Healthcare

GENERATIVE AI / THE RACE IS ON

FOUR ADOPTION

innovation
PATTERNS FORwith

YOUR
Generative
ENTERPRISE AI
The Future

is Now
Pundits have called generative AI technology a once-in-a-lifetime SoftServe’s adoption patterns
moment — a transformative technology to rival the commercialization aim to provide you — business
of the Internet. The hype is deafening, even confusing.

leaders — with just that: the voice

of clarity to navigate the fog of
Industries across the business spectrum have entered the race to emerging technology with a
understand technology and determine where it can provide the most chance to avoid pitfalls and pass
value. And while the generative AI journey is not a sprint, executives through the trough of
must approach generative AI with a heightened sense of urgency.

disillusionment, reaching the

plateau of Gen AI productivity
While many have spent countless words talking about what Generative faster and more successfully than
AI is and why it matters, few explain where to start and how to make your competition.
Generative AI real in the enterprise.

Generative AI Adoption Patterns

The multitude of options and specialized terminology in the Gen AI field can be perplexing.
To help you puzzle out and choose the right approach for your Gen AI endeavor, we describe
four adoption patterns here and depict them in the diagram below:
Pay
Integrate
Enrich with
Train on

and Use

Your Apps

Your Data

Your Data

Rapidly adopt third-party Seamlessly integrate Customize third-party Develop custom Gen AI model
application, but with third-party Gen AI Gen AI model with your trained on your data & integrate
limited customization model into your data and integrate it into into your apps for full IP ownership,
and control application(s) your apps control, and customization

3rd Party
3rd Party
3rd Party
3rd Party
In-house

Gen AI Model Gen AI Model Gen AI Model Gen AI Model Gen AI Model

3rd Party Corporate Corporate Corporate Corporate Corporate

Apps Apps Apps Data Apps Data

Users Users Users Users

Shorter time Higher value

Pay

and U se

ChatGPT by O k
penAI is undoubtedly the most well- nown 3rd party Gen AI service nowadays, having

J
reached 100 million monthly active users in anuary, just two months after its launch. Today, the field of

Gen AI SaaS applications is booming, o ff k

ering users a range from universal chatbots li e ChatGPT and

Claude.ai by Anthropic, to more specialized tools such as Git H ub Copilot for drafting code. The following

table shows key aspects of the Pay and Use Generative AI Tools approach. This table will help you compare

the SoftServe approach with other options and better understand if it is right for your business.

U se case e x amples The most suitable use cases

Code drafting for Software Engineering

Content drafting for M k

ar eting and Sales .

Somewhat suitable use cases

Extracting information from documents for various corporate roles .

The least suitable use cases

online customer support.

L imitations Accuracy, truthfulness, and inherent bias are common issues of Gen AI model

App specific limitations, e.g. ChatGPT cannot access data or current events after

September 2021

The primary limiting factor is that in this business pattern 3rd party Gen AI apps

are not integrated into the corporate ecosystem, which restricts their value to

individual usage. They provide limited or even transient competitive advantage,

as such tools are widely available.

Proprietary Data Proprietary data might be used in manual prompts or when uploading

documents to SaaS apps, such as Claude.ai.

Security/Privacy User input may be stored for the purposes of future model training. As for copyright

on the generated content, laws can vary depending on the jurisdiction and the specifics

of the use case, and policies may change over time. For example, under current U.S.

law, AI-generated content is not considered to be an original creation of the human

mind and is therefore not protected by copyright law.

Technology Foundational or specialized models available for end-users as

web, mobile apps, or plugins.

Cost
Typically fixed-fee subscriptions of $10 to $30 per user per month,

some products have usage-based pricing.

Implementation Team This adoption pattern does not require a technical team.

Our recommendation

Evaluate 3rd party Gen AI apps, adopt them for corporate roles where they display value, and monitor their

further development. The field is emerging, and for example, it is anticipated that GPT-5 will be 10 to 20 times

larger than the current GPT-4, which could enable new opportunities for your business.

