. Furthermore, greater emphasis should be placed on designing audit support systems that minimize the deskilling that affects accounting professionals (Arnold, Collier, Leech, and Rose 2023).


Trust in client AI-generated reporting

  • Among the common cases where genAI is used or expected to be used are writing disclosures of financial statements and writing code for various reports (CAQ 2024). Some of the risks reported are completeness, accuracy, and appropriateness of these disclosures as well as model performance (CAQ 2024). Echoing concerns expressed in the CAQ report, P-20 mentions that code generated by ChatGPT could be used by clients to make a change in the system that would introduce a vulnerability if a person using the code lacks programming competence. Currently, some of the risks can be mitigated by placing a human in the loop for manual checks and reviews. In the future, however, genAI-related risks are likely to become more difficult to assess as reliance on such technology increases.

  • In the context of auditor–client communication, some interviewees discussed the dilemma of trusting information provided by clients if genAI becomes more common in generating responses to auditor requests, especially in the remote work environment. For example, P-20 states:

  • How do I know that what it came up with is the right stuff? How do you know the response you got wasn’t generated by ChatGPT?

  • In light of the risks discussed, determining the source of the information may be challenging resulting in more difficulty for auditors to trust the information provided by clients. Additional skepticism may be required.


Need for regulator guidance

  • Some interviewees express concern over the lack of frameworks that govern simple automations which in turn makes them question the readiness to govern AI especially because of the difficulty with transferring human thought process and judgment to AI. P-1 states the following:

  • You don’t have a framework that even governs simple automation, you definitely are not ready […] as I bring in an AI engine that’s going to drive decisions that affect the control framework. […] I’m going to process it to bring in all this other qualitative factors where it’s now taking over somebody’s judgment that we already don’t do a good job of being able to translate into, how did I apply that judgment, what was my judgment thought process. We’re nowhere near ready to let the machines do that and cover that in a risk framework. (P-1)

  • In addition to not having a governance framework for automations, challenges that seem to resonate with several interviewees are the pressure of regulatory scrutiny (P-13) and the absence of an AI auditing standard (P-3). P-3 shares the following:

  • The European Union right now is actually focused on AI. Again, everybody begins with governance because you cannot just do this haphazardly […] but nobody’s actually written a standard on it yet. They’re all starting to focus on it. So, the IASB, which is your AICPA or PCAOB, has a working group focused on technology. How are we using it? How are clients using it? What does it mean? Everybody is recognizing the need for definitive guidance in all of these areas. (P-3)

  • P-17 calls for the need to modernize the standards in the U.S. and points out that “there’s nothing in the standards that prevents you from doing it [using AI], but it also doesn’t help you do it or encourage you to do it”. P-18 adds that “regulation is not catching up with the developments”. P-18 further explains that they are informing the PCAOB of their use of AI in audits by providing them with use cases to demonstrate that “hey, regulation needs to catch up […]; by enabling you to run, you guys are actually developing the next generation auditor.” (P-18).


Responsible use of AI in accountancy


Strategies for Ethical Excellence

  • From conducting thorough risk assessments to implementing robust data governance frameworks, we'll equip you with the tools and knowledge needed to navigate ethical considerations effectively in the age of AI. We'll explore real-world examples and case studies, highlighting best practices for integrating AI into accounting processes while maintaining ethical standards.

  • Implement data governance frameworks to ensure the ethical collection, storage, and usage of data in AI applications.

  • Establish clear guidelines and protocols for the use of AI tools in your office, including protocols for detecting and addressing algorithmic bias.

  • Provide ongoing training and education for your tax and accounting staff on ethical considerations related to AI.

  • Foster a culture of transparency and accountability within your organization, where ethical considerations are prioritized and openly discussed in team meetings.

  • Regularly review and update AI systems and processes to adapt to evolving ethical standards and regulatory requirements.

  • Engage with industry peers and regulatory bodies to stay informed about emerging ethical guidelines and best practices in AI adoption within the accounting profession.


Importance of an AI policy

First and foremost, prior to introducing AI into the firm, a clear AI policy is essential. It establishes guidelines for how and when AI tools can be used as well as who should be using them, which ensures accountability and consistency.

Key elements of an AI policy include:

  • Type of data: knowing which data can be entered into AI tools is important for the whole team, think about spilling it between non-confidential and confidential

  • Software overview: a comprehensive list of permitted software that can be used in the firm. This could be split into sections of AI specific software (such as ChatGPT) and/or software that has AI capabilities built into it (such as using Co-Pilor in Sage)

  • Data security: ensure compliance with UK GDPR and other data protection laws

  • Human oversight: maintain human involvement in critical decisions, avoiding full reliance on AI

  • Employee training: regularly train staff on ethical AI practices and security protocols




Guidelines for Responsible AI

  • Accurate

    • Have quality in your grounding material

    • Convey uncertainty when the answer isnt clear

    • Enable fact-checking when possible


  • Safe

    • Mitigate bias, toxicity and harmful content

    • Protect PII and prevent data leakage

    • Guard against nefarious use


  • Honest

    • Respect data provenance & attribution where possible

    • Make clear that content is AI-produced when delivered

    • Align use of AI to your company's values


  • Empowering

    • Recognize & supercharge human capabilities

    • Make accessible to all

    • Engage in responsible labor practices


  • Sustainable

    • Right size models to reduce carbon and water footprint

    • Balance value of applying high power queries to expected benefit / gains

    • Icnlude IEEE metrics for responsbile 


  • Accurate

    • Always provide min. 2 citations in customer facing content

    • Always show confidence score

    • Never display results with confidence score <90% exteranally

    • Never display results with confidence score <75% internally


  • Safe

    • Always design prompts 


  • Honest

    • Respect data provenance & attribution where possible

    • Make clear that content is AI-produced when delivered

    • Align use of AI to your company's values


  • Empowering

    • Recognize & supercharge human capabilities

    • Make accessible to all

    • Engage in responsible labor practices


  • Sustainable

    • Right size models to reduce carbon and water footprint

    • Balance value of applying high power queries to expected benefit / gains

    • Icnlude IEEE metrics for responsbile