CPA.com Generative AI Toolkit
A roadmap for accounting and nance
professionals to understand and leverage
the transformative impact of GenAI
2 CPA.com Generative AI Toolkit
Table of contents
CPA.com's Multi-Part
Generative AI Initiative
Over the past year, we'vewitnessedthe tremendous potential of
generative AI (GenAI) to reshape how businesses work. In response,
CPA.com is leading a multi-part GenAIinitiative to support CPAs’
understanding and utility of the transformativetechnology. This
toolkit is part of a broader collection of resources developed to
empower professionals to assess the possibilities and navigate
the associated risks of implementing GenAIinto their practices.
We encourage readers to explore the additional resources,
including our AI-focused sessions at the Digital CPA conference,
videos featuring expert insights, and more. We will also have an AI-
specic cohort as part of the 2024 Startup Accelerator program
and will hold our rst AI Symposium that will bring together subject
matter experts, rm leadersand AI solution providers to discuss
the real and potential impacts of the transformative technology on
the profession.Learn more about the broader initiative and nd the
latest resources at CPA.com/Gen-AI.
2 CPA.com Generative AI Toolkit
We believe generative AI’s impact will
be much more profound and happen
a lot faster than other technology
advancements, such as cloud computing.
— Erik Asgeirsson, CEO, CPA.com
Introduction
3
Challenges and limitations of generative AI
14
Seven steps to building an AI strategy
8
Security and risk considerations:
Safeguarding validity, privacy and transparency
18
Driving digital transformation
6
The great equalizer
17
Use Cases: Practical application of
generative AI for accounting and nance
10
3CPA.com Generative AI Toolkit
Introduction
Very shortly after it was rst introduced, ChatGPT became the fastest-
growing consumer application in history. In less than ve days, the
program reached a million users; a technology rollout record. Within two
months, it had been used by over 100 million users. By contrast, it took
roughly nine months for TikTok to hit 100 million users and over two years
for Instagram to do so. That’s a testament not only to the excitement
about the technology but also to its accessibility and ease of use. Just
a few months after ChatGPT’s release, 77% of business leaders surveyed
cited generative AI (ChatGPT is one of many examples) as tech they
expected to have the largest impact on their businesses out of all emerging
technologies over the next three to ve years. That was nearly double the
number for the next highest technology according to the KPMG survey. A
total of 71% of leaders plan to implement their rst generative AI solution
within the next two years.
Articial intelligence (AI) is an umbrella term that encompasses a
range of technologies that are generally intended to perform tasks that
might otherwise require human intelligence. It is already very much a
part of our everyday lives, powering search engines, self-driving cars,
facial recognition and virtual assistants like Siri and Alexa as just a few
examples. Accountants use it to automate data entry and reconciliation,
perform nancial modeling and forecasting, detect anomalies in auditing,
risk management and in many other use cases.
Working with generative AI is often a conversation. When we
asked ChatGPT to compose a message to accountants about
how it could assist them, its rst output was much longer
than we wanted, so we asked the system to shorten it. The
next response was better, but still a little wooden, so we asked
it to turn it into a rhyming verse. Here’s the nal product:
Hey there, accountants, listen up, it's me, your AI
friend, ChatGPT! With generative AI, we've got a
plan to revolutionize your work—and we can!
Automating tasks with a swift hand, tedious
chores, I'll take a stand. Data entry, reconciling,
tax prep too, all in a flash, I'll do for you!
Predictive insights, I'll unveil, financial trends,
I won't fail. With time freed up, you'll soar so
high, advisory roles, you'll reach the sky!
Personalization, I've got it down, customers
with smiles, not a frown. Chatbots aid, support
in sight, answering queries day and night.
Digital transformation, we'll lead the way, a
partnership that's here to stay. Together we'll change
the game, embracing AI, we'll rise to fame!
So, accountants, let's unite and soar, with AI's help,
we'll do much more. In this tech-driven world, we'll
stand tall, making accounting a blast for all!
A message from ChatGPT,
I’m here to help!
3CPA.com Generative AI Toolkit
4 CPA.com Generative AI Toolkit
Generative AI is a subset of the broader articial intelligence category that can perform more complex steps. It has reshaped the technological
landscape by offering advanced capabilities across a variety of domains. Beyond just Large Language Models (LLMs, see Glossary) like ChatGPT,
generative AI encompasses a broad range of modalities:
Image and Video
AI can produce, modify or interpret images
and videos, enabling applications such as
understanding visual tables, flowcharts, and
worksheets as well as art creation and deepfake
generation.
Sound
These models generate or analyze
sound waves, useful in voice synthesis,
music production, and audio anomaly
detection. Speech to text models are
also gaining in popularity.
Text
These models generate or understand human
language, enabling tasks like conversation,
summarization and translation.
Computer Code
AI can take a description of a task and
write code to achieve it, or teach the
user how a code example works.
5CPA.com Generative AI Toolkit
Natural Language Processing has been around for a while, most
recognizably in the chatbots that pop up on many websites. However,
generative AI has advanced quickly in recent months due to new
innovations in the LLMs on which it is based.