In the meantime, let’s explore more value that lies in integration with your apps and corporate ecosystem.
Integrate

Your Apps
The good news for businesses is that popular LLM foundational models are available through APIs,
which can be integrated with existing applications or used for building custom software. In the table
below, we will explore key aspects of the “Integrate Your Apps” pattern:

Use case examples The most suitable use cases

Document processing and information extraction for various corporate
functions, such as procurement, risk, compliance, and legal;
Sentiment analysis;

Somewhat suitable use cases

Online customer support, virtual agents;

The least suitable use cases

Corporate knowledge base assistant.

Limitations See limitations in “Pay and Use”. In addition there are technical limitations related to APIs
The size of the context window (the data that the model can "see" or consider) is limited and
varies from service to service. For instance, as of now, GPT-4 supports up to 32K tokens. As a
rough estimate, you can think of 1 token as roughly equivalent to 1 word. This corresponds to
about 100 pages
While this pattern allows for integration with an application's data context, the primary limitation
is that it does not leverage corporate data stores such as wikis, document management, data
lakes, or data warehouses. The next two patterns aim to solve this limitation.

Proprietary Data Proprietary data from corporate apps are passed to a 3rd party
Generative AI service in the form of prompts.

Security/Privacy 3rd party terms can vary but most Cloud vendors do not transfer customer’s
data to other customers and do not use it for improving their models.

Technology Azure OpenAI Service (GPT-4, gpt-3.5-turbo), Google Cloud (PaLM 2), Amazon
Bedrock (Titan, Claude 2, etc.), Nvidia (NeMo). Most services provide REST API and
client libraries for Python, Java, Java Script and other popular languages.

Cost The monthly cost of usage 3rd party Gen AI service heavily depends on the amount of passed
and generated information (tokens), for example a virtual 24/7 agent answering around 10,000
questions per day would cost from $210 to $6,000 monthly (based on GPT-3.5 and GPT-4 Turbo
current pricing). At the same time, there are options for using open-source models that can
noticeably decrease the total cost of infrastructure for high-load applications.

The cost of integration with corporate apps will depend on the number of those apps and their
complexity; thus, it is not included in the calculation.

Implementation Team Solutions Architect, Software Engineers, DevOps Engineer

Our recommendation
Carefully assess your existing applications and identify where Gen AI integration can provide the most value.
Consider the technical limitations, such as the context window size and the cost based on the leveraging selected
model and API usage. Depending on the option, the cost difference can vary by up to 30 times. When the need
arises in the corporate knowledge base, then the next pattern is worth looking into.
Enrich with

Your Data
?
Looking to use existing LLM solutions with your own data No worries, it doesn't necessarily re uire training q
your own LLM. LLMs introduce a cost-effective approach to knowledge utilization called embeddings.
Through AI-powered indexing and search, GenAI systems can access information from corporate systems
such as portals, wikis, data warehouses, data lakes, and other sources of structured or unstructured data.
The integration of LLMs with proprietary data opens the door to leveraging the distinct knowledge amassed
q
by an organization — a resource uni uely their own.

Use case examples The best use cases for this pattern are the ones that rely on corporate data and knowledge,
e.g., Customer Service Automation, Marketing Content Generation, Auto-reporting, Enterprise
Search, Job Specific Assistants/Copilots, and Decision Support systems.

Limitations LLMs enable enhanced content indexing using AI-powered embeddings (indexes).
However, it also introduces the need for an additional data infrastructure for
processing and storing the embeddings
Fitting corporate knowledge into the model's context could be challenging.
However, adding an extra processing logic may help extract concise summaries
from the most relevant documents.

Proprietary Data Additional proprietary data from corporate data sources is passed to an
embedding model for content indexing. Self-hosted open-source embedding
models can be used to secure sensitive data.

Security/Privacy In addition to the embedding model, corporate data is also passed to a 3rd party LLM,
potentially exposing sensitive data to users via an LLM response. Proper guardrails and
response filtering must be implemented to prevent sensitive data exposure.

