A PRACTICAL GUIDE
TO BUILDING GENERATIVE
AI APPLICATIONS
G INTAR E S AALI
FINNISH AI REGION PUBLICATION
Everything you need is
within your reach
Brought to you by:
Written by Gintare Saal.
Layout and graphic design
by Martti Asikainen
[Link]
Contents
Introduction to Generative AI 03
Overview 04
Importance and Benefits for Business 04
Use Cases of Generative AI 05
Components of a Generative AI Application 06
Choosing the Right Generative AI Tools and Frameworks 09
Large Language Models 10
Vector Databases 11
Embeddings 13
LLM Orchestration 14
Risks and Challenges of Generative AI and Their Mitigation Strategies 16
Unpredictable Output 17
Mitigation Strategies 17
Hallucinations 17
Mitigation Strategies 18
Data Privacy and Security 19
Mitigation Strategies 20
Bias and Fairness 20
Mitigation Strategies 21
Technical Complexity 21
Mitigation Strategies 21
Getting Started with Generative AI in Business 23
Additional Resources 24
A PRACTICAL GUIDE
TO BUILDING GENERATIVE
AI APPLICATIONS
INTRODUCTION TO
GENERATIVE AI
FINNISH AI REGION PUBLICATION
Overview
Generative AI refers to a subset of artificial intelligence that focuses on creating new content and data
rather than analysing or acting upon existing data. Unlike traditional AI, which primarily processes and
interprets information, generative AI can produce novel text, images, music, and even entire simulations
based on patterns and inputs it has learned during its training.
The core technologies enabling generative AI include advanced machine learning models, particularly
neural networks such as Generative Adversarial Networks (GANs) and Transformer-based models like GPT
(Generative Pre-trained Transformer). (Martineau 2023)
Generative AI models are trained on vast datasets and learn to understand the underlying structures and
patterns within the data. Once trained, these models can generate outputs that mimic the style, content,
and context of the training data, often with exceptional coherence and creativity.
Importance and Benefits for Business
Generative AI holds transformative potential for businesses by providing them with advanced tools
and capabilities. Here are some ways in which generative AI can benefit businesses:
• Enhanced customer engagement: Generative AI can power intelligent chatbots and virtual
assistants that provide personalized, instant customer service and support. These AI systems can
handle a wide range of customer inquiries, offer product recommendations, and resolve issues
efficiently, leading to improved customer satisfaction and loyalty. (Ihnatchyck 2023)
• Cost efficiency and scalability: By automating repetitive and time-consuming tasks such as
content creation, data entry, and report generation, generative AI allows businesses to reduce op-
erational costs and reallocate human resources to more strategic activities. This not only increases
efficiency but also enables businesses to scale their operations without a proportional increase in
costs. (Marr 2023)
• Creative content generation: Generative AI can assist in creating high-quality marketing ma-
terials, including text, images, and videos. AI-driven content generation tools can produce engaging
social media posts, articles, and promotional materials. (Ihnatchyck 2023)
• Data-driven decision making: Generative AI can analyse large datasets to generate insights
and predictive analytics, providing businesses with valuable information for decision-making. By
understanding customer behaviour, market trends, and operational performance, businesses can
make informed decisions that drive growth and competitiveness. (Microsoft 2024)
• Innovation and competitive advantage: Generative AI can assist in the ideation and design
of new products by generating innovative concepts and prototypes. This accelerates the develop-
ment process and helps your business bring new products to market faster, maintaining a competi-
tive edge over your competitors. (Ihnatchyck 2023)
4
Use Cases of Generative AI
Generative AI has the potential to transform various aspects of business by automating processes,
enhancing creativity, and improving customer engagement.
Here are some of the use cases of generative AI identified by Deloitte:
• Marketing content assistant
• Code assistant for developers
• Customer support on demand
• Product design assistant
• Research-based report generation
• Synthetic data generation
• Enterprise-wide data search and access
• Game content development
• Language translation
• Simulation of urban planning scenarios
• Hyper-personalised education
• Onboarding assistant
References
Deloitte. The Generative AI Dossier: A selection of high-impact use cases across six major industries.
Deloitte. London. [Link]
html#.
Ihnatchyck, Y. (2023). Top 4 generative AI benefits for business. Published on Data Science Central
on August 22, 2023. TechTargetr. Newton. [Link]
tive-ai-benefits-for-business/.
