ChatGPT (Traditional Chatbot)
Its primary role was to generate text responses to user questions and instructions. While AI agents are “AI subordinates that work autonomously behind the scenes toward objectives,” traditional ChatGPT is likened to a “smart tool that only answers.”

Operator (Released January 2025)
Specialized in automatically operating web browsers and performing online tasks. It could execute restaurant reservations and online shopping but couldn’t perform deep analysis or report creation. Realized through the “Computer-Using Agent (CUA)” model based on GPT-4o, it was trained to analyze GUI (Graphical User Interface) and replicate mouse and keyboard operations.

Deep Research (Released February 2025)
Deeply investigates vast information on the web and generates expert-level research reports in just minutes. It can analyze and integrate hundreds of web sources and return detailed reports with citations, but it was web-browsing only and couldn’t take actions based on that information (like logging in to obtain additional information).

1. Complex Research and Content Generation

Understands users’ ambiguous requests, intelligently navigates the web to collect and filter information, and generates editable slides and spreadsheets summarizing research results. For example, it can create briefing materials for client meetings or analyze three competitors and create slides.

2. Web Browser Operation

Can browse and operate websites just like humans through graphical user interfaces (GUI). Beyond scrolling, clicking, and form filling, it can access information requiring login. This enables automatic execution of complex tasks like airline and hotel reservations and online shopping.

3. Code Execution

Has a Linux-based virtual terminal environment and executes Python and other code for data analysis, file downloads, and processing. This allows AI agents to complete data processing and calculations previously done manually.

4. External App Integration

Integrates with external applications like Gmail and GitHub through OpenAI’s “ChatGPT Connectors.” For example, it can summarize Gmail emails or read calendar appointments for scheduling, executing tasks based on user-specific data.

5. Flexible Planning and Multi-step Execution

Dynamically creates internal plans to achieve given goals and sequentially selects necessary tools to progress tasks. Has advanced ability to reach solutions through trial and error according to situations.

6. Human Collaboration and Safety

User control is always maintained, requiring permission before important operations. Users can intervene to directly operate browsers or interrupt tasks midway. The execution process is displayed on screen as a “thinking process,” visualizing what the AI is doing. Additionally, it’s trained to resist malicious instructions and fraud, with security and privacy-conscious design.

AI Agent Mechanism: Three-Step Cycle

1. Task Creation and Management: Breaks down goals into detailed tasks.
2. Reasoning: Devises optimal methods for executing each task.
3. Action: Executes decided methods according to plan.

By flexibly repeating this cycle, AI agents autonomously accumulate optimal actions toward goals. Additionally, they have the ability to learn independently, utilizing feedback information for subsequent tasks.

Levels of AI Agent Autonomy

Simple Processor
Simple question-answering systems that respond based on pre-set FAQs and scripts, like chatbots. Limited in handling complex tasks and situational adaptability.

Router
Uses LLM output for if/else branching decisions, determining pre-defined workflow branches. For example, systems that route to appropriate workflows based on customer inquiry content.

Advanced AI Agents
• Purpose Understanding and Decomposition: Understands user objectives and breaks them into multiple processes
• Autonomous Action: Can act autonomously toward complex goals without detailed step-by-step instructions
• Learning and Improvement: Utilizes feedback and experience for subsequent tasks, independently learning and improving capabilities
• Reasoning: Enhanced ability to progress reasoning through “multi-step reasoning” that breaks down complex problems step by step, drawing conclusions while referencing external knowledge

1. Dramatic Evolution of LLMs

As of 2025, the evolution of LLMs (Large Language Models) is accelerating AI agent implementation. LLMs are rapidly advancing through the “scaling law” where performance improves as data scale, computation, and model parameters increase.

• OpenAI GPT-2 (2019): 1.5 billion parameters
• OpenAI GPT-3 (2020): 175 billion parameters (approximately 120x scale increase)
• Google PaLM (2022): 540 billion parameters

Previous LLMs had issues with low planning accuracy, complex multi-step reasoning, and long processing times, but these are now being addressed.

• Improved Planning Accuracy: Evolution of reasoning models (models that perform deep analysis and consideration) and hardware accelerators (TPUs, specialized AI chips) enables more accurate real-time response generation.
• Improved Complex Multi-step Reasoning: Ability to break down and reason through complex problems step by step has improved through technological developments like Chain of Thought (CoT).
• Reduced Processing Time: Hardware accelerator evolution has shortened AI inference processing time and improved speed.

2. Expansion of General AI Use and Increased Agent-type Demand

The spread of generative AI has led to recognition of AI’s potential for various tasks, increasing demand for automating more complex operations.

In Japanese companies, generative AI adoption is progressing centered on large companies and IT industries:

• 32% of IT companies have adopted (as of December 2024)
• 23% of finance/insurance companies have adopted (as of December 2024)
• 13.1% of Japanese companies overall have adopted generative AI (as of December 2024)
• Predicted to exceed 50% by around 2030 at this pace

In the future, AI agent use is expected to become standard when discussing AI utilization, with “AI use ≈ AI agent use” anticipated in the 2030s.

