The Theme for Generative AI This Year: Building the Environment

This Year's Generative AI Theme is "Environment Building"

Last year, the performance of Large Language Models (LLMs) improved dramatically, and the practical application of generative AI accelerated all at once. However, just having excellent models isn't enough to extract true value within business and society—this is why "environment building" is attracting attention.
I will delve into this theme from two perspectives: the "AI environment" and the "human environment."

The AI Environment

Importance of the Application Layer to Leverage LLMs

Until last year, high-performance LLMs such as GPT-4o and Claude-Sonnet-3.5 appeared one after another. In addition to functional progress like text summarization and code generation, they are starting to handle more advanced task execution, such as integration with external systems and multi-step reasoning.
However, even if models become high-performance, utilization will be limited without a mechanism to seamlessly incorporate them into operations or services.
This is where the application layer that connects LLMs with external tools gains attention.

• MCP (Model Context Protocol)
  A standard protocol that allows AI assistants to access external data sources safely and flexibly. It is expected to control access scope and clarify data handling rules.

https://github.com/modelcontextprotocol/servers

• Dify
  An open-source LLM application development platform. Since applications and workflows using generative AI can be designed on a GUI basis, it lowers the barrier to entry.

https://dify.ai/jp

• Browser Use
  A framework that allows LLMs (agents) to operate web browsers. It automates everything from search site navigation to inputs on SaaS screens, enabling labor savings in information gathering and form submission.

https://github.com/browser-use/browser-use

• OpenHands
  A platform where AI can handle tasks related to software development, such as command execution and code modification. By automating part of the development process, humans can focus on higher-level decision-making.

https://github.com/All-Hands-AI/OpenHands

Or Devin

https://devin.ai/

Key Points for Environment Building

1. Standardized Connection Methods
   By managing which data or APIs an LLM can access at the protocol level, like with MCP, both safety and scalability are achieved.

2. Workflow Visualization and Verification
   By utilizing tools like Dify, the LLM's reasoning results and the status of integration with external services can be displayed and verified in an easy-to-understand format.

3. AI Agents as Executing Entities
   By creating an environment where AI can actually take action (executing commands, operating browsers, etc.) through tools like Browser Use or OpenHands, it becomes possible to improve operational efficiency beyond the scope of mere chatbots.

I believe that providing such environments where AI operates applications will become increasingly common.

The Human Environment

Necessity of Changing Organizations and Business Processes

Even if AI agents become more sophisticated, they will not be able to demonstrate their full potential if existing organizational structures and business flows remain unchanged. Rather, "AI-centric" process design and the development of APIs for internal systems are essential. For example, significant efficiency can be expected simply by reorganizing tasks so that AI agents handle initial inquiries while humans focus on exceptions and final decisions.

• Redesigning Business Flows

  • Inquiry handling systems with AI integrated by default
  • Automation of internal documents and code reviews, accompanied by human final approval flows

• Permissions and Access Management

  • Establishing rules on the scope of operations allowed for LLMs from the perspective of security and privacy
  • Clarifying risk mitigation measures, such as conducting log audits and providing emergency stop mechanisms (Kill Switches)

• Literacy and Governance

  • Sharing the limitations and possibilities of AI with everyone from management to frontline staff and providing education for its appropriate use
  • Clarifying organizational strategies, such as choosing between in-house development and external vendors, or open source and SaaS

Rethinking Business Models

Beyond business flows, we expect to see more movement toward redefining business itself as "co-creation with AI agents." Rather than building all applications in-house, an increasing number of companies may emphasize speed by actively integrating tools from cloud vendors and startups. Such a horizontal division-of-labor ecosystem may become the mainstream in the future.

Even if the AI environment is ready, I believe it will be difficult to apply it to business unless the human environment that sits on top of it is also prepared.

Summary

Last year, the rapid evolution of LLMs was a hot topic, and many companies and developers went through the stage of "trying out" generative AI. However, this year, "environment building" to achieve full-scale operation and connect it to business results will likely be a major theme.

• AI Environment
  The application layer that safely and efficiently connects LLMs with external tools is drawing attention, and platforms and protocols such as Dify, MCP, and OpenHands are expected to be further developed.

• Human Environment
  Business flows and organizational structures that facilitate AI agent activity, along with security, risk management, and improved literacy, are essential. It will be necessary to comprehensively transform these elements.

In other words, I believe that as "environment building on both the technical and organizational sides" progresses, the possibilities for generative AI will expand significantly.

While people are talking about "AI agents" and the like, I suspect this will be a year where we focus more on the environment surrounding them instead.