Marr, B. (2023). Boost Your Productivity with Generative AI. Published on Harward Business Review
on June 27, 2023. Harvard Business Publishing. Brighton. [Link]
ductivity-with-generative-ai.
Martineau, K. (2023). What is generative AI? Published on IBM Research website on April 20, 2023.
IBM. New York. [Link]
Microsoft 365 Team. (2024). Six Generative AI Benefits for Small Businesses. Published on Micro-
soft’s Business Insights and Ideas on April 29, 2024. Microsoft Corporation. Redmond. https://
[Link]/en-us/microsoft-365/business-insights-ideas/resources/generative-ai-bene-
fits-for-small-businesses.
5
COMPONENTS OF
A GENERATIVE AI
APPLICATION
FINNISH AI REGION PUBLICATION
Before diving into specific tools and frameworks, it’s important to understand
the different components that make up a generative AI application. These
components work together to create sophisticated AI systems capable of
generating content, answering questions, and more.
1. Large Language Models (LLMs): LLMs are deep learning models trained on vast amounts of text
data. They can generate coherent and contextually relevant text based on given prompts.
Examples: GPT-4, Gemini, Llama 3, Claude 3, Poro and Gemma.
2. Knowledge Bases: A knowledge base is a centralized repository that stores information, enabling
easy access and retrieval for various applications. It supports the integration and retrieval of data neces-
sary for generative AI systems to function effectively.
a. Vector databases (Pinecone, Weaviate): Store data as high-dimensional vectors, allowing efficient
similarity searches and semantic queries.
b. Relational databases (MySQL, PostgreSQL): Store structured data in tables with defined relation
ships, ideal for complex queries and transactions.
c. Document stores (MongoDB): Manage unstructured data in document formats (e.g., JSON), of
fering flexible schema and easy scalability.
d. Graph Databases (Neo4j): Represent data as nodes and edges, suitable for exploring relation
ships and connected data.
e. In-Memory Databases (Redis): Store data in memory for rapid access and real-time processing,
beneficial for high-speed transactions.
3. Embedding Models: These models convert data into numerical vectors that capture semantic mean-
ing, facilitating tasks like similarity search and clustering.
Examples: OpenAI embeddings (text-embedding-3-large), Cohere embeddings (embed-
multilingual-v3.0), Google embeddings (text-multilinguel-embedding-preview-0409).
4. Backend Logic: The backend logic integrates various components and handles the business logic, API
calls, data processing, and more. It ensures that the application functions smoothly and efficiently.
Examples: Flask, [Link].
5. Orchestration and Workflow Tools: These tools help in chaining together different AI models and
services to create end-to-end applications. They manage the flow of data and tasks between components.
Examples: Langchain, LlamaIndex.
7
6. User Interface: The front-end component that interacts with users, providing an accessible and us-
er-friendly way to interact with the AI application.
Examples: React, HTML/CSS, Javascript.
The diagram illustrates the flow and components of a generative AI application, detailing how different
elements interact to process user inputs and generate outputs (Figure 1).
Figure 1. The architecture of an LLM application showing the major components. (Choi 2023)
References
Choi, Nicole. (2023).The architecture of today’s LLM applications. Published on Github’s blog on
October 30, 2023. Github. San Francisco. [Link]
todays-llm-applications/.
8
CHOOSING THE RIGHT
GENERATIVE AI TOOLS
AND FRAMEWORKS
FINNISH AI REGION PUBLICATION
Generative AI applications rely on various tools and frameworks to function effectively. These tools
manage different aspects of the application, from data processing, orchestration, and response generation.
Selecting the right tools for your generative AI applications involves understanding the specific needs of
your project.
Large Language Models
When choosing a generative AI model for your business needs, it is essential to consider
the strengths, weaknesses, and specific capabilities of each model.
Here are some of considerations (Valenzuela 2024):
1. Use case specificity: Determine the primary task, such as text generation, coding assistance, or
comprehension. For instance, GPT-4 is highly versatile, suitable for a wide range of applications, while
specialized models like Code LLaMa are optimized for programming tasks.
2. Interfacing methods:
a. Playground interfaces are simple to use but limited in customization (e.g. ChatGPT, Google’s
Gemini)
b. API access allows for integration into applications without infrastructure management
but incurs costs (e.g. OpenAI’s GPT-4 API is easy to integrate)
c. Hosting locally offers full control and customization but requires significant technical effort
and resources (e.g running LLaMa models using Ollama).
