CHECKLIST 2024

Generative AI in Practice: Exploring Use

Cases to Harness Enterprise Data

By Fern Halper, Ph.D.

 Generative AI in Practice: Exploring Use
Cases to Harness Enterprise Data
By Fern Halper, Ph.D.

I
t is an exciting time for artificial intelligence.
Generative AI—a subset of artificial intelligence
Five considerations for generative AI:
that involves systems designed to generate
outputs such as images, music, text, or other forms of 1 Understand when generative AI is the
media based on its training data—is top of mind for right choice
many organizations. Generative AI promises to
revolutionize businesses by potentially enabling 2 Consider content creation use cases first
unprecedented levels of creativity, efficiency, and
personalization in content creation, product 3 Evaluate chatbots
development, and operational processes. In
November 2023, OpenAI announced that its 4 Use generative AI to derive insights
generative AI system, ChatGPT, had 100 million
5 Don’t forget about back-end deployment
weekly active users.1 It is not surprising, then, that
issues
in TDWI surveys the majority of respondents state
that they are either experimenting with or exploring
generative AI.

1
“OpenAI unveils personalized AI apps as it seeks to expand its ChatGPT consumer business,”
https://www.reuters.com/technology/openai-enables-customized-gpt-bots-offers-cheaper-more-powerful-models-2023-11-06/

1   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

At TDWI, we see organizations exploring generative

AI to unlock new levels of productivity using large
language models (LLMs) alongside related tools. In a
Understand when
recent TDWI survey, for instance, the top use cases for
generative AI included creating chatbots for customer
1 generative AI is the
support (39%), generating marketing content (29%),
onboarding new employees (26%), and acting as a
right choice
front end for analyzing company data (22%).2 All of
these use cases may involve using company data, There is often confusion around terms such as
including structured and unstructured data. AI, machine learning, and generative AI. AI is an
umbrella term that describes a broad field of computer
For instance, content creation or chatbots may utilize science focused on creating systems capable of
unstructured, text-heavy assets such as corporate performing tasks that typically require human
product information, troubleshooting manuals, intelligence. It consists of a number of technologies
website content, and other documentation that holds including machine learning, deep learning, natural
valuable company information; analytics assistants language processing, and generative AI. For instance,
may use structured and other kinds of tabular data machine learning methods originated in the field
typically stored in data warehouses, lakes, and other of computational science in the 1990s. In machine
analytics platforms. For organizations, it’s important learning, systems learn from data to identify patterns
to consider use cases that target all kinds of data with minimal human intervention, i.e., the computer
because with all enterprises having access to the learns from examples.
same models, company data is what provides their
competitive advantage. Because AI technologies have been used in practice
for decades, there is a wide range of use cases that
Of course, generative AI isn’t necessarily the best can be deployed without having to use generative AI.
option for all AI use cases. In this ever-evolving Although generative AI is an exciting technology, it
environment, it is important to have well-defined may not be the best AI technology for every use case.
business objectives and then evaluate the technology The choice between generative AI and other kinds
that will help the organization successfully achieve of AI, such as predictive AI, should be driven by the
them. As part of implementation, it’s important specific needs of the project or the decisions needed
to have a clear path to production, one focused on to be made.
maintaining data security and a low operational
infrastructure burden. For instance, predictive analytics techniques are
often used in retention analysis, cross-selling,
This TDWI Checklist Report explores several popular forecasting, and fraud detection. Predictive analytics
generative AI use cases as well as how to approach has also been successfully used for predictive
getting started with generative AI and putting it into maintenance. Predictive analytics uses statistics
production with company data. and machine learning techniques to determine the
probability of future outcomes using historical data,

2
Unpublished 2024 TDWI Data and Analytics Survey

2   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

often with known outcomes. For instance, in the case All this also presupposes that your organization has
of retention analysis, companies can use machine a data foundation in place to support these use cases.
learning to train a model using a historical data set of
customers who churned or didn’t churn and use it to
predict that behavior for future customers. To build
such models, structured and unstructured data have
Consider content
2
to be manipulated in specific formats the model can
understand—a process that requires expertise often creation use cases first
possessed by data scientists.

