The Ultimate Guide to Training Generative AI: Unleash Your Company’s Data and Transform the Future

Training Generative AI Using Your Company’s Data: Unleashing the Power of Knowledge

Training Generative AI Using Your Company's Data

Numerous companies have been experimenting with large language or image models like ChatGPT, discovering their remarkable ability to articulate complex ideas. However, these models are primarily trained on internet-based information, making it challenging for them to respond to proprietary content or knowledge-related queries.

In today’s volatile business environment, leveraging a company’s proprietary knowledge is vital for competition and innovation. Effective knowledge management plays a crucial role in organizational innovation, involving the creation, management, application, recombination, and deployment of knowledge assets and know-how. However, comprehensive knowledge within organizations is often scattered across various sources and forms, making it difficult to organize and deploy efficiently.

Emerging technologies, such as large language and image generative AI models, offer new opportunities for knowledge management, enhancing company performance, learning, and innovation capabilities. For instance, a Fortune 500 provider of business process software conducted a study that showed a generative AI-based system for customer support significantly increased productivity, retention, and positive feedback from customers. Moreover, the system accelerated the learning and skill development of novice agents.

Inspired by success stories like these, an increasing number of organizations are leveraging the language processing and reasoning abilities of large language models (LLMs) to provide broad internal or customer access to their intellectual capital. They use LLMs to inform customer-facing employees about company policies, offer product/service recommendations, solve customer service issues, or capture departing employees’ knowledge.

Although the objective of capturing and disseminating knowledge has existed since the knowledge management movement in the 1990s and early 2000s, previous technologies were inadequate. However, generative AI is reviving the possibility of effectively capturing and disseminating important knowledge throughout organizations and beyond their boundaries. One manager described the experience of using generative AI for this purpose as feeling like “a jetpack just came into my life.” Despite recent advances, challenges similar to those faced in the past still remain.

Technologies for Generative AI-Based Knowledge Management

The technology to incorporate an organization’s domain-specific knowledge into an LLM is rapidly evolving. Currently, there are three primary approaches to achieving this:

  1. Training an LLM from Scratch: Creating and training a domain-specific model from scratch is a less common approach due to the massive amount of high-quality data and considerable computing power and expertise required. Bloomberg is an example of a company that took this approach, using over 40 years’ worth of financial data and a large volume of text to create their finance-specific language model. However, most companies lack the necessary resources for this approach.
  2. Fine-Tuning an Existing LLM: Fine-tuning involves training an existing LLM with specific domain content. It requires less data and computing power compared to training from scratch. Google, for instance, used fine-tuning to enhance their medical knowledge model. By retraining their general language model on curated medical knowledge, they achieved improved performance in answering medical licensing exam questions. Although cost-effective compared to training from scratch, fine-tuning still demands data science expertise and has certain constraints, such as compatibility limitations with some LLM vendors.
  3. Prompt-Tuning an Existing LLM: The most common approach for customizing an LLM’s content is prompt-tuning. This method involves modifying the model’s responses using user prompts containing domain-specific knowledge. It is computationally efficient and requires a relatively small amount of data to train the model in a new content domain. Morgan Stanley, for example, used prompt-tuning to train OpenAI’s GPT-4 model with carefully curated financial and investment knowledge. The prompt-trained system operates within a private cloud accessible only to Morgan Stanley employees.

Content Curation and Governance

Similar to traditional knowledge management, generative AI-based knowledge management requires high-quality content. In some cases, curated databases of domain-specific knowledge are readily available. Otherwise, companies need human curation to ensure accurate and timely content that avoids duplication. Morgan Stanley, for instance, has a team of knowledge managers who constantly score documents based on suitability for incorporation into their generative AI system. Companies without well-curated content may face challenges in this regard.

Maintaining high-quality knowledge becomes easier when content creators are equipped with effective document creation skills. Companies like Morgan Stanley and Morningstar provide training to their content creators on writing and tagging content, facilitating content governance processes.

Quality Assurance and Evaluation

Ensuring the quality of generative AI content is crucial, as these models occasionally produce inaccurate information. However, companies that fine-tune LLMs with domain-specific knowledge have observed fewer errors compared to out-of-the-box models. Evaluation strategies play a vital role in managing generative AI content. Companies like Bloomberg and Google evaluate their models on specific criteria, such as factuality, precision, reasoning, and safety, involving human experts and comprehensive assessments. Similarly, Morgan Stanley employs a set of “golden questions” to test the accuracy of their system whenever changes are made.

Legal and Governance Issues

Deploying LLMs involves complex legal and governance issues, including intellectual property, data privacy and security, bias and ethics, and false/inaccurate outputs. Although the legal status of LLM outputs is still unclear, incorporating legal representatives in the creation and governance processes is advisable. Confidentiality concerns can be addressed by fine-tuning LLMs on private instances inaccessible to the public. Advanced safety and security features are also being developed by vendors to prevent unauthorized access and protect propriety data.

Shaping User Behavior

To effectively incorporate generative AI capabilities into tasks, companies should develop a culture of transparency and accountability. Users need to understand how to safely and efficiently utilize generative AI, maximizing performance and productivity. By leveraging generative AI’s contextual awareness, content aggregation, and data-driven predictions, employees can streamline information-intensive search processes and focus on complex decision-making and problem-solving aspects of their jobs. Training content creators on best practices and content suitability for generative AI usage can further enhance the system’s performance.

The Ever-Evolving Landscape

The field of generative AI is rapidly evolving, with new models and tuning approaches emerging frequently. Companies embarking on embedding their knowledge into generative AI systems should be prepared to adapt their strategies regularly. Despite the challenges, the long-term vision of enabling easy access to knowledge for employees and customers alike is a powerful incentive. Generative AI is finally making this vision a reality.


Can generative AI models be trained on confidential company information?

Yes, when using private instances of models and implementing measures like turning off chat history collection, confidentiality can be maintained.

How can generative AI models be used in customer service?

Generative AI models can assist customer-facing employees by providing company policy information, product/service recommendations, and solutions to customer service problems.

What evaluation strategies can be employed to ensure the accuracy of generative AI models?

Evaluation can involve testing the models with known questions, employing human reviewers, and assessing criteria such as factuality, precision, and reasoning.

Is training an LLM from scratch a common approach?

No, it is not common due to the massive amount of data and computing power required.

How can generative AI models enhance employee productivity?

Generative AI models can automate information-intensive tasks, allowing employees to focus on complex decision-making and problem-solving aspects of their roles.