Although the content it can generate is important, the real power of
generative AI lies in its ability to understand complex instructions and
break them down into individual tasks it can perform in response,
according to Aaron Harris, CTO of Sage. There are also limits and
serious risks associated with generative AI. The security of any data
that is shared with this technology is a major concern, particularly when
it comes to sensitive client information. At least in its initial iterations,
generative AI has also been found to get things wrong and even make
things up—or hallucinate. Its condent, articulate and well-reasoned
answers may mask this signicant failing.
This toolkit is intended to help accounting and nance professionals to
understand generative AI along with its risks, challenges, opportunities
and limitations. It includes various use cases that demonstrate
how generative AI can be put to work for accounting and nance
professionals. It also provides a 7-step plan to help build your AI
strategy, security & risk considerations to help rms anticipate and
minimize related threats, and a glossary of terms.
Be sure to visit CPA.com/Gen-AI for explainer
videos on each use case, additional learning
opportunities and more.
Nearly every accounting technology provider
today is exploring if and how generative AI can be
integrated into their existing solutions to address
rms' top challenges. When harnessed the right
way, these capabilities present an enormous
opportunity for the accounting profession to
deliver higher value services.”
— Erik Asgeirsson, CEO, CPA.com
6 CPA.com Generative AI Toolkit
Organizations have embraced a variety of forms of digital transformation in recent years, many of those being driven by technological advancements,
as well as the disruptions and innovations caused by global events. Generative AI is producing new applications and opportunities that are enabling
wider and speedier transformation. The emerging tech has been compared to the internet and smartphones in terms of the impact it will have on
daily life and business. In fact, Open AI released a study highlighting occupations with the greatest risk of being disrupted, noting that 100% of tasks
for accountants are currently exposed.
Driving digital transformation
• Automating workflows from beginning to end.
• Handing off numerous tasks to digital assistants that function
as resources within the organization or for clients.
• Hyper-personalizing marketing content, taking on routine
sales tasks and conducting better lead identication to
optimize business development and growth.
• Allowing companies to design, test and bring new products to
market more quickly.
• Speeding uses of tools, such as 3D printing and
computer-aided design, and identifying and enhancing design
options.
• Minimizing the accounting profession’s stafng shortage by
taking on a wide range of tasks, even beyond the repetitive
steps that many AI tools are already handling.
• Analyzing and identifying potential instances of fraud and
other risk anomalies.
• Potentially making it easier to address environmental, social
and governance (ESG) issues.
When it comes to digital transformation, among the
many things that generative AI can enable are:
6 CPA.com Generative AI Toolkit
7CPA.com Generative AI Toolkit
“Part of the challenge is that it’s very difcult to see all the ways that
generative AI will evolve and emerge,” Harris said. “It is so powerful
and it’s moving really fast. There are already hundreds of companies
getting venture funding and doing really impressive things with it.”
CB Insights reports that investment in generative AI startups reached
$14.1 billion in equity funding in 86 deals during the rst half of 2023.
How will it work in practice? Within CPA rms, customized generative
AI systems could be designed to immediately provide a list of new
changes to the tax code or compare the details of a new standard
with an old one. Firms could use it to create personas for internal use
that have expertise in particular service lines, technologies or other
areas, thereby signicantly cutting down research time. To prevent
inaccuracies or hallucinations, a rm can limit the content that the
technology is trained to pull from so that it is referencing only completely
reliable sources, such as the Internal Revenue Code or the text of other
standards, laws or regulations. Sage’s Harris envisions specialized
generative AI agents that will be trained specically to manage the
accounts payable workflow or interact with vendors. An AI agent
designed for ofce productivity might help a rm perform analysis
Part of the challenge is that it’s very difcult to see
all the ways that generative AI will evolve and and
emerge. It is so powerful, and it's moving really fast."
— Aaron Harris, CTO, Sage
and research, then turn its ndings into a professional document or
build PowerPoint slides. To create a more thorough, personalized and
competitive proposal, rms can use generative AI to gather publicly
available information on a prospect’s industry, market and competitors,
making it easier to highlight the rm’s understanding of the prospect’s
business and the many ways that it can meet their needs.
7CPA.com Generative AI Toolkit
8 CPA.com Generative AI Toolkit
CPA.com Seven steps to building an AI strategy
Imagine that someone who plays baseball suddenly switches to basketball. Their underlying athletic ability remains the same, but they have to
relearn what it takes to succeed in the new arena. When we learn a new sport, we not only learn the rules of the game, but we must also become
comfortable with new movement and develop new muscle memory. “That change is similar to the mindset shift necessary with generative AI,” said
innovation expert and CPA.com collaborator Pascal Finette, co-founder of be radical. It’s not just a new way of using technology. It’s a new way of
thinking about technology and work in general. That’s why, in understanding the rules of the game with articial intelligence–and more specically
generative AI–it’s important to avoid trying to slot them into existing expectations. Instead, we need to be open to developing new ideas, uses and
capabilities.