Generative AI, on the other hand, is useful for creating Content creation use cases are often good ways to get
new content or data that mimics the distribution of a easy wins and advance on the generative AI journey.
given data set. Foundation models for generative AI, Organizations may start off using popular generative
such as LLMs, are trained on a corpus that is, in many AI systems with prompt interfaces to help create
cases, the internet. Generative AI is useful for tasks an email or an article on a certain topic. They may
that require the generation of new ideas, designs, use a generative system to create illustrations for
or information based on learned data patterns. This presentations. However, soon they may want to move
may include content creation (text, images, videos, to a use case that utilizes their own company data
music), data augmentation (generating synthetic data (often company documents).
for training models), or creating realistic simulations.
Use cases include generating marketing emails or There are a number of content creation use cases
computer code. In the case of text, it’s useful not that we’re seeing at TDWI that make use of company
only for generating but also analyzing text to capture content data in the form of documents. These include:
sentiment, summarize content, or even translate it. • Document summarization is a growing use case
It is going to be important to determine which for generative AI, although it might initially seem
technique to use and when to use it. If your company more like a content reduction or simplification
has historical data about customer retention and task. In this context, generative AI is creating
wants to classify customers into certain categories, it new content in the form of a summary that
can use a predictive model rather than generative AI. distills information from longer documents into
If it is trying to understand why an event occurred, a concise form. This involves understanding
it might use the emerging technique of causal AI. If the content, extracting key points, and then
it is looking to create new content, it might employ generating a summary that accurately reflects
generative AI. the important points of the original text(s). This
process leverages generative AI’s capabilities in
Of course, different types of AI are not mutually natural language understanding and generation,
exclusive, and some use cases may require multiple making it a content use case. For instance, it can
kinds of models. For example, if a company wants to be used to summarize call center notes. Some
generate personalized emails to different customer LLMs can even provide sentiment analysis (see
segments, it may use a machine learning or clustering Number 4) so users can understand the customer
algorithm to generate the segments but then utilize a sentiment when reporting a problem or providing
generative AI model to create the marketing content. a review. This can help the company further

3   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

enhance voice-of-customer analytics by doing data such as wikis and internal FAQs. Chatbots are
such things as translating, summarizing, and software applications designed to simulate conversation
capturing sentiment of call center notes, trouble with human users, typically via text or voice interfaces.
tickets, or product reviews. They can be deployed on websites, messaging platforms,
mobile apps, and through virtual assistants to provide
• Document standardization means using AI
customer support, gather information, or facilitate
to extract the same set of fields across a large
transactions. Chatbots can change the paradigm of
corpus of documents so the enterprise can run
having users manually search for information and let
analytics or comparisons across documents. A
them just talk to the data instead.
good example is résumé processing. The system is
trained to understand multiple possible formats Two popular examples of chatbots include customer
and then recreate all the résumés in the user’s support chatbots and employee onboarding
chosen standard format. chatbots. To create a customer support chatbot,
an enterprise would use information from product
• Document generators create other kinds
documentation, user guides, and FAQs to fine-tune a
of documents. For instance, some document
pre-trained LLM so it can answer questions and cite
generators use company material and system code
relevant documents. The customer asks a question
to build technical documentation. Companies
with natural language and the chatbot returns the
are also making use of previous analysis about
answer. More complex chatbots might use customer
specific customers to generate product pitches or
transaction history (data structured in tables) to
marketing content.
deliver even more personalized content. For instance,
To make these use cases effective, enterprises must if an existing customer is asking a chatbot a question,
have a data foundation that can store documents the system might also pull customer data to know
along with metadata about the original and whether the customer is a loyal customer or has had
generated content. It will also require a private LLM. the product for a certain amount of time and use that
Because much of the information is confidential, information to provide a personalized experience.
the LLM will need to be governed. Although
Employee onboarding is another popular use case for
organizations may start by using pre-trained,
a chatbot that acts as an assistant. Here, a company
off-the-shelf LLMs, fine-tuning or customizing that
utilizes generative AI to answer questions about its
LLM may be needed over time to produce outputs in
products and services to help new employees. For
specific formats.
instance, a financial services company might build a
chatbot that is trained on information about its entire
portfolio. The new employee can ask questions and
get responses using this system to help bring them up

3 Evaluate chatbots to speed much faster.

Creating a chatbot requires a way to manage

Some extremely popular use cases for generative documents, load them into the system as text, and
AI involve chatbots and assistants that sit on top of provide information to the LLM. Depending on the
company documents and other large amounts of text chatbot, it might be necessary to turn the prompt as
well as the data into vector embeddings (numerical

