How to manage two types of generative artificial intelligence

As companies continue to experiment with and realize business value from generative artificial intelligence, executives are implementing the technology in two different ways.

According to a new research briefing from researchers Nick van der Meulen And Barbara H. Wixom At the MIT Center for Information Systems Research, organizations distinguish between two types of generative AI implementations. The first broadly applicable generative AI tools serve to increase personal productivity. The second, tailored generative AI solutions are designed for use by specific groups of organizational stakeholders.

The study, based on roundtable discussions with members of the MIT CISR Data Research Advisory Board and interviews with executives, outlines the two approaches and highlights unique challenges and management principles for each.

Broadly applicable generative AI tools

Generative AI tools such as conversational AI systems and digital assistants embedded in productivity software are inherently broadly applicable. They are versatile and their use is typically defined and refined by their users, the researchers write.

“This is AI for everyone,” said JD Williams, vice president and chief data and analytics officer at global animal health company Zoetis, who is a member of the MIT CISR Data Board. “This is where you bring in external products and privatize them within the company so that your data is protected.”

According to the researchers, generative AI tools present companies with four major challenges:

1. Because generative AI tools rely on large language models trained to predict the most likely word sequence in a given context, they often produce output that is commonplace. Therefore, the quality and relevance of the output depends on the specificity of the prompts entered by the user.
2. Generative AI tools can lack context, contain biases, present false or misleading information as fact, and fail to compute. Consequently, users must continually critically evaluate a tool’s results to avoid accepting biases or inaccurate claims.
3. Untested, publicly available generative AI tools can pose significant risks, especially when employees use them professionally. These risks include loss of data, loss of intellectual property, copyright infringement and security breaches.
4. Generative AI tools are expensive. Providing users with licenses for tools from multiple vendors can quickly become costly once free trials and early adoption incentives expire.

To address these concerns, companies should give their employees sanctioned access to a select number of generative AI tools to create a safe space for experimentation. To enable safe and successful deployment of generative AI, the researchers suggest that leaders do the following:

• Develop clear usage guidelines. These policies should be developed by cross-functional leadership teams with representatives from technology, legal, privacy and governance. Policies should specify which tools are permitted under which conditions and outline the associated risks and possible consequences. Williams said mitigating risk, protecting data and ensuring regulatory compliance are critical to any AI governance framework. “You want to be innovative and fast, but also risk aware, data secure and compliant,” he said.
• Invest in training. Organizations should establish AI policies and assessment practices, including training employees to effectively instruct and query generative AI tools, understand the underlying models, and use the tools responsibly.
• Standardize with a select group of vendors. Build a cross-functional team of likely users of generative AI tools to identify which tools offer the most potential for your business.

Generative AI as a tailor-made solution

Generative AI solutions are business case-driven development initiatives that address strategic business objectives and create value for specific groups of organizational stakeholders, ideally at scale, the researchers write. Organizations fund these solutions after they meet innovation criteria related to end-user attractiveness, technical feasibility and business viability.

“(Companies) deploy these solutions in specific functions that perform specific tasks,” Williams said. “In manufacturing, for example, it could be about monitoring processes and products to ensure they are moving in the right direction during production. There are a lot of great applications here.”

Although generative AI solutions share some similarities with other AI initiatives, they present three unique challenges, according to the researchers:

1. As more employees begin to realize the potential of generative AI, companies risk developing “shadow generative AI,” where groups of stakeholders independently pursue unapproved solutions with the help of eager vendors.
2. A few vendors own and control most of the base models that support generative AI solutions. This complicates organizations’ understanding of models and their own ability to assess bias and predict model behavior, which can pose various risks, including data leaks and inaccurate results. Uncertainty about future usage, model performance and pricing also makes it difficult for companies to estimate the long-term operating costs of generative AI solutions.
3. The value that companies derive from generative AI solutions depends on whether the company purchases a solution, extends a vendor model, or develops its own solution. Depending on the approach, there are trade-offs in terms of transparency, contextual awareness and cost.

Companies can benefit from generative AI solutions by deploying them across functions. To be successful with targeted generative AI solutions, organizations can also do the following, the researchers write:

• Establish a formal, transparent generative AI innovation process. Organizations need clear governance structures, early and consistent stakeholder engagement, and a focus on scalable solutions.
• Formulate guidelines for generative AI development decisions. Leaders need to differentiate between the different approaches to generative AI development to help teams make informed decisions, as there are different advantages and disadvantages when purchasing, building, or improving generative AI models.
• Create a generative partnership strategy for AI vendors. Effective partnerships with generative AI providers are based on mutual understanding and long-term collaboration. This promotes adaptability and continuous improvement, which benefits both parties.

AI Executive Academy

In person at MIT Sloan

Register now