3. Performance and cost:
a. Compare performance against cost, especially for extensive usage. GPT-3.5 models are
cheaper than GPT-4, making them a viable option for less demanding tasks.
b. Higher parameter models like GPT-4 require more computational power, impacting costs
and deployment strategies.
4. Customization and control:
a. If specific tuning is required, models like OpenAI’s fine-tuned GPT models or Meta’s LLaMa with
instruction tuning can be beneficial.
b. Hosting models like LLaMa locally provides customization but demands substantial infrastructure
management.
10
Hugging Face is a leading platform in the AI community, providing tools
and resources for developing, deploying, and sharing machine learning mod-
els. It is particularly known for its extensive repository of pre-trained mod-
els for natural language processing (NLP) and other AI tasks. Hugging Face
facilitates AI development and deployment, making advanced models more
accessible.
Spotlight on Poro Model: Poro 34B is an open-source, multilingual large
language model developed through a collaboration between Silo AI, the
TurkuNLP group at the University of Turku, and the High Performance
Language Technologies (HPLT) project. Trained on the LUMI supercomputer,
Poro 34B excels in English, Finnish, and code. (LumiOpen)
Vector Databases
A vector database is a specialized database optimized for storing and querying high-dimensional vector
data, commonly used in machine learning and AI applications. Vectors represent data in a numerical
format that captures semantic meaning, making them ideal for tasks like similarity search,
recommendation systems, and natural language processing.
Vector databases operate by transforming data into vector embeddings, which are high-dimensional rep-
resentations of the data. These embeddings are stored in the database and can be quickly queried using
various algorithms to find the nearest neighbours or most similar items. The process typically involves:
1. Raw data is converted into vector embeddings using machine learning models.
2. The embeddings are stored in the database, often with additional metadata for filtering and
enhanced search capabilities.
3. When a query is made, the database uses Approximate Nearest Neighbour (ANN) algorithms to
find and return the most relevant vectors quickly. (Huang 2023)
Selecting a vector database requires a thorough evaluation of several key factors to ensure it aligns with
your specific use case. Key considerations include performance, relevance, scalability, cost efficiency,
developer experience, enterprise readiness, and the capabilities for database management and hosting
(Huang 2023).
• Data freshness: Ensure the database can quickly update and query new data to maintain
relevance, crucial for real-time applications.
• Query latency: Look for low query response times to support interactive applications like
chatbots and search engines.
• Query per second (QPS): High QPS capabilities are essential for handling large volumes of
queries efficiently.
11
Relevance:
• Accuracy: Balance speed and accuracy by selecting a database with robust algorithms for
approximate nearest neighbour searches.
• Hybrid search: Consider databases that support both semantic and keyword search for more
versatile querying.
• Metadata filtering: Ability to filter queries based on metadata enhances search relevance and
specificity.
Scalability:
• Scaling capabilities: Assess whether the database can scale vertically (adding more resources to
existing servers) and horizontally (adding more servers) to handle growing data volumes.
• Automatic scaling: Prefer databases that support automatic scaling to simplify management.
Cost-effiency:
• Cost performance: Evaluate the cost against the performance benefits, especially as you scale
up the number of embeddings.
• Predictable pricing: Look for transparent pricing models to avoid unexpected costs.
Developer experience:
• Ease of use: Check for user-friendly APIs and SDKs to facilitate quick integration and
development.
• Documentation: Comprehensive and clear documentation helps developers get started and
troubleshoot issues effectively.
• Integrations: Ensure the database integrates well with your existing tech stack and preferred
cloud providers.
Enterprise readiness:
• Security and compliance: Verify that the database meets necessary security standards and
compliance requirements (e.g., SOC 2, GDPR).
• Support and expertise: Consider the availability of technical support and expertise from the
provider to assist with deployment and scaling.
• Monitoring and availability: Robust monitoring tools and high availability are critical for
maintaining uptime and performance.
Database management and hosting:
• Managed services: Some vector databases are fully managed by the provider, hosted either
on their platform or the client’s cloud, offering ease of use and reducing the need for in-house
management.
• Self-managed options: Others can be managed by the user, providing flexibility to host on cloud
or local infrastructure, which may require more technical expertise and resources.
12
Examples of vector databases (Ali 2023):
• Vector libraries (e.g., FAISS, HNSWLib, ANNOY): Suitable for static data, lacking CRUD (Create,
Read, Update, Delete) support.