4   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

representation of data) using an embedding model, as requirements for a chatbot would apply in terms of
well as a way to utilize the vectors with the LLM and managing unstructured data and building a
pass the results to the chatbot so it has relevant and production-ready application.
contextual information. This may require technology
Analyzing traditional company data with a
such as retrieval-augmented generation (RAG) to
generative-AI-enabled tool as a companion to (or
retrieve data about the customer. Some vendors provide
replacement for) traditional BI tools is increasing.
toolkits and templates for building chatbots that can
Many BI vendors are providing natural language
help organizations get started. Some even provide tools
interfaces to their platforms so users don’t
that offer a way to provide RAG without integrations,
necessarily need to know SQL to query a database.
infrastructure management, or data movement.
The idea is for the user to ask the question in a
natural way and get the response from the BI system.

Organizations should ask if their BI vendors are

incorporating natural language interfaces to their
Use generative AI to
4 derive insights products before trying to use generative AI (rather
than a BI tool) to analyze data. Using AI for such
analysis requires the generative AI tool to act as a
semantic layer and generate SQL to retrieve data,
Interestingly, a top use case being considered for
vectorize the data to find data with similar qualities,
generative AI is as a front end for analyzing company
and visualize this data in reports, dashboards,
data. The key objective is to democratize access to
graphs, and charts. These capabilities are now
insights by making structured data in a database
being incorporated into data platforms where the
available for anyone’s questions without having to
generative AI front end can generate SQL for the
write any code or use filters or sorting.
user. The user can interact with the system to tune
Organizations are considering generative AI as a the query and ensure they are happy with the result.
front end to both structured and unstructured data
Is this the future of BI? That remains to be seen. This
analysis. Here are two examples:
presupposes that the data in your organization is in
Analyzing conversations across websites, social a form to be analyzed, is of good quality, and all the
media channels, and applications can provide other requirements needed for BI and analytics to
organizations with insight into their customer work. Generative AI is not yet a silver bullet here.
interactions and requirements. This capability allows
businesses to interact with their text data in natural
language. A simple example of this is deriving
sentiments from data. An LLM can be trained on
several examples (called few-shot prompting when
the examples are provided within the prompt) of
customer statements with associated sentiments.
The LLM can then use this context to provide the
sentiment of other phrases/statements. If this
were to be done on a continual basis, the same

5   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

will also need to be concerned about “shadow AI”

implementations popping up that may use private
data on public systems.
Don’t forget about
5 back-end deployment • Security and privacy. Generative AI models must
also comply with security and privacy regulations.
issues This will include the ability to classify sensitive
information so it can be handled properly in
No matter how your company uses AI technologies, generative AI models. Additionally, only those
putting models into production comes with who are authorized to make use of certain data
concerns that organizations need to address, should be able to do so for generative AI. That
especially when dealing with private company data means that at a minimum, it is important to
and new model types. These include: implement robust role-based access control
(RBAC) and policy-driven security measures.
• Model governance. As with traditional AI In role-based access, each user has a role and
models, generative AI models and applications belongs to a group; this may be a department,
will need to be governed to make sure the models physical location, or user type. Traditional security
put into production are trusted and retain their mechanisms such as robust authentication,
integrity. This may include putting operations encryption, and masking will continue to be
(Ops) personnel in place to ensure that the important for data both at rest and in transit.
models are versioned and documented. The Ops
team may also track models and take corrective • Performance. Data pipelines will feed LLMs; in
action if the models drift or become stale. In some cases, data will have to be transformed as
the case of generative AI, the team should also the model is operating in production. Additionally,
ensure that the model doesn’t start to produce some models will need to be fine-tuned, which
invalid outputs (called hallucinations) and that can be computationally expensive and require
high-quality data is fed to the model. additional GPUs. Organizations will need to
look at techniques such as semantic caching to
It will be important to have a way to classify AI help maintain performance standards. Semantic
systems and applications according to risk and caching stores query results along with their
other parameters that may emerge with new semantic descriptions, enabling efficient handling
regulations. Some organizations have begun to of subsequent queries by leveraging previously
develop and advocate for AI model cards. Just as cached data. This technique optimizes data
the government requires nutritional fact labels retrieval by checking if new queries can be
on packaged goods in the U.S. for promoting answered using data already in the cache based
overall health, AI model cards provide relevant on its semantic context. It’s especially beneficial
documentation about an AI asset or application. in environments where data access is costly or
slow, enabling enhanced performance by reducing
Of course, given the relative ease of using certain unnecessary data access.
simple generative AI applications, organizations • Cost. Generative AI can also increase costs
for computational resources, data acquisition

