McKinsey Direct

Generative AI in finance:
Finding the way to faster,
deeper insights
Generative AI technologies can automate time-consuming tasks for finance
professionals, but can they be trusted to give the right answers?

This article is a collaborative effort by Steve Eklund, Lisa Kaufman, Avani Kaushik, Andrii Kurdiuk, Jan Svoboda,
and Edward Woodcock, representing views from McKinsey’s Operations Practice.

© Getty Images

January 2023
 Business leaders are excited about generative The proof of concept was designed to address
AI (gen AI) and its potential to increase the a key pain point for the business. Finance
efficiency and effectiveness of corporate analysts were often overloaded with requests for
functions such as finance. A May 2023 survey information from managers, and the managers
of around 75 CFOs at large organizations found were frustrated that it could take several days
that almost a quarter (22 percent) were actively to get answers to relatively straightforward
investigating uses for gen AI within finance, questions. Solving that problem helped to
while another 4 percent were pursuing pilots of create interest in, and enthusiasm for, gen AI
the technology. technology (Exhibit 1).

That enthusiasm has been tempered by 2. Assemble the right tools and capabilities for
concerns over safety, privacy, accuracy, the job
copyright, and social manipulation. In the A diverse array of gen AI solutions is available,
case of finance, where numerical data and ranging from on-premises options to cloud-
the accuracy of mathematical operations are based solutions and the choice of deployment
fundamental, there is also concern over the should reflect the data sensitivity and specific
possibility that gen AI systems could produce requirements of the use case. Taking advantage
inaccurate or misleading information, a of existing IT and digital infrastructure may
phenomenon known as “hallucination.” shorten development time and can also ensure
that gen AI solutions integrate existing workflows
with minimal disruption.
Developing a finance assistant using
generative AI After aligning on its preferred technology stack,
One European consumer goods company the company assembled a highly integrated cross-
recently developed a proof-of-concept gen AI functional team to drive the development process.
assistant for finance professionals and business In addition to data scientists and IT engineers,
users. The new tool was built in about six weeks that team included senior members of the finance
by a team of data scientists, engineers, and team, who helped to shape the design and acted
finance experts. It allows users to ask questions as superusers during the development and testing
about financial performance in everyday phases. Their input was critical to ensure that the
language and rapidly receive answers that aid new tool met the organization’s quality, reliability,
them in understanding and interpreting the data. and usability requirements.

This proof-of-concept exercise was only the 3. Put humans in the loop, and keep them there
company’s first step in applying gen AI in the The company also paid significant attention
finance function, but it offers several useful to design of the user interface, maintaining
lessons for organizations seeking to capture the a consistent structure and “look and feel” to
benefits of these technologies while managing provide a straightforward and streamlined user
the risks. experience. One early request from users was for
the ability to drill down rapidly through finance
1. Start with high-impact, internally facing data to get to the information they needed. To
use cases. facilitate this, the tool offers instant graphical
Gen AI pilot projects should solve meaningful visualization of results, so staff can quickly absorb
problems for the organization without creating and interpret its answers. All data is also made
new ones. That requires a high degree of control available in a tabular format that can be shared
over the data that will feed into the model and with other analytical tools.
a design that minimizes security challenges or
the potential for data misuse. In this case, the During development, the company recruited a
company chose a single in-house dataset and a group of superusers from its finance function as
user group that was already familiar with it. product testers. They were encouraged to apply

2 Generative AI in finance: Finding the way to faster, deeper insights

Web <2024>
<Gen-AI-in-finance>
Exhibit
Exhibit <1> 1of <1>

The
Thegen
genAI
AIfinance
financeassistant
assistanthelps
helpsusers interrogate
users financial
interrogate data
financial with
data with
speed and precision.
speed and precision.
Illustrative examples

Gen AI SG&A Copilot Gen AI SG&A Copilot

What is the total opex by country for 2021 What are the total cost element drivers in
in Region 1? Country A for 2021?