• SQL databases with vector support (e.g., pgvector, Supabase): Convenient for integrating
vector data but less efficient for large-scale vector similarity searches.
• Full-text search databases (e.g., ElasticSearch, OpenSearch): Excellent for text search but
underperform in vector similarity searches.
• NoSQL databases with vector support (e.g., Redis, MongoDB): Basic vector capabilities, not
optimized for vector similarity search.
• Dedicated vector databases (e.g., Pinecone, Milvus, Weaviate, Qdrant): Designed for vector
operations, supporting high query workloads and advanced indexing algorithms.
For a continuously updated comparison of vector databases, refer to the comprehensive list provided by
Superlinked, available at [Link]
Embeddings
Embeddings are numerical representations (vectors) of data such as text, images, or audio. These
vectors capture the semantic meaning of the data, enabling efficient retrieval and comparison. In
natural language processing (NLP), embedding models transform text into vectors, allowing similar
texts to be placed close together in vector space.
Retrieval-Augmented Generation (RAG) relies heavily on embeddings to retrieve relevant information
from knowledge bases, enhancing the context and accuracy of generated outputs.
Evaluate the model using benchmarks like the MTEB Leaderboard on Hugging Face, which provides
performance statistics across various tasks.
• Focus on metrics like Retrieval Average (or Normalized Discounted Cumulative Gain (NDCG),
which measures retrieval accuracy.
• Larger models typically offer better performance but come with increased computational
requirements and latency. Balance the need for performance with available resources.
• Check the maximum number of tokens the model can process in a single embedding.
Generally, a limit of 512 tokens is sufficient for most applications.
• The length of the embedding vector affects both performance and storage. Shorter vectors
are faster and more storage-efficient, while longer vectors capture more data nuances. Aim
for a good trade-off based on your application’s needs. (Joshi 2024)
13
LLM Orchestration
LLM Orchestration involves managing and optimizing large language models (LLMs) to enhance their
performance and effectiveness. It includes tasks such as generating effective prompts, combining
outputs from multiple LLMs, efficiently managing resources, and monitoring performance.
LLM orchestration increases developer productivity by automating complex tasks, improves application
performance through optimized resource usage, reduces development costs, and enhances scalability
and reliability. This structured approach allows to focus on core functionalities, making applications
more robust and efficient. (Otten 2023)
LangChain and LlamaIndex are two prominent LLM orchestration frameworks. Each offers unique
features and benefits suited to different use cases (Padhy 2024).
1. LangChain is designed to simplify the creation and management of complex workflows
involving multiple AI models and services. It focuses on integrating LLMs with various data sources and
orchestrating their interactions. Key features are:
• LangChain provides a high-level API that allows to define intricate workflows and chains of AI
models and services.
• It integrates well with existing tools and services, making it adaptable to a wide range of
applications.
• The framework excels at managing prompts, ensuring that they are effectively used to elicit
the desired responses from LLMs.
14
LangChain is ideal for building sophisticated applications such as advanced chatbots, virtual assistants,
and AI-driven content creation tools that require integrating multiple AI models and external data
sources, such as social media platforms, online communities and websites and more.
2. LlamaIndex is focused on bridging the gap between powerful LLMs and private or domain-specific
data. It provides a structured way to ingest, organize, and query various data sources using LLMs.
Key features are:
• LlamaIndex allows for seamless integration with diverse data sources such as APIs, databases,
and document files.
• Facilitates natural language interactions with private data without the need to retrain models.
• Offers both high-level and low-level APIs to cater to different expertise levels among
developers.
LlamaIndex is perfect for applications that require integrating LLMs with specific datasets, such as
internal knowledge bases, proprietary documents, and specialized databases. It excels in scenarios
where natural language querying of private data is essential.
When choosing between LangChain and LlamaIndex, it’s crucial to assess the specific and unique
requirements of your project:
• If the application requires sophisticated orchestration involving multiple AI models and servic-
es, LangChain’s comprehensive tools and flexibility make it the better choice.
• For projects focused on integrating LLMs with private or domain-specific data sources,
LlamaIndex offers a more straightforward and efficient solution.
Both frameworks offer powerful capabilities, but their suitability depends on the complexity and specific
requirements of the generative AI application.