6   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

and processing, and ongoing maintenance

and updates. As mentioned, an organization
wishing to train or fine-tune its foundation
model may require advanced GPUs, which can Concluding thoughts
lead to significant expenses. Data acquisition
involves both the financial costs of gathering and In the domain of artificial intelligence, generative
storing large data sets as well as the potential AI is capturing the imagination of organizations
costs associated with ensuring data quality worldwide. Organizations are now navigating
and addressing privacy issues. Additionally, a landscape where the integration of such AI
maintaining and updating AI models to keep technologies is becoming more important, from
them effective and relevant can incur ongoing enhancing customer interactions with chatbots
costs, including the need for continuous to generating insightful marketing content.
monitoring, retraining with new data, and Organizations are already starting to put these use
adapting to evolving compliance and ethical cases into practice.
standards. These factors combined make the
This journey into generative AI has challenges. For
total cost of ownership for generative AI projects
instance, the importance of a solid data foundation
something that organizations examine closely.
cannot be overstated. The successful deployment
of generative AI use cases—whether for content
creation, engagement through chatbots, or deriving
analytical insights—hinges on the ability of the
data infrastructure to handle the complexities
and demands of these applications. It will also
require a way to utilize private LLMs, support new
technology (such as vector databases), and support
data and model governance tools. The good news
is that vendors are starting to incorporate this
functionality into their platforms to make it easier for
organizations to get started.

7   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

About our sponsor About the author

Fern Halper, Ph.D., is vice president
and senior director of TDWI Research
for advanced analytics. She is well
Snowflake enables every organization to mobilize known in the analytics community,
their data with Snowflake’s Data Cloud. Customers having been published hundreds of
use the Data Cloud to unite siloed data, discover times on data mining and information
and securely share data, power data applications, technology over the past 20 years. Halper is also
and execute diverse AI/ML and analytic workloads. coauthor of several Dummies books on cloud
Wherever data or users live, Snowflake delivers a computing and big data. She focuses on advanced
single data experience that spans multiple clouds and analytics, including predictive analytics, machine
geographies. Thousands of customers across many
learning, AI, cognitive computing, and big data
industries, including 639 of the 2023 Forbes Global
analytics approaches. She has been a partner at
2000 (G2K) as of July 31, 2023, use Snowflake Data
industry analyst firm Hurwitz & Associates and a
Cloud to power their businesses.
lead data analyst for Bell Labs. She has taught at both
Colgate University and Bentley University. Her Ph.D.
Learn more at snowflake.com. is from Texas A&M University.

You can reach her by email (fhalper@tdwi.org), on X/

Twitter (x.com/fhalper), and on LinkedIn (linkedin.
com/in/fbhalper).

8   TDWI RESEARCH tdwi.org

 TDWI CHECKLIST REPORT: GENERATIVE AI IN PRACTICE: EXPLORING USE CASES TO HARNESS ENTERPRISE DATA

About TDWI Checklist Reports About TDWI Research

TDWI Checklist Reports provide an overview of TDWI Research provides industry-leading research
success factors for a specific project in business and advice for data and analytics professionals
intelligence, data warehousing, analytics, or a related worldwide. TDWI Research focuses on modern data
data management discipline. Companies may use this management, analytics, and data science approaches
overview to get organized before beginning a project and teams up with industry thought leaders and
or to identify goals and areas of improvement for practitioners to deliver both broad and deep
current projects. understanding of business and technical challenges
surrounding the deployment and use of data and
analytics. TDWI Research offers in-depth research
reports, commentary, assessments, inquiry services,
and topical conferences as well as strategic planning
services to user and vendor organizations.

A Division of 1105 Media

© 2024 by TDWI, a division of 1105 Media, Inc. All rights reserved. Reproductions in whole or in part
6300 Canoga Avenue, Suite 1150 are prohibited except by written permission. Email requests or feedback to info@tdwi.org.
Woodland Hills, CA 91367 Product and company names mentioned herein may be trademarks and/or registered trademarks
of their respective companies. Inclusion of a vendor, product, or service in TDWI research does not
E info@tdwi.org constitute an endorsement by TDWI or its management. Sponsorship of a publication should not be
construed as an endorsement of the sponsor organization or validation of its claims.
tdwi.org

9   TDWI RESEARCH tdwi.org