Tools like ChatGPT and Bard are very easily accessible, so pick
one, create an account and start experimenting to understand the
technology’s potential. Considering the rapid pace of change in this
space, it’s a good idea for rms to get their feet wet as soon as possible,
just don't put personal identiable information in these...yet.
Experiment
STEP 1
Given the accelerated advancement in this space, attempting to create
a ve-year plan for generative AI use is futile. At least at rst, a two-
year strategic plan may be optimal.
Focus on the near-term
STEP 2
On the one hand, many companies are likely still in the learning phase
themselves, so rms should not shortchange their learning curves
because they worry about staying ahead of client expectations.
However, rms should expect that generative AI will quickly become
clients’ user interface of choice. Get a sense of where clients are, then
plot out where the rm needs to be and how generative AI can address
both rm and client needs.
Consider client expectations
STEP 3
8 CPA.com Generative AI Toolkit
The steps below outline a way to shift your mindset and help your organization to develop its own AI strategy.
9CPA.com Generative AI Toolkit
Since the systems available right now work well with text, identify
the ways that the organization is spending time and energy on text-
based work, such as responses to clients, business development
and marketing emails. Then determine how generative AI can reduce
current effort and costs. Revisit this step regularly as the technology’s
capabilities mature.
Some organizations have questioned whether it’s best to ban generative
AI use to limit the associated risks to privacy and condentiality.
Unfortunately, prohibitions may backre if staff members are accessing
the technology on their own, unsecure devices. In addition, rms
should want to promote staff capabilities with articial intelligence.
To establish guardrails, circulate an acceptable use policy as early
as possible in the process. Among other things, the policy should
prohibit uploading or asking questions about client data within a public
generative AI tool. Exposing sensitive client or rm information is a
risk because anything entered into ChatGPT, for example, becomes
public domain and is used in the system’s continuous training of its
machine learning model. To address staff research questions related
to client data, a better ultimate choice might be to develop an internal-
use bot that can provide answers based on reliable authoritative or
locally stored data. (For more information and advice on acceptable
use policies, see the Risk & Security Considerations later in this toolkit.)
Put the technology’s best features to work
Dene acceptable use
Trust will be vital when offering products or services
associated with AI, as clients are handing over responsibility
for completing a task that’s critical to their businesses. If
there’s not perfect condence in an outcome, generative AI
projects will require human oversight.
Because generative AI output may include biases, errors or
hallucinations, a human should review and ensure the appropriateness
and accuracy of any content that is being used in decision-making
or shared with clients. As an example, the system could be asked to
identify emails that are marked as urgent and to analyze and report on
the content. The user can then ask the system to create responses,
stipulating their length and tone. The nal, and most important step,
would be for someone to review the emails before they are sent to
ensure they have the right tone and accurate content for each recipient.
Ensure that humans maintain a key role in the process
Before building your own generative AI tools, ask current and potential
vendors about their existing offerings or plans. “I can guarantee that
nearly every solution provider has something in the hopper regarding
use of AI in the accounting, tax or audit software space,” Finette said.
Determine how new or emerging tools can be incorporated into the
rm’s strategic planning.
Research what solution providers are building
STEP 4 STEP 6
STEP 5 STEP 7
10 CPA.com Generative AI Toolkit
CPA.com Use Cases: Practical application of
generative AI for accounting and nance
Organizations can make the most efcient use of generative AI if
they understand how to ask the right questions. These use cases
illustrate the kinds of requests users might make in different situations
and examples of how to phrase them to get the best outputs. These
examples are referred to as PROMPTS. For each use case, you
can watch an explainer video and access swipeable prompts at
CPA.com/GEN-ai. Note that these use cases can be used with most any
GPT system; we have identied specic solutions like Claude, ChatGPT
Code Interpreter, etc. due to functionality. There are hundreds, if not
thousands, of generative AI tools on the market. For these use cases,
we have leveraged low-cost/no-cost solutions that anyone can easily
access.
The use cases have been categorized based on the practice area or
business subset that would most likely benet from its implementation.
Finance Use Case: Structured Data Extraction
In a perfect world, systems would talk to each other and sync data
seamlessly. Unfortunately, working with many systems still requires
exporting, importing and reformatting data. Language models are
particularly adept at taking unstructured data—either from an export
or even simply copied and pasted from a PDF report—and turning it
into structured data. Here’s how to use ChatGPT to take unstructured
data in various formats, quickly reformat it into a structured
comma-separated values (CSV) le, and validate that output is correct.
Example prompts:
•
The following is an accounts payable detail report. Structure the
data into a CSV le with columns for Invoice Date, Vendor, Amount
and Currency.
• The following is a bank statement. Structure the transactional data
into a CSV le with columns for Transaction Date, Description, Amount
(which is the sum of Debit minus Credit), and Available Balance.
Please note, any examples that reference uploading data will require
all identiable information to be removed prior to use.
Companion Videos
CPA.com/genai-practical-use-cases
10 CPA.com Generative AI Toolkit
11CPA.com Generative AI Toolkit
We just have to start. Find a way to make
the use of AI a habit."
— Jason Staats, Founder, Realize
• The following is an audit trail report. Return all the data from John
Smith user into a CSV le with the columns Date, Time, Action and
IP Address.