References
Ali, M. (2023). Optimizing RAG: A Guide to Choosing the Right Vector Database. Published on Medium on
December 3, 2023. Medium. San Francisco. [Link]
to-choosing-the-right-vector-database-480f71a33139
Joshi, A. (2024). How to Choose the Right Embedding Model for Your LLM Application. Published on
MongoDB on June 18, 2024. MongoDB Inc. Github. [Link]
las/choose-embedding-model-rag/
LumiOpen. Poro 34B Model Card. Published on HuggingFace. [Link]
Huang, X. (2023). An (Opinionated) Checklist to Choose a Vector Database. Published on Pinecone on
September 13, 2023. [Link]
base/
Padhy, R. P. (2024). Exploring Popular LLM Orchestration Frameworks. Published as LinkedIn article on
April 15, 2024. [Link]
bi-prasad-hjxkc/
Valenzuela, A. (2024). LLM Classification: How to Select the Best LLM for Your Application. Published on
Datacamp on March 2024. [Link]
best-llm-for-your-application.
Van Otten, N. (2023). How to manage Large Language Models (LLM) — Orchestration Made Sim-
ple [5 Frameworks]. Published on Spot Intelligence on November 17, 2023. [Link]
com/2023/11/17/llm-orchestration-frameworks/
15
RISK AND CHALLENGES
OF GENERATIVE AI AND THEIR
MITIGATION STRATEGIES
FINNISH AI REGION PUBLICATION
While generative AI offers numerous benefits, it also comes with its own set of risks and challenges that
businesses need to consider. Understanding these challenges is crucial for effectively implementing and
leveraging generative AI technologies.
Unpredictable Output
Unpredictable output refers to the inherent
variability in the quality and appropriateness of
content generated by generative AI models. These
models can produce a wide range of outputs, some
of which may not align with brand guidelines,
cultural norms, or the intended use case.
(Vartak 2023)
Mitigation Strategies
1. Human-in-the-Loop Review: Implementing a robust human review process is essential to assess the
quality and appropriateness of AI-generated content. Human reviewers can provide the necessary context
and judgment that AI models lack.
2. Guideline Enforcement: Establishing and enforcing strict brand and content guidelines can help
mitigate risks. Clear guidelines on tone, style, and cultural sensitivity can aid human reviewers and AI
systems in producing consistent and appropriate content.
3. Feedback Loops: Implementing feedback mechanisms where users and stakeholders can report
inappropriate or low-quality outputs allows for continuous improvement of the AI models. This feedback
can be used to improve future outputs.
Hallucinations
AI hallucinations refer to instances where generative AI models produce information that appears plausible
but is incorrect or nonsensical. These models generate content based on patterns in the data they were
trained on, but when they encounter gaps in their knowledge or ambiguous prompts, they may fabricate
information to fill in those gaps. (Tonmoy et al. 2024)
Hallucinations can lead to the dissemination of false or misleading information, which can significantly
harm a company’s credibility and decision-making processes. Hallucinations pose a particular risk in
scenarios where accuracy and reliability are critical. They can undermine the perceived reliability of AI
systems, necessitating rigorous human oversight to verify the correctness of AI-generated outputs.
Without proper mitigation, the risk of hallucinations can outweigh the benefits of using generative AI.
(IBM)
17
Mitigation Strategies
1. Retrieval-Augmented Generation (RAG): Integrating information retrieval with text generation can
help reduce hallucinations. By retrieving relevant documents or data from a verified corpus before
generating responses, RAG ensures that AI outputs are based on accurate and relevant information,
enhancing their reliability. (15)
2. Human-in-the-Loop Review: Implementing a system where human reviewers verify AI-generated
content can significantly reduce the risk of hallucinations. Humans can cross-check the information
provided by the AI against reliable sources, ensuring accuracy before the content is published or used.
(IBM)
3. Training on High-Quality Data: Ensuring that AI models are trained on high-quality, accurate, and
diverse datasets can help reduce the frequency of hallucinations. High-quality training data enables the
models to generate more reliable and contextually appropriate content. (IBM)
4. Prompt Tuning: Prompt tuning involves customizing prompts to guide large language models towards
producing desired outputs. This approach helps avoid extensive retraining by providing carefully designed
prompts that lead to more accurate and contextually appropriate responses. (Tonmoy et al. 2024)
Here are several prompting tips:
• Ask the LLM to explain how it arrived to the conclusions or to provide evidence to support its claims;
• Provide the LLM with output examples or a response structure;
• Break down a complex task to smaller steps;
• Be more specific, do not leave room for ambiguity.