3. Advancement of Technological Competition

Major IT companies like Microsoft, Salesforce, and Amazon Web Services are accelerating AI agent development and deployment. OpenAI CEO Sam Altman has stated, “2025 is the year AI agents truly enter society,” citing factors accelerating their spread:

• Improved technological maturity
• Significant reduction in implementation costs
• Accumulation of success cases
• Enhancement of development tools

4. Deepening Labor Shortage and Rising Business Automation Needs

Japan faces limits in expanding its labor force due to declining birthrate and aging population, forcing companies to advance business automation.

According to Ministry of Internal Affairs estimates:

• Japan’s working-age population (15-64 years) will decrease from 75.09 million in 2020
• To 52.75 million in 2050, a approximately 30% decrease

AI agents are expected to contribute to business automation and productivity improvement as a means to solve such labor shortage challenges.

1. Predictive Power (Predict)

The ability to predict the future.

Examples: AI is utilized for diagnosing and predicting water pipe breakage risks and natural disaster risks using satellite image data. Research on medical systems that predict future illnesses humans couldn’t foresee is also advancing, demonstrating AI’s expanding predictive power. In commerce, predictive sales-type commerce may emerge.

2. Discriminative Power (Distinctify)

The ability to find specific events or patterns humans couldn’t notice from massive data.

Examples: Cases where AI identified a hidden phantom self-portrait in Van Gogh’s painting underdrawing, or where AI decoded valuable text information from burned ancient documents that researchers couldn’t read. This demonstrates the ability to extract and classify features from data.

3. Individualizing Power (Individualize)

The ability to optimize according to individual specificity and peculiarities.

Examples: In healthcare, research is advancing where AI proposes personalized treatments from individual genetic and activity data. In education, AI grasps each student’s learning progress and understanding level, utilized for lesson planning, learning guidance, and grade evaluation. In commerce, this connects to the ability to individualize products and services according to user preferences.

4. Communicative Power (Communicate)

The ability to interpret and translate, including understanding user needs through dialogue and providing appropriate information.

Examples: Known medical interview AI where doctor avatars conduct symptom interviews and explain treatment flows before visits, aiming to shorten consultation time. Examples of conversational commerce broadly presenting related products users weren’t explicitly searching for also show expanded communicative power.

5. Structuring Power (Model)

The ability to structure knowledge.

Examples: AI programmers that generate and explain code, or companies where all members except the CEO are AI, are prime examples of expanded structuring power. Recently, AI CEOs have appeared, taking responsibility for overall management, decision-making, and risk management strategy execution with fair judgment prioritizing organizational profit.

6. Creative Power (Create)

The ability to generate new knowledge and create new ideas by combining existing knowledge.

Examples: In Denmark’s Slagelse municipality, AI with creative power attempts a new approach collecting citizen discussions, insights, and proposals to provide information to policymakers. At NASA, AI was used for some satellite equipment, resulting in adopted equipment with shapes humans couldn’t conceive yet structurally superior to conventional ones.

Current State (Late 2024-2026)

AI agents enable rapid customer response and human resource reduction through automated chatbot responses.

• A major retail company reduced recruitment officer inquiry response time by 60% and improved candidate evaluations through recruitment chatbot implementation.
• GMO Media developed an autonomous AI agent specialized in inquiry responses, achieving over 70,000 inquiry reductions, 68%+ work hour reduction, and 12.5 point customer satisfaction improvement in about 1.5 years.

Future Evolution (Post-2030 Predictions)

AI agents will collaborate to enable more advanced customer service. They’ll comprehensively judge customer emotional states, past history, and situational complexity, independently devising and executing solutions even for exceptional cases. Through collaboration with early AGI versions, non-routine inquiry handling will evolve, with routine inquiries fully automated and humans only making final judgments for non-routine responses.

Current State (Late 2024-2026)

Autonomous processing of routine tasks like schedule management and data entry, task priority adjustment, and routine work optimization through integration with multiple business tools are advancing. About 50% of back-office operations are expected to be covered by AI agents.

Future Evolution (Post-2030 Predictions)

Cross-functional collaboration will advance, enabling comprehensive business management. AI agents will handle desk work, significantly reducing corporate workload, and autonomously design and improve optimal workflows by understanding entire business processes. In planning and creative work, expert-level AGI may create optimal plans with humans only making minor adjustments.

Current State (Late 2024-2026)

AI agents analyze customer data to suggest optimal approaches to sales representatives and automate contract management for procedural efficiency and error reduction.

Salesforce’s “Agentforce” provides functions including:

• Natural language customer interactions
• Prospect appointment acquisition
• Sales representative coaching
• Campaign building support

“AI supervisor” functions where AI agents participate in actual customer meetings and provide real-time advice to sales representatives have also emerged.

Future Evolution (Post-2030 Predictions)

Full sales process automation will advance, with AI agents autonomously planning and executing optimal customer responses and contract proposals through comprehensive analysis of market trends, competitive information, and customer psychology. Sales expert AI agents will create negotiation scenarios and provide optimal proposals according to customer challenges, creating immersive proposal experiences personalized for each purchasing decision-maker.