Tax Use Case: Legislation Research Assistant
Navigating large documents like legislation can be difcult when
relying solely on text-based search. While Ctrl+F may get you to the
right section, it will only work if you use the exact keywords. Large
language models, however, enable a more general semantic form of
search that can surface relevant sections more flexibly. By prompting
the model to return verbatim excerpts, you can mimic the way a
human researcher gathers information in order to reach conclusions
more efciently. This leverages an LLM called Claude for its ability to
handle large documents up to 75,000 words. This human-in-the-loop
approach combines the breadth of AI with human judgment.
Example prompts:
•
Attached is the text of Secure Act 2.0. I’m researching the impact
it has specically on 401(k) plans. In a bullet point list, provide the
most relevant verbatim excerpts from the document, with a citation
at the end of each quote to the exact section to nd the excerpt in
the bill.
• Attached is proposed legislation HR3937. I’m researching the impact
it has specically on 1099 reporting. In a bullet point list, provide the
most relevant verbatim excerpts from the document, with a citation
at the end of each quote to the exact section to nd the excerpt in
the bill.
• Attached is the text of IRS Revenue Procedure 2023-8. I’m researching
whether there is any guidance surrounding audit protection in the
rev proc. In a bullet point list, provide the most relevant verbatim
excerpts from the document, with a citation at the end of each quote
to the exact section to nd the excerpt in the bill.
11CPA.com Generative AI Toolkit
12 CPA.com Generative AI Toolkit
Advisory Use Case: AI-enriched KPI Analysis
The programming capabilities of ChatGPT's Code Interpreter model
make it well-suited for analyzing large volumes of data. This can be
particularly useful in advisory contexts when evaluating a dataset
against key performance indicators (KPIs). Code Interpreter can help
ensure nothing gets overlooked by thoroughly examining the full
breadth of the data, even aspects that may not be directly captured by
the predetermined KPIs.
Specically, Code Interpreter could be prompted to review the complete
dataset and identify any notable trends, patterns or insights that the
existing KPIs fail to highlight. Based on this analysis, Code Interpreter
could then suggest potential new KPIs to better encapsulate these
additional ndings. Taking advantage of ChatGPT in this way enables
the development of KPIs that more fully and accurately reflect the
meaningful results within a dataset.
Example prompts:
•
Attached are several accounts receivable reports. The company has
two accounts receivable KPIs: Days Sales Outstanding, which was
41 this month, and Average Days Delinquent, which was 6. Review
the data to identify any advisory insights that these two KPIs may
not take into account and, if there are any omissions, how these KPIs
could be improved.
• Attached is the general ledger detail from last month. I would like
you to analyze three KPIs: Salaries as a percentage of revenue, which
was 20% this month, Asset Turnover, which was 0.79, and Operating
Cash Flow, which was 1.29. Review the data to identify any advisory
insights that these three KPIs may not take into account, and if there
are any omissions, how these KPIs could be improved.
• Attached is a set of sales reports. The sales team has two KPIs:
Average Purchase Value, which was $19,582 this month, and Lead-
To-Close %, which was 17.52. Review the data to identify any advisory
insights that these two KPIs may not take into account, and if there
are any omissions, how these KPIs could be improved.
13CPA.com Generative AI Toolkit
Audit Use Case: Interview Transcript Assistant
Audit software centralizes workpaper documentation, but important
client conversations often are not captured for the workpaper file.
Now, AI meeting assistants can provide near real-time, accurate
meeting transcriptions. By leveraging interview transcripts and using
large language models to search them, audit teams can quickly find
specific information across extensive transcripts.
Example prompts:
•
In this transcript, was anything discussed surrounding inventory
receiving procedures?
• In this transcript, was an update provided on the conrmation letter
from National Bank?
• In this transcript, were user permissions for the accounts receivable
module discussed?
Internal Business Use Case: Recruiting Agent
AI agents use LLMs to reason through tasks. Some are even capable
of navigating applications. Human users prompt the agent to perform
a task, and the agent plans out how to complete it, critiques its own
plan and continuously course-corrects until the task is complete. This
use case leverages Hyperwrite's Personal Assistant agent (a Chrome
browser extension) to navigate hiring websites to identify qualied
candidates for an open role at the organization.
Example prompts:
•
Navigate this job board and nd me 10 applicants whose
qualications are similar to those of Jane Smith. They must be CPAs
with a minimum of 5 years experience.
• Review J o hn S mith’s con nections on LinkedIn to identif y any potential
candidates for the position I’m hiring for. I’m looking for junior- and
senior-year students studying accounting within 50 miles of Austin.
Prioritize candidates who actively engage on LinkedIn. Once you’ve
found the three best candidates, draft a message to John for me to
review, asking John to make an introduction.
• Locate the following individuals on LinkedIn: [list]. Identify if they’re
connected with anyone in my network. If so, provide a couple people
in my network who could make an introduction to each individual.
Join the conversation
These are just a few ways that accounting and nance
professionals can use the power of generative AI.
What are some ways you would train the platform to
be more resourceful for your organization or clients?