(Arsanjani 2023)
18
Data Privacy and Security
Generative AI systems often require large amounts of data to function effectively. Handling and processing
this data raises significant privacy and security concerns, particularly when dealing with sensitive or person-
al information. Ensuring that data is protected from unauthorized access and breaches is vital. (Marr 2024)
Inadequate data protection measures can lead to severe consequences such as data breaches, legal
repercussions, and loss of customer trust. For businesses, a data breach can be particularly damaging,
leading to financial losses, regulatory fines, and a tarnished reputation.
Mitigation Strategies
1. Compliance with Regulations: Complying with data protection regulations such as the General Data
Protection Regulation (GDPR) is essential. These regulations provide stringent requirements and guidelines
for maintaining data privacy. (Baig 2023)
The new EU AI Act categorizes AI applications by risk levels, imposing strict requirements on high-risk AI
systems to ensure transparency, data privacy, and security. Compliance with this Act will be crucial for busi-
nesses operating within or engaging with the EU market. (European union)
2. Access Controls: Implementing robust access control mechanisms limits access to generative AI mod-
els and the data they generate to only authorised personnel. (Baig 2023)
3. Data Anonymization: Removing personally identifiable information before feeding it to generative
models protects individual privacy. (Baig 2023)
4. Secure Data Storage: Utilizing secure cloud storage solutions with robust security measures such as
encryption, regular security updates, and secure access protocols safeguards data. (Baig 2023)
5. User Consent: Informing users of the intended use of their data and requesting their consent, this way
ensuring a transparent system and an informed decision-making process by the users. (Baig 2023)
6. Data Governance Policies: Implementing a comprehensive data governance framework establish-
es clear policies and procedures for data management, including data handling and protection standards.
(Baig 2023)
19
Bias and Fairness
Bias in generative AI occurs when models produce outputs that reflect prejudiced or unfair perspectives,
often derived from the biased data on which they were trained. Fairness refers to the need to ensure that
AI systems treat all users and scenarios equitably, without discrimination or favouritism. (Ferrara 2023)
Bias in AI can lead to unfair treatment of individuals or groups, perpetuating inequality and potentially
causing harm. For instance, biased AI outputs can affect hiring decisions, customer service interactions, and
product recommendations, leading to legal and reputational risks for businesses. Ensuring fairness is critical
to maintaining trust and promoting ethical AI use. (Guvvala 2023)
Table 1. Bias in AI systems. (Ferrara 2023)
20
Mitigation Strategies
1. Diverse and Representative Training Data: Ensuring that training datasets are diverse and
representative of all relevant user groups and scenarios. This helps the AI model learn a more
balanced perspective. (Guvvala 2023)
2. Human-in-the-Loop (HITL) Approaches: Establishing robust feedback loops where users can report
biased or unfair outputs. This feedback can be used to continuously improve the AI model. (Guvvala 2023)
3. Transparent and Explainable AI: Ensuring transparency in how AI models are trained and make de-
cisions. Clear documentation and communication about the model’s workings can help users understand
and trust the AI system. Additionally, developing explainable AI models that can provide insights into how
decisions are made. This helps in identifying and addressing potential biases. (Ferrara 2023)
Technical Complexity
Technical complexity in generative AI refers to
the challenges involved in implementing,
maintaining, and optimizing AI systems. High
technical complexity can result in increased
costs, longer implementation times, and
potential failures in deploying AI solutions
effectively. Business may struggle with limited
technical expertise and resources, hindering
their ability to leverage AI fully. (Vartak 2023)
Mitigation Strategies
1. User-Friendly Tools and Platforms: Leveraging AI tools and platforms that offer user-friendly
interfaces and require minimal technical expertise. These tools often provide pre-built models and easy-
to-use frameworks that simplify AI implementation. Additionally, utilizing managed AI services from cloud
providers that handle the underlying infrastructure and maintenance can significantly reduce technical
complexities associated with AI applications.
2. Partnerships and Collaboration: Collaborating with AI experts, consultants, or technology
partners who can provide the necessary technical support and guidance as well as engaging with
academic institution to stay updated on the latest AI advancements and best practices. This can provide
access to cutting-edge research and technical expertise.
3. Incremental Implementation: Implementing AI solutions in phases, starting with proof-of-concept
projects and gradually scaling up. This allows businesses to learn and adapt to the complexities of AI with-
out overwhelming their resources.