Supply Chain

AI agents advance real-time demand forecasting and logistics planning optimization. Post-2030, they’ll comprehensively understand global supply chains, analyze geopolitical risks, weather fluctuations, and market changes to fully automate inventory management and delivery planning. Complete on-demand production systems may realize hyper-personalized experiences tailored to individual consumer needs.

Medical/Healthcare

Remote diagnosis accuracy improvement and individual treatment proposals based on patient data are advancing. Post-2030, expert-level AGI will fully automate diagnostic support and handle medical consultations. They’ll comprehensively analyze patient genetic information, lifestyle data, medical history, and latest medical research to autonomously propose optimal treatments and preventive measures for individual patients.

Financial Institutions

In loan underwriting processes, multiple specialized AI agents share tasks and efficiently process credit risk scenarios, improving efficiency by shortening review cycles by 20% to 60%. Dynamic pricing and personalized promotions also become possible.

Software Development

AI agents streamline software migration processes and improve productivity by analyzing, documenting, and translating old code efficiently, with quality assurance agents generating test cases to improve accuracy. Code generation and explanation by AI programmers is also possible.

Manufacturing/Warehouses

Fujitsu’s video analysis AI agent analyzes video in real-time in warehouses and factories, providing alerts and suggestions to managers and workers. This autonomously supports work efficiency and safe workplace creation.

Marketing

• Marketing Research and Analysis: Attempts have begun to generate “AI personas (virtual consumers)” from large consumer data and proprietary consumer research databases to verify consumer needs and marketing measure responses. Kirin Beer began verification where they train AI personas with consumer interview voices for new product development, asking AI personas about product concepts and flavors to pseudo-extract customer insights and suppress development period extension.
• Advertising Content Creation: Efficiently creates advertising content using LLMs and image/video/audio generation AI.
• Marketing Planning: Utilizes “multi-AI agent systems” where multiple AI agents work collaboratively, cooperating to achieve common goals and advance marketing planning.

Future Technology Trends

Multimodal LLM

Evolution of multimodal AI that processes multiple data formats including text, images, audio, and video integratively will enable AI to handle more complex tasks. This enables capturing complex emotions beyond surface emotional understanding and contributes to long-term information memory.

Evolution of Reasoning Models

Reasoning models that decompose complex problems and discover hidden patterns and intentions by generating long Chains of Thought (CoT) are evolving. This improves generalization through analogy, with expectations for development into AI capable of PhD-level human-like advanced problem-solving.

Large Action Model (LAM)

AI models that execute actions based on user input, where LLMs understand and generate language while LAMs understand human intentions and execute physical or digital operations. This is expected to apply to AI agents, autonomous robots, smart home control, and more.

• With AI adoption, routine tasks will be replaced by AI, allowing humans to focus on more creative and strategic work.
• “Decrease in white-collar workers” and “increase in no-collar workers (creative jobs as new positions supporting professionals wearing collarless clothes)” are predicted.
• Emergence of AI assistants and increase in “AI labor” taking on human work are also predicted.
• As Japan’s “middle” (young labor population central to work) decreases due to declining birthrate and aging, new organizational forms where AI handles this “middle” while experienced business people handle “entrance” (design) and “exit” (evaluation/execution) portions are thought to enhance overall organizational competitiveness.

For example, situations like 30 humans and 70 AI agents among 100 employees may become realistic.

• AI’s expanded intelligence including predictive, discriminative, communicative, individualizing, creative, and structuring powers realizes improved decision-making quality, early problem and opportunity detection, and business precision and real-time capabilities.
• The “Singularity Enterprise” concept predicts that post-2035, AI will dynamically reorganize business flows in holacracy organizations, with AI agents autonomously taking diverse roles as main actors, causing not only efficiency and advancement of corporate activities but fundamental transformation of organizational structures and business processes.
• 70% of business leaders respond that humans and AI will become partners complementing and enhancing each other’s capabilities, recognizing that business processes and organizational structures will change significantly toward 2030.

Realization of Longevity Escape

Dr. Aubrey de Grey, central to longevity research, proposed the concept of “longevity escape,” and US inventor Ray Kurzweil stated “modern humans can live to 500 years old if diligent.” Insilico Medicine, BioAge Labs, and others use AI to search for aging substances and treatments, while Retro Bio funded by OpenAI’s Sam Altman and Google-founded Calico also research anti-aging and longevity.

Emergence of Transhumans

Historian Yuval Noah Harari points to the emergence of “transhumans” in his book “Homo Deus,” predicting that some wealthy individuals will become transhumans (Homo Deus) through AI and biotechnology and dominate the majority of current humans.

Business Efficiency: Many routine tasks including report creation, market analysis, data processing, accounting, and code generation will be streamlined by AI agents. This allows employees to focus on more creative and high-value work.

Democratization of Expertise: Even without advanced research analysis or programming skills, high-quality results can be enjoyed through AI agents, enabling output beyond human and budget