Join the conversation at #DigitalCPA.
#DigitalCPA
13CPA.com Generative AI Toolkit
14 CPA.com Generative AI Toolkit
Challenges and limitations of generative AI
As the security checklist in the next section makes clear, generative AI
poses a number of challenges that rms must be prepared to mitigate.
The newer and more popular an open-source solution is, the less mature
its security is. It's especially important to address security concerns
when working with vendors whose products embed generative AI
because, in the headlong rush to bring new tools to market, some may
not have addressed all security issues or be working in private data
pools. Systems can be especially vulnerable because:
• Bad actors can use generative AI to guess passwords, enhance
phishing attempts, perform CAPTCHA-cracking and build better
malware.
• Hackers can invade a company’s system through unsanctioned
generative AI tools used by employees.
Questions to consider if the rm is working
with a SaaS provider:
• What protections will be in place if staff upload client data?
• Is data identiable?
• Is the data encrypted?
• Where is the data stored?
• How is data shared with public LLMs? As a general rule, no client
or business information should be entered into a public LLM.
• What kind of verication is done on facts and math calculations?
The wide-open generative AI universe promises a range of possibilities, but with it lies challenges. Here are some of the current limitations and
security risks that rms need to consider when looking to integrate the technology into their practices.
Security is paramount
14 CPA.com Generative AI Toolkit
15CPA.com Generative AI Toolkit
Communication and critical thinking
skills will matter
Developing communication and critical thinking
skills has long been a priority for rms, and they
will be especially important when using generative
AI. Since the technology is language-based, rms
will have the best outcomes if their people have the
necessary skills to review and improve the content
that the technology creates and spot incorrect
nuances, biases or gaps in logic included in the
responses. They should also be able to identify, for
example, that a stiffly written communication is not
aligned with a rm’s tone and client relationship, or
that an email or blog is too long or doesn’t contain
enough detail to be valuable to clients.
Today’s LLMs produce generated output, not computed
answers, which means you shouldn’t trust them with
math or nancial analysis. Expect this to improve in the
next 12-18 months.”
— Jeff Seibert, CEO, Digits
15CPA.com Generative AI Toolkit
16 CPA.com Generative AI Toolkit
Be more specic
If you don’t like the rst answer, decide if
you need something that’s shorter or more
detailed or that is written in a different tone.
You can use terms that are broad like “more
relaxed” or targeted (e.g. “in the tone of my
last letter to client X”). A perfect example of
this is our own message from ChatGPT at
the opening of this toolkit.
Be specic
Don’t just ask the system to write a client
proposal. Instead, ask it to set out a specic
number of points for the proposal using
information from particular sources such
as recent presentations your rm has given
or available client or industry and market
information. Providing the model with
examples helps get a better output.
Have a dialogue
The systems will keep delivering new
iterations based on your instructions, so
don’t give up if the rst couple of tries don’t
work. You can ask the system to tell you what
information or instructions it needs to create
content—such as a proposal or cover letter—
and you can then supply the details it needs.
Prompts are important
Generative AI is only as good as the prompt that drives it. With that in mind, some organizations are hiring prompt engineers—people with the skills
and backgrounds needed to communicate with generative AI systems, according to the Wall Street Journal. Prompt engineering will come in handy
at rms of any size. Zunie Nguyen’s seven-person rm, Yogi CPA, has provided a training session on prompt engineering, sharing useful examples
and offering guidelines on how to write new ones.
Among the best practices for prompts:
16 CPA.com Generative AI Toolkit
17CPA.com Generative AI Toolkit
McKinsey predicts that by 2030, activities that required
up to 30% of the hours currently worked across the U.S.
economy could be automated, a shift that generative
AI will accelerate. The consultants expect, however,
that generative AI will enhance the way that business
professionals work rather than replacing many jobs in
this area outright.
The great equalizer
AI will make you a superhuman, you will become
even more valuable to clients. The only problem
is if you don’t evolve yourself and use it.”
— Pascal Finette, co-founder, be radical
18 CPA.com Generative AI Toolkit
Security and risk considerations:
Safeguarding validity, privacy and transparency
As the utilization of generative AI continues to expand, understanding
the associated security implications becomes crucial for businesses
across industries.
There are 5 key risk considerations for the accounting
profession, both for internal and client use:
a. Understand the information you’re ingesting into AI and its
sensitivity. Public tools such as ChatGPT often rely on user input
for its own training.
b. “De-identify” (sanitize) personal information before ingesting it
into both internal AI systems and public tools. Personal identiable
information is anything that can be used to identify an individual
person.
c. Have a conversation with the provider of your enterprise AI
platform to exercise control over usage, retention, and disposal of
information stored on the platform.
a. Engage legal counsel to identify any laws that may apply which
require disclosure use of AI.
b. If a client’s condential information will be shared with a third
party vendor, including AI, provide written disclosure to and obtain
specic consent from the client in the appropriate format before
the condential information is shared.
c. If client information is used to train the AI model, then consider
explicitly stating such in your disclosure to and consent from the
client.
d. If disclosure is not specically required, consider still disclosing
a rm's use of AI in the engagement letter. Include AI in the
disclosure if it is anticipated AI tools will be used to render the
agreed upon services.
e. An addendum to an engagement letter or other written consent
from the client can be created if it is later determined that AI tools
will be used.
Privacy
Transparency
1
2
19CPA.com Generative AI Toolkit
The way in which we ask questions are often biased, so
sometimes an AI can respond to a question in a way which
conrms our bias. We need to be thoughtful about both inputs
and outputs (responses).
Avoiding Bias
3
4
a. Accounting professionals have a responsibility to monitor
answers that are coming out of generative AI to ensure
accuracy.
b. Discuss with general counsel necessary documentation
of a review process for AI output.
Human Review
a. Begin with the rm’s existing “acceptable use policy,”
which deals with access and use of company hardware
and software, VPN access, client privacy concerns, etc.
Limiting Use Cases
5
Data Exposure and Leaks
The transformative capabilities of generative AI have many risks. One
primary concern is the exposure of sensitive information:
Employees Exposing Condential Information
Case Study: In March 2023, Samsung employees inadvertently exposed
condential data to OpenAI via ChatGPT. This included source code,
internal meeting notes, and product roadmaps. ChatGPT uses these
inputs to further train their models, which creates the risk that this specic
information might leak to the wider ChatGPT user base. Samsung has
since banned the use of ChatGPT and other generative AI systems on
company-owned devices and internal networks. The leaks have raised
concerns about the security risks of using AI tools in the workplace
without proper safeguards.
19CPA.com Generative AI Toolkit
b. Dene ordinary use for AI tools, such as
writing marketing language vs. extraordinary
use such as using AI to produce counsel or
professional advice.
c. Ask, who in the rm should have access
to AI tools.
Additional risk considerations,
examples and recommendations
for mitigation:
20 CPA.com Generative AI Toolkit
Mitigation: Reviewing the terms of use for the AI systems
– some vendors explicitly do not train their models on
your data, for others it is part of the T&Cs. Clear policies
and procedures in place for the use of AI tools. Regular
audits of interactions with external systems, coupled with
stringent access controls, to prevent unintentional data
sharing.
Data Leakage from LLM Responses
Large Language Models can sometimes disclose sensitive
or proprietary information in their outputs. This may
result in unauthorized data access, privacy violations, and
potential security breaches.
Mitigation: A review layer to vet AI responses, along with
the usage of data sanitization protocols and implementing
strict user guidelines, can signicantly reduce the risk of
unintended data disclosure. It’s worth noting that some
vendors start to indemnify their users from any claims
(e.g. Adobe with their visual GenAI).
Insufcient Access Controls
An oversight in the deployment of LLM models can lead
to unauthorized users gaining access and model and data
exposure.
Mitigation: Multi-factor authentication, role-based access
controls, and periodic access reviews ensure only authorized
personnel can interact with the AI systems.
Infrastructure Vulnerabilities
The intricacy of enerative AI systems can sometimes introduce specic
vulnerabilities that attackers can exploit:
Prompt Injections
Attackers can supply crafty inputs to manipulate LLMs, causing
unintended actions or even bypassing lters.
Mitigation: Rigorous input validation and ltering mechanisms.
Consider the use of a whitelist approach where only certain
predened types of inputs are permitted.
Inadequate Sandboxing
AI models, when inadequately sandboxed, can allow unauthorized
access to the underlying IT environment. Consequences can include:
• Insecure Output Handling
When an LLM output is accepted without scrutiny, exposing
backend systems.
• Unauthorized Code Execution
Using malformed natural language inputs to execute malicious
code or commands.
• Server-side request forgery (SSRF) Vulnerabilities
Accessing restricted resources, internal services, APIs, or data.
• Exploitable Plugin Design
Some LLM plugins can have insecure inputs and insufcient
access control.
20 CPA.com Generative AI Toolkit
21CPA.com Generative AI Toolkit
Mitigations:
• Deploy AI models within isolated environments, ensuring they
don’t have unwarranted access to external systems.
• Regularly update and patch systems to defend against known
vulnerabilities.
Improper Error Handling
Errors, when not properly masked, can reveal sensitive data,
including personal information.
Mitigation: Customize error messages and debugging output to
ensure they remain generic and non-revealing. Conduct thorough
testing to identify and rectify information leak points.
Model Denial of Service
Given the resource-intensive nature of generative AI models,
attackers might initiate resource-heavy operations causing service
degradation or high costs.
Mitigation: Rate-limiting and monitoring systems to detect and
alert on unusual activity patterns.
Data Integrity Threats
Generative AI models are only as good as the data
they're trained on. Any compromise in data integrity
can lead to serious ramications:
Training Data Poisoning
Deliberate tampering with training data can introduce
biases, backdoors, or vulnerabilities.
Mitigation: A secure and controlled environment for
data collection and curation. Regularly review and
validate datasets to ensure their quality and integrity.
Employ anomaly detection to identify unusual patterns
in training data.
Model Theft
Theft of proprietary AI models not only results in
economic losses but also exposes sensitive information
and intellectual property.