4. Training and Skill Development: Investing in training programs to upskill employees on AI technol-
ogies and best practices. Encouraging participation in AI workshops, courses, and certification programs to
21
enhance technical knowledge and competencies. This helps build in-house expertise and reduces reliance
on external resources.
5. Scalable Infrastructure: Leveraging cloud computing resources to scale AI infrastructure as needed.
Cloud platforms offer scalable and flexible environments that can handle the computational demands of AI.
References
Arsanjani, A. (2023). Navigating the Challenges of Hallucinations in LLM Applications: Strategies
and Techniques for Enhanced Accuracy. Published on Medium on November 20, 2023. Medium.
[Link]
tions-strategies-and-techniques-for-ab2b5ddc4a63.
Baig, A. (2023). Navigating Generative AI Privacy : Challenges & Safeguarding Tips. Published
on [Link] on September 21, 2023. [Link]
European union. (2024). Up-to-date developments and analyses of the EU AI Act. European
union. Brussels. [Link]
Ferrara, E. (2023). Fairness and bias in artificial intelligence: A brief survey of sources, impacts,
and mitigation strategies. University of Southern California. Los Angeles. [Link]
pdf/2304.07683.
Guvvala, S. (2023). Bias Mitigation in Generative AI. Published on Analytics Vidhya on No-
vember 29, 2023. Gurgaon. [Link]
tion-in-generative-ai/
IBM. What are hallucinations?. IBM. Armonk. [Link]
Marr, B. (2024). Generative AI And Data Protection: What Are The Biggest Risks For Employers?.
Published on Forbes on May 27, 2024. Forbes. Jersey City. [Link]
nardmarr/2024/05/27/generative-ai-and-data-protection-what-are-the-biggest-risks-for-em-
ployers/
Tonmoy, S. M. T. I., Zaman, M., Jain, V., Rani, A., Rawte, V., Chadha, A. & Das, A. (2024). A Com-
prehensive Survey of Hallucination Mitigation Techniques in Large Language Models. Cornell
University. Ithaca. [Link]
Vartak, M. (2023). Six Risks Of Generative AI. Published on Forbes on June 29, 2023. Forbes.
Jersey City. [Link]
tive-ai/.
22
GETTING STARTED WITH
GENERATIVE AI
IN BUSINESS
FINNISH AI REGION PUBLICATION
Implementing generative AI involves several steps to ensure
success and value realization (Falck 2024):
1. Education and motivation: A thorough understanding of generative AI is essential before imple-
mentation. This involves reading books, articles, and taking courses to stay updated with the rapidly
evolving technology.
2. Focusing on real problems: Generative AI should be applied to solve specific, minor problems initially.
Starting small allows for internal testing and skill development before scaling to more significant, custom-
er-facing projects.
3. Selecting tools and partners: Choosing the right AI tools and platforms is critical. Initially, partner-
ing with experts can help in building a realistic AI roadmap and navigating through various available tools.
4. Creating an AI-ready culture: An AI-ready culture is characterized by employees viewing AI as an
opportunity, willingness to integrate AI across functions, and focusing on augmenting human abilities
rather than replacing them. Transparency, training, and incentivizing AI use cases are vital for this culture.
5. Learning prompt engineering: Prompt engineering is crucial for utilizing generative AI tools effec-
tively. It involves crafting precise prompts to guide AI models in generating the desired outputs. Under-
standing and improving prompts can significantly enhance the results.
References
Falck, J. (2024). 5 steps to getting your organization started with generative AI. Published on
Qvik on March 14, 2024. Qvik. Helsinki. [Link]
zation-started-with-generative-ai/
Additional Sources
1. McKinsey: The economic potential of generative AI: The next productivity frontier: [Link]
ties/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier#introduction
2. Github: A continuously updated list of Large Language Models that are licensed for commercial use: [Link]
eugeneyan/open-llms
3. [Link]: A continuously updated detailed list of LLMs with information about parameters, training datasets and
other notes: [Link]
4. Superlinked: A continuously updated comparison of vector databases: [Link]
son
5. Nvidia: What is Trustworthy AI?: [Link]
is%20an%20approach,people%20who%20interact%20with%20it.
6. McKinsey: Why businesses need explainable AI—and how to deliver it: [Link]
black/our-insights/why-businesses-need-explainable-ai-and-how-to-deliver-it
24