Technology Azure OpenAI, Amazon Bedrock, Google Cloud Embedding API, Cohere, and Nvidia NeMo are
the most popular 3rd party services for embedding models. HuggingFace Hub provides a wide
range of open-source embedding models. Milvus, FAISS, Chroma, and Weaviate are the most
popular open-source embedding databases, and Pinecone is a good 3rd-party alternative.

Cost “ Y ”
Like in the abovementioned pattern, Integrate our Apps, the monthly cost of usage
depends on the amount of passed and generated information. However, in this case,
indexing data through embeddings also incurs additional costs .

Although the cost of embeddings seems low ( . $0 0001

per 1K
tokens for GPT), and
1 0 j
indexing million 3 -page documents would cost ust about $1600 1
, indexing petabyte
of text may easily exceed $10
M when using a 3rd-party service .

Thus, using open-source models for embedding can drastically decrease the cost when
supporting massive data.

Implementation Team Solution and Data Architects, Software, Data, and ML Engineers, DevOps Engineer.

Our recommendation
Leverage corporate data as a powerful differentiator. Consider utilizing open-source models to mitigate
expenses, especially when dealing with large volumes of data. Implement robust guardrails and response
filtering to prevent potential exposure of sensitive data. In case your system relies on specific domain expertise,
the next approach — Train on your Data — can be a better option.
Train on

Your Data
The most value can be extracted from LLMs when it’s trained on the specific data relevant to the business
and subject domain. This pattern is more expensive to implement but provides more accurate responses
for specific domains or skills. The good news is that the initial model training can be eliminated by using
pre-trained Foundational Models (FMs) and fine-tuning them on the proprietary data. Some third-party
party LLMs allow fine-tuning the model on the customer's data, while others do not. In the latter case, the
model can be fine-tuned on the customer's data using self-hosted open-source FMs.

Use case examples The use cases for specialized skills or specific domain (Legal, Medical, Finance,
etc.) are the best candidates.

Limitations Training and fine-tuning LLMs requires a significant amount of training data and extensive
compute resources (including GPUs/TPUs)
Data quality is a critical factor for the success of this pattern – smaller but higher quality data
brings better results. Additional data curation and governance processes may be required
Many 3rd-party LLMs are limited in training data customization (e.g., GPT-3.5 and GPT-4 do
not support fine-tuning on the customer's data, only older versions)
LLMs trained on a specific domain may not be able to generalize to other domains.

Proprietary Data In addition to user and corporate search data, fine-tuning a 3rd party LLM requires passing a
significant amount of data relevant to the business and subject domain for knowledge extraction.

Security/Privacy 3rd party terms may vary but most Cloud vendors do not share customer’s
data and fined-tuned models between tenants.

Technology Google Vertex AI, Amazon SageMaker, Azure Machine Learning, HuggingFace Hub,
and Nvidia NeMo are the most popular services for training and fine-tuning LLMs.
HuggingFace Hub provides a wide range of open-source Foundational Models.

Cost The cost of fine-tuning varies heavily depending on the approach: managed service vs. open
source. For example, fine-tuning foundational models with biomedical data (~50B tokens)
using the OpenAI Davinci model costs $1,500,000, compared to $43,550 when using the
open-source MPT-30B. For comparison with fine-tuning, here are some examples of training
foundational models from scratch:

Scenario 1 (commodity model Scenario 2 (state-of-the-art model

Architecture: GPT3-styl Architecture: Llama-
Number of parameters: 70 Number of parameters: 70
Hardware: A cluster of 256 x Nvidia Hardware: GCP, a2-ultragpu-1g, NVIDIA
A100-40G A100 80G
Training dataset size: 1.4T token Training dataset size: 2T token
GPU Hours: 1,084,723 hour GPU Hours: 1,720,320 hour
Approx. Cost: $2,500,000 Approx. Cost: $8,719,958

Implementation Team Solution and Data Architects, Data Scientist, Software, Data, and ML Engineers, DevOps Engineer