Mitigation: Encryption techniques for model storage
and transmission. Use secure enclaves or trusted
execution environments to protect models during
inference. Additionally, utilize model watermarking
techniques to trace unauthorized usage.
21CPA.com Generative AI Toolkit
22 CPA.com Generative AI Toolkit
Compliance and Ethical Concerns
The wide-ranging capabilities of generative AI also introduce potential
compliance and ethical challenges:
Over-reliance on Generated Content
Generative AI models can sometimes produce outputs that aren't
grounded in their training data, "hallucinations". Excessive trust in LLM-
generated content can result in misinformation, miscommunication,
legal issues or security vulnerabilities.
Mitigation: Manual review processes for critical AI-generated content.
Educate users on the potential risks and encourage a culture of
verication.
AI Alignment and Excessive Agency
LLMs can exhibit behaviors misaligned with intended use cases,
leading to unintended consequences.
Mitigation: Regularly test and validate AI systems against real-world
scenarios to ensure alignment. Restrict excessive permissions and
autonomously granted functionalities to LLMs.
Bias and Discrimination Risks
AI responses may sometimes demonstrate biases, risking violations of
anti-discrimination laws.
Mitigation: Regularly audit models for bias. Use diverse training
datasets and employ fairness-enhancing interventions during model
development.
Intellectual Property and Copyright Implications
AI-powered tools are often trained on massive amounts of data and
are usually unable to provide sources for their responses. Copyrighted
resources, such as books, magazines, and academic journals, may
be included in some of the training data. Using AI output based on
copyrighted materials without citation could result in legal penalties.
Mitigation: Provide disclaimers when using AI-generated content.
Implement fact-check mechanisms to detect potential copyright
infringements.
Licensing Restrictions and Content Use
Another consideration involves using content that may be under a
license or other agreement. The terms of these agreements might not
explicitly permit or could even prohibit the use of such content with AI
tools, either fully or without citation.
Mitigation: While this area remains legally ambiguous, it's worth
consulting with legal counsel on this matter.
Legalities Surrounding Chatbot Use
Using AI-powered chatbots to answer customer inquiries without
proper disclosures may invite legal penalties.
Mitigation: Always disclose AI interactions to users. Stay updated on
regional laws and regulations concerning chatbot and AI use.
22 CPA.com Generative AI Toolkit
23CPA.com Generative AI Toolkit
Open-Source Models Risks
Unvetted Code & Components
Open-source models and their infrastructure are more likely
to contain code or components that haven't undergone
rigorous security checks. Malicious actors can introduce
vulnerabilities or backdoors.
Mitigation: Regularly audit the open-source code for
vulnerabilities. Utilize static and dynamic code analysis tools
to inspect for potential threats.
Inconsistent Updates & Maintenance
Some open-source projects might not receive regular
updates, leaving them exposed to known vulnerabilities.
Mitigation: Ensure you're using actively maintained open-
source projects. Regularly check for updates and patches,
and integrate them promptly.
Exposure to Pre-trained Data
Open-source models, especially pre-trained ones, might have
been exposed to unvetted data sources. This could introduce
biases or make the model respond in unpredictable ways.
Mitigation: When possible, ne-tune the open-source models
using your own vetted datasets. This can help in overriding
potential biases introduced during the initial training.
Additional Attack Vectors and Misuse
The vast potential of generative AI opens avenues for misuse, both in
terms of external attacks and malicious applications:
Supply Chain Vulnerabilities
The lifecycle of generative AI can be compromised by vulnerable
components or services. The AI ecosystem is rapidly evolving, creating
multiple opportunities for malicious actors to attack.
Mitigation: A thorough vetting process for third-party datasets, pre-
trained models, and plugins. Ensure a regular update and patching
schedule. Conduct security audits on all integrated components.
24 CPA.com Generative AI Toolkit
Data Privacy Concerns
For enterprises who develop their own generative
AI models, the process requires large amounts of
training data, raising concerns if breaches occur
and threat actors gain access to sensitive data,
including personal information.
Mitigation: Robust data encryption, both in transit
and at rest. Implement strict access controls on
data storage locations, and ensure compliance with
relevant data protection regulations for the usage,
retention, and disposal of stored information.
Malicious Use of Deepfakes
AI's capability to create realistic fake video and
voice cloning can be used to impersonate key
personnel and obtain sensitive information.
Mitigation: Educate employees about the risks
and signs of deepfakes. Implement detection tools
that leverage AI to recognize manipulated media.
Personalized Social Engineering Attacks
AI can craft highly targeted phishing or scam
messages, enhancing their efcacy.
Mitigation: Train employees and stakeholders on
the latest social engineering tactics.
Mitigation Strategies
Beyond addressing specic vulnerabilities, a broader, proactive approach is
essential to safeguard against evolving threats in the generative AI landscape:
Continuous Security Training
Educate teams about the evolving threat landscape and best practices in handling
and deploying AI tools. A well-informed team can be the rst line of defense against
potential breaches.