The total opex by country for 2021 in The total cost element drivers in Country A
Region 1 is the following: for 2021 are the following:

Total opex by country, Total cost element drivers,

Region 1, 2021 Country A, 2021

Total Total
opex, opex,
€ €

A B C D 1 2 3 4 5 6
Country Cost element driver
View explanation View explanation
Source: Finance data Source: Finance data

Send a message… Send a message…

McKinsey & Company

the tool in real-world situations and to assess computations. Engineering teams are working
and provide feedback on its reliability and hard to address this limitation. One approach is
usability. to build a hybrid model, using the LLM in ways
that, build on its strengths and other data tools
Everyday users also play a central role in to perform mathematical manipulations. For its
the model’s quality assurance and risk proof of concept, the company used the gen AI
management. Up front, they are given clear system to translate user requests into database
guidance about the types of questions they can queries, then gave the calculation work to a
ask and how to formulate them. Additionally, to dedicated data analytics platform (Exhibit 2).
make it easier to check and validate outputs, The output from the analytics system is passed
the model is designed to generate a plain- back to the AI platform, which uses that data to
language explanation of the calculations produce several end products, including a text
underlying each set of results it produces. precis of the results, a summary of the calculation
methodology, and the computer code used to
4. Play to the strengths of gen AI create the graphical visualization.
While gen AI is a powerful solution for many
tasks, other technologies may be a better fit 5. Optimize the prompts
for some elements of the overall solution. By The team behind the AI assistant also took great
designing solutions that move functions that care to minimize the opportunity for error and
can be handled independently of gen AI to other hallucination by engineering the prompts used
technologies, organizations can enhance the to trigger the gen AI system. Shorter, simpler
stability and performance of the tools they build. prompts encourage large language models to
produce more predictable results. They are more
One limitation of the large language models efficient, too, since the computational workload
(LLMs) that underpin many modern gen AI of a gen AI system is closely related to the
systems is innumeracy. These systems were number of language “tokens” it handles in each
designed to operate on natural language, not interaction. For the finance assistant, the team
to perform precise calculations, and have divided the work into three distinct interactions,
been observed to struggle with mathematical each handled separately by the gen AI model:

Generative AI in finance: Finding the way to faster, deeper insights 3

 Exhibit 2
The gen AI platform provides the interface to financial data but does not
The gen underlying
perform AI platformcalculations.
provides the interface to financial data but does not
perform underlying calculations.

Illustrative example
User interface Gen AI Data management Gen AI User interface

Which five Gen AI engine Database Gen AI transforms

countries had transforms the returns that data into easily
the highest user’s question relevant interpreted information
indirect spend into instructions data • Summarize results
in 2022? for the database • Visualize numbers
• Explain logic

McKinsey & Company

translate the user’s question into a data query, maintain quality and manage risks at scale. These
generate a summary and explanation, and create might include ongoing monitoring and auditing
a visualization . mechanisms to assess AI system behavior,
ensuring it aligns with established ethical
The company developed examples of effective guidelines. Cross-functional teams of AI experts,
prompts for interacting with the system plus ethicists, and legal advisors should evaluate AI
guidance and training materials to help users models and applications for potential biases or
generate their own. Those training materials ethical concerns. After the success of its finance
were another chance to emphasize the assistant proof of concept, the consumer goods
importance of the human-in-the-loop, reminding company is now exploring appropriate governance
users of their responsibility to check the structures that will support the wider deployment
assumptions in the language of their queries and of gen AI tools.
behind the results of every interaction.

6. Build a robust test environment

Fine tuning a gen AI model is a highly iterative Generative AI is already changing the way
process. The ability to test and validate rapidly professionals do their work in the finance function
is key to successful development. To achieve it, and beyond. To capture the benefits of these
the team built a comprehensive automated test exciting new technologies while controlling the
suite, which operated using a list of high-priority risks, companies must invest in their software
queries provided by senior finance analysts. This development and data science capabilities. And
test was run after every update of the model they will need to build robust frameworks to
during development, with the results used for manage data quality and model engineering,
further refinement and operation. human–machine interaction, and ethics. However,
as these case examples show, these technologies
7. Put effective governance systems in place can accelerate and enable access to critical
As companies move from pilot projects to business information, giving human decision
mainstream adoption of gen AI tools, they will makers the information to make thoughtful and
need appropriate governance frameworks to timely choices.

Steven Eklund is a partner in McKinsey’s Stamford office, where Edward Woodcock is a senior expert; Lisa Kaufman is a
specialist in the London office; Avani Kaushik is an expert in the Southern California office; and Andrii Kurdiuk is a senior data
scientist in the London office, where Jan Svoboda is an associate partner.

Copyright © 2024 McKinsey & Company. All rights reserved.

4 Generative AI in finance: Finding the way to faster, deeper insights