Our recommendation
The “Train on Your Data”approach is more complex and expensive but offers the greatest benefits. Leverage pre-trained
foundational models and fine-tune them on proprietary data to reduce initial model training costs. Pay attention to data
quality and consider additional curation and governance processes, as smaller but higher quality data may yield better
results, and be mindful of the significant costs and resources required for training from scratch.
Get Started on
your Generative AI
Journey
Let SoftServe Accelerate Your
Generative AI Journey with 3 Offerings
Don't be left behind in the race to harness the disruptive innovation of Generative AI. SoftServe's adoption
patterns are designed to help you navigate the complexities of this emerging technology. We will help you
select an appropriate adoption pattern or its combinations for building tailored solutions to meet your
enterprise's unique needs. Expedite your journey towards a new way of productivity and creativity with
SoftServe Generative AI offerings, positioning your business as a leader in the industry:

AI Discovery AI launchpad AI adoption

Interest to Discovery: Launchpad to Innovation: Insight to Impact:

Generative AI Ecosystem and

Evidence-Based Exploration Rapid Scaling and Adoption
Implications for My Business and Deployment in My Organization

Use Cases & Business Generative AI Lab Generative AI Adoption

Impact Priorities Roadmap
AI Launchpad Program

Data Quality and for rapid experimentation Technology Strategy

Availability
Value Stream Mapping & Data Strategy
Technology Trade-offs Use Cases
and Architecture Change Management

POC/POV Pipeline and AI Governance

Technical Feasibility

with POC
Generative AI Solution
Development

Generative AI in Product
and Engineering Teams

It's urgent and the time to start is now. But you know that Generative AI is not a deployment sprint. With
SoftServe's experts to address your unique enterprise needs and guide you through possible adoption
options, you will move past the hype and harness its benefits for your enterprise.
Learn more about SoftServe’s Generative AI Lab, POVs, offerings, and partners.

Visit us at our Generative AI website page.

Your Partner
for the Future
Digital advisors, software innovators,
and experience builders
30 Years of

Experience 98% Client

Retention 14 Countries of
Operations 58 Offices
Worldwide

UK Sweden
Germany Poland
Canada Bulgaria Ukraine
Romania
USA
UAE
Mexico
Singapore
Colombia

Chile

Google Cloud Expertise

As a Google Cloud Partner and the winner of the 2020 Google Cloud Partner of the Year Award for
Machine Learning, SoftServe is committed to helping customers solve their most pressing business
challenges. SoftServe utilizes Google Cloud's AI ecosystem, including tools such as Google Cloud
Vertex AI and Generative AI App Builder. Learn more
120+ 690+ 100+ 3
GCP Enabled GCP Certified Data Science
Anthos

Customers Recources Experts Fellows

Ama zon W eb Services

As an APN Premier Services Partner, SoftServe acts as an exceptional cloud guide, vastly decreasing
the time to achieve cloud value. By doing so, SoftServe ensures that your AI initiatives unleash the full
potential of AWS Machine Learning services, such as Amazon Bedrock and SageMaker, and that they
are deployed in accordance with AWS Well-Architected best practices. Learn more
500+ 100+ 50+ 11
AWS Certified AWS Certified AWS Professional AWS

Resources Solution Architects Certifications Competencies

Microsoft
Innovate with purpose, rationalize costs, and drive efficiencies with Microsoft Azure's open and
flexible cloud computing platform. Leverage Azure OpenAI and Machine Learning services to deliver
next-generation AI solutions. As a Gold Microsoft Partner, SoftServe enables your business to build
and deploy on your terms — both today and in the future. Learn more
250+ 19+
Azure Certified
Years as a

Professionals Microsoft Partner

N vidia
As an NVIDIA Service Delivery Partner, SoftServe harnesses NVIDIA's cutting-edge technologies, like
GPU-accelerated compute infrastructure, to deliver robust AI solutions. Leveraging NVIDIA's NeMo
Service, SoftServe streamlines the development of Generative AI products, driving rapid digital
transformation. Learn more
deep technology expertise in AI/ML

human-first approach to designing

effective experiences

industry expertise and partnerships

with cloud solutions providers

Iurii Milovanov,

AVP AI/Data Science, [email protected]

Serge Haziyev,

Advanced Technologies CTO, [email protected]

Alex Chubay,

Chief Technology Officer, [email protected]