Robust Infrastructure Security
While the focus might be on AI, ensuring that the foundational IT infrastructure is
secure is paramount. Regularly update, patch, and monitor underlying systems to
deter potential breaches.
Third-party Vendor Audits
Regularly review and audit third-party AI vendors. Ensure they adhere to security
best practices and are transparent about their data handling and processing
methods. Consider obtaining an assurance report from a licensed CPA rm.
Incident Response Plan
Prepare for potential security incidents by having a response plan in place. This
plan should detail steps to contain the threat, communicate with stakeholders, and
recover operations.
Feedback Loops
Establish feedback mechanisms with users of AI systems. They can often provide
valuable insights into potential system misbehaviors or vulnerabilities.
25CPA.com Generative AI Toolkit
Limit AI Autonomy
For high-risk scenarios, consider limiting the autonomy
of AI systems. Manual oversight or approval mechanisms
can act as safety checks against unintended AI actions.
AI Adoption and Risk Levels
The risks associated with generative AI vary depending
on a rm's engagement with the technology. Different
threats emerge based on whether a company merely
uses tools like ChatGPT for basic tasks, leverages
vendor-specic applications for specialized operations,
or actively develops and trains their own models using
client data.
25CPA.com Generative AI Toolkit
The Double-Edged Sword of AI
As generative AI tools enhance productivity and open new possibilities, they also
present novel risks. It's imperative to approach them with a blend of enthusiasm and
caution.
Hacker Productivity
Just as businesses benet from advancements in AI, so do malicious actors. The tools
and methods that enhance legitimate operations can also streamline illicit activities.
Traditional Cybersecurity
Embracing new AI-driven solutions should not come at the cost of sidelining traditional
cybersecurity measures. Both old and new defenses must work in tandem to ensure
robust protection.
26 CPA.com Generative AI Toolkit
Algorithms
In computer science, this refers to tools that can perform calculations,
data processing and automated reasoning tasks.
Application Programming Interface (API)
A way for two or more software programs to talk to each other.
Articial Intelligence (AI)
Computer systems that can perform tasks that would typically require
human intelligence to complete.
Bots
A computer program that can interact with users or other systems.
Since they are automated, they can follow specic instructions without
continuing human oversight.
Data lakes
A central repository that can store large volumes of raw data in its
native format, whether structured, semi-structured or unstructured,
and a place to organize large volumes of data from a variety of sources.
Data can be ingested faster into a data lake than into a data warehouse.
Deep Learning (DL)
This subset of machine learning learns from multiple layers of neural
networks, enabling it to perform more complex tasks.
Generative AI (GenAI)
A type of AI that can create new content. GPT or generative pre-trained
transformers. A transformer is a neural network that enables AI tools
such as ChatGPT to generate content similar to what a human might
create.
Hallucination
This can occur when an AI system is asked a question it can’t answer,
so it makes up false or nonsensical information instead. In one case,
when asked to cite precedents to be used in a court case, the system
responded with cases that did not exist. This tendency is one major
risk in working with generative AI.
Large Language Models (LLMs)
The foundation of generative AI tools related to language, these models
are trained on vast amounts of data, enabling them to create original
output in response to prompts, such as text, images and computer
code.
Glossary of key terms
Here are a few key terms related to generative AI that will help you navigate the emerging tech:
27CPA.com Generative AI Toolkit
Machine Learning (ML)
Describes how machines can perform and rene tasks without specic
instructions, using algorithms and statistical models to nd patterns in
the data.
Natural Language Processing
This makes it possible for LLMs to understand and respond to text or
voice data, rather than computer code.
Neural Networks
Systems modeled on the workings of the human brain. Generative AI
uses them to recognize data patterns and structures and then create
new content based on what it learns.
Robotic process automation (RPA)
RPA can mimic human actions by taking on basic, repetitive tasks,
such as data processing. generative AI, on the other hand, can simulate
human intelligence, with an ability to learn and reason.
27CPA.com Generative AI Toolkit
Additional Resources
Access videos and additional
resources at
CPA.com/Gen-AI
© 2023 CPA.com. All rights reserved. CPA.com and the CPA.com logo are trademarks and service marks of CPA.com. The Globe Design is a trademark of the Association
of International Certied Professional Accountants and is licensed to CPA.com. All rights reserved.
About CPA.com
CPA.com’s mission is to empower accounting professionals in the digital age—to be the advisor’s
advisor.Technology’s impact on business model transformation and practice evolution is undeniable.
It’s creating new practice areas and driving rapid change across rms. We aim to help practitioners
successfully navigate this change. We do this through numerous strategic initiatives focused on driving
the transformation of practice areas, advancing the technology ecosystem for the profession and leading
technology research and innovation efforts for practitioners.
CPA.com’s research and innovation efforts include the CPA.com Startup Accelerator, ecosystem networking
events, category symposiums and the curation of future-focused developmental and educational resources
like toolkits and special reports intended to demystify emerging technologies within the context of their
impact on the accounting profession.
A subsidiary of the American Institute of CPAs, CPA.com is also part of the Association of International
Certied Professional Accountants, the world’s most influential organization representing the profession.
For more information, visit CPA.com.
