The Business of AI: Current Applications and Data Fuel AI Innovation and Solutions

How businesses can stay grounded in strategy while leveraging existing assets to train and deploy AI solutions

Brian Silverman February 3, 2025

This is the fourth article in the series, “The Business of AI.” In case you missed them, here are the first three articles:

1. Focus on ‘The Business of AI’ to Move From Hype to ROI
2. A Look at the Different Types of AI and Their Value
3. From Foundation Models to Large Language Models

Recent news about DeepSeek and the advances it represents in training AI models, as well as lowering the compute requirements to support it, are remarkable. To learn more about the different training techniques for large language models (LLMs) and foundation models, please review the addendum at the end of this article.

These advancements in AI are accelerating every day. Last week, OpenAI announced its own agent, Operator, and at the end of 2024, it announced its soon-to-be-released o3 model, which promises improvements in reasoning and capabilities for its LLMs and foundation models. At CES in January, Nvidia announced several innovations, including a supercomputer that can run AI and fit in the palm of your hand.

This makes our focus in these articles on “The Business of AI” even more important. Organizations should keep their AI strategy grounded by focusing on the real value that AI—across different types, models and technologies—can bring to their operations.

The Critical Role of Current Applications and Data for AI Success

The heart of any organization’s strategy will be how to leverage its current applications and data, as these are the lifeblood of business operations. According to recent studies, the average enterprise uses over 100 applications and manages hundreds of databases, which hold critical business information and power essential workflows.

For AI solutions to succeed, these existing assets must be leveraged to train and fine-tune AI models, to develop and deploy AI applications and solutions that support AI business use cases and return value of an organization’s time and investment.

Most organizations depend on applications ranging from enterprise resource planning (ERP) systems, for managing finances and supply chains, to customer relationship management (CRM) tools, for tracking customer interactions. Also critical are databases storing transactional records, product inventories and historical data.

Additionally, productivity apps, such as collaboration tools and task managers, play a role in generating unstructured data—including emails, meeting notes and shared documents—that can be vital for generative AI (GenAI) use cases.

These applications and databases not only support the organization’s operations but also embody business rules, logic, processes and data that will be needed to train AI models.

The Impact of Different AI Types in Applications and Data Integration

Traditional AI relies on structured datasets and predefined logic, making it well suited for tasks such as fraud detection or predictive maintenance. These applications are typically integrated directly into existing systems, automating repetitive or rule-based processes with high accuracy and reliability.

In contrast, GenAI requires a mix of structured and unstructured data—like manuals, emails and customer feedback—to fine-tune models and generate creative, adaptive solutions. For example, GenAI can power customer service chatbots or content generation tools, which interact dynamically with both users and back-end systems.

Building upon Generative AI, Agentic AI represents the next phase in AI evolution, enabling autonomous decision-making and task completion with minimal human intervention. For Agentic AI to succeed, high-quality data and seamless access to applications are essential. These applications must support real-time interactions and provide actionable insights while ensuring that the AI agents can securely access and utilize business-critical systems and workflows.

To underscore the importance of high-quality data for training AI models, organizations should consider a dedicated strategy for preparing and managing this data. The addendum below explains different types of AI training, such as supervised, unsupervised and reinforcement learning, emphasizing the unique data requirements and their role in successful AI implementation.

Example: AI-Powered Customer Support Chatbot

Consider a business prioritizing a GenAI-powered customer support chatbot to offload most customer support requests, allowing their human counterparts to focus on complex challenges and high-profile customers. For this solution to succeed, the organization needs the right information to train its own large language model (LLM) or fine-tune a foundation model with specifics for the organization. They will require quality data, including years of customer support case history, as well as current and past manuals, technical reports and other unstructured data, all of which ensure the chatbot is accurate and effective in addressing customer needs.

In addition, the chatbot must access key services and applications to deliver effective support. This includes verifying entitlement for solutions or premium support, accessing inventory of spare parts, scheduling service calls, or shipping replacements. The chatbot will also need to log interactions and resolutions in a CRM or support application.

The foundation of a successful AI strategy lies in how well organizations leverage and integrate their existing applications and data into AI solutions. These elements are not merely supporting factors; they are critical enablers of innovation, ensuring AI systems can deliver meaningful business value while seamlessly integrating into operational processes.

Strategic Steps for Incorporating Applications and Data Into Your AI Strategy

1. Prioritize AI Applications and Use Cases

The first step is to understand the prioritized applications and use cases. This will keep the focus on key applications, and the data that is required for their success.

Focus on AI use cases that align with business goals and processes.
Understand whether the use case relies on Traditional AI (e.g., predictive models) or GenAI (e.g., LLMs), as this will impact the applications and data required.

2. Build the Right Team

The right team needs to include AI, IT and business professionals that understand the business processes the AI solution will support, and the different applications and data that need to be leveraged to train the models as well as support the AI solution being developed.

Include AI specialists, data engineers, IT staff and business stakeholders.
Ensure the team understands the applications, data and processes the AI solution will support.

3. Inventory and Assess Applications for AI Success

Applications generate the data needed to train AI models. They also provide the essential services and capabilities that enable AI solutions to succeed in delivering meaningful outcomes. Organizations should consider the following steps when incorporating applications and data into their AI strategy:

Map and validate application capabilities to the business processes supported by the AI solution.
Ensure applications provide necessary access, such as APIs, for on-premises, cloud or SaaS-based solutions.
For legacy applications lacking API support, consider middleware solutions for modern access and integration.
Evaluate whether to replace applications to ensure long-term viability and support for future AI requirements.

4. Leverage Data for AI Success

Quality data is the fuel that will enable success with an organization’s AI strategy and solutions. From the training of AI models to accessing data required by the business application, the importance of data cannot be exaggerated for AI success.

Inventory and Assess Databases
Identify all databases and silos containing business-critical information required for AI solutions.
Evaluate data quality, as training AI models demands high-quality, clean data.
Include unstructured data, such as manuals and reports, particularly for GenAI models.

5. Modernization Requirements for Applications and Data

When inventorying and planning access to applications and data, organizations should assess their fit for supporting evolving needs, including future AI solutions, to ensure alignment with long-term business objectives.

Application Modernization:

Refactor or replace applications to support current and future AI workloads. For example, refactoring an application might involve updating its codebase to ensure compatibility with modern AI APIs, or migrating it to a cloud-based platform to enable scalability. In scenarios where legacy applications cannot be adapted, replacement may be necessary, such as adopting an SaaS solution with built-in AI integration capabilities.
Migrate application infrastructure to cloud-based architectures for scalability and flexibility.

Data

Implement ETL (Extract, Transform, Load) pipelines or data virtualization solutions to centralize legacy system data.
Use middleware or APIs to integrate data from legacy applications.

By taking these steps, organizations can ensure their existing applications and data fuel AI innovation effectively, driving transformative outcomes such as competitive differentiation, measurable ROI and seamless integration into core business processes.

Training Techniques for LLMs and Foundation Models

Whether using publicly available AI models or open source, or developing your own AI models for your specific needs, these models will require training.

For the training to be successful and ensure the highest-quality AI solution, it requires the right data, and quality is key. Bad data can train the models to produce flawed responses.

The table below describes the different types of training leveraged by different AI types, the role of data and why it matters for training of AI models. Due to DeepSeek’s recent impact on the AI industry, that model is highlighted in this table.

Beyond initial training of models, organizations need to consider the model and how to maintain and update the data. This includes leveraging fine-tuning and retrieval augmented generation (RAG) technologies and techniques.