How to address the "proliferation" of generative AI in companies: Balancing AI utilization and governance.

Three obstacles to the use of generative AI in businesses

I believe there are three major obstacles to using generative AI in businesses.

1. The initial hurdle: We don't know where to begin as a company.
2. The accuracy barrier: Expectations rise without a clear definition of what constitutes a practical, usable answer.
3. The hurdle of proliferation: After just starting out, various systems are operating in a disjointed manner.

What I want to draw particular attention to here is the "proliferation problem" that results from prioritizing speed.

Also known as "AI sprawl," this term refers to a situation where AI tools and applications proliferate within an organization, making them unmanageable.

Many companies find themselves in a situation where each department independently builds its own RAG (Ragnadynamic Agencies) or chatbots, making it impossible to track who is using what AI and where.

▼I want to know more about RAG
⇒ Retrieval Augmented Generation (RAG) | Glossary

Back to Top

Why is there such a proliferation of AI systems?

Many of these proliferations stem from a positive attitude on the ground.

There are three underlying factors.

Low implementation costs

Many AI-generating services are offered as SaaS or APIs, making them easy to start using immediately.

The system is structured in such a way that departments can easily complete the approval process at their discretion, making it easier for implementations to proceed by bypassing the regular procurement and approval processes.

Low technical barriers to entry

Because LLM provides plausible answers when prompted, you can build chatbots and RAGs even without specialized development skills.

Locally optimized incentives

In a culture where "just try it" is valued, it seems more rational to solve one's own department's operational problems immediately rather than waiting for the company-wide infrastructure to be established. Because speed takes precedence over the widespread sharing of results, each department pursues its own optimization, which ultimately leads to a proliferation of different approaches throughout the company.

These factors combine to create a situation where, rather than anyone intentionally causing a proliferation of AI systems, a series of rational decisions by different companies have resulted in AI systems scattered throughout the organization that are beyond the company's control.

Back to Top

Why is the proliferation of generative AI dangerous for businesses?

The reasons can be broadly summarized into three problems: "lack of control," "duplication of costs," and "data inconsistency."

First, it becomes difficult to ascertain whether AI is being used in accordance with the rules established by the company, leading to a lack of control.

Furthermore, multiple departments end up paying double or even triple costs for similar systems.

Furthermore, because each system operates independently, there is a risk that data sources will not be standardized, leading to the generation of responses based on outdated company regulations or incorrect data. If different answers circulate from department to department, it could lead to confusion and incorrect decision-making on the ground.

Back to Top

To maintain an offensive stance, we need to systematize our defense.

Is restricting usage again the right solution to curb this proliferation? The answer is NO.

To maintain control without sacrificing the speed of each department, and to connect individual initiatives to company-wide assets rather than letting them end as isolated points, I believe a centralized management and distributed execution model for promoting AI is necessary.

So, what exactly should you keep in mind? Here are three key points:

Centralized model management

Instead of each department individually contracting for a wide variety of models, the company will centrally manage them and provide the most suitable model according to the application.

Visualization of usage

This will clarify who consumed how many tokens, when, and how many, allowing for better identification of shadow AI and cost optimization.

Early detection of risks

We will create a system that allows us to understand usage patterns from logs and detect early instances of confidential information being entered due to unintentional mistakes or inappropriate usage patterns.

So, how are these actually being implemented? In the next section, we will introduce practical examples from Saison Technology.

Back to Top

Examples of Saison Technology's Practices

From here, I will introduce the governance of the generative AI platform that Saison Technology is actually working on.

Based on the three points mentioned above, our company operates our governance not as a one-off initiative, but as a continuously evolving system by running a feedback loop of "visualization → analysis → improvement" centered around a dashboard.

Specifically, we implement it in the following three cycles:

① Analysis of usage history

We analyze data not only by the number of active users and interactions, but also by criteria such as model, application, and department. This allows us to replicate successful patterns from departments where the system is being used effectively, and also helps us review initiatives that are not proving successful.

② Expanding knowledge through prompt analysis

We analyze prompt trends and organize frequently asked questions into FAQs and templates.

In the initial stages of implementation, many people said, "I don't know what to ask the AI," so we played a role in broadening the entry point for AI utilization by visualizing frequently asked questions as a UI.

Now that its use has matured, we have shifted the focus of our analysis to extract common business challenges across departments and identify areas with low response accuracy. By feeding these analysis results back into improving prompts and reviewing RAG reference data, we are contributing to improving the overall accuracy of the system.

③ Autonomous system updates

We check the prompts for any rules that should be added as guardrails (prohibited actions), and continuously update the guardrails themselves. For example, if we find an unexpected trend of confidential information being entered, we reflect this in the guardrails.

The evolution of generation AI is rapid, and rules established once quickly become outdated. We aim to create a system that continuously updates the rules themselves to match actual usage.

Back to Top

Finally

When we hear about AI governance in a company, we tend to think of it as something that "restricts its use."

However, the true purpose of governance is not to impose restrictions, but to create an environment where we can confidently accelerate the use of AI.

We become cautious because we don't know where the risks lie. Conversely, if companies understand what's happening within their organization, they should be able to push forward with AI adoption more boldly.

First, we need to organize our scattered AI systems and establish a management system that allows us to see the whole picture. I believe that's what's necessary to take the next step.

If you have questions or concerns such as "We don't have a grasp of how AI is being used within our company" or "We want to consolidate the disparate AI infrastructure across departments," please feel free to contact Saison Technology. We will support you in building a foundation that balances the speed of AI adoption with effective governance.

Back to Top

Back to Top

The person who wrote the article

Affiliation: Technovation Center, R&D Engineer

Kawaguchi Himeka

Since joining the company, I have participated in projects to build data analysis infrastructure for clients as a data engineer. Subsequently, I was in charge of creating a system for improving business processes using generative AI for internal use. Currently, as an R&D engineer, I am working on the development of a generative AI platform. I am striving every day to grow into a leader who can lead a team, leveraging my past experience. On my days off, I enjoy traveling and dancing.
(Affiliation is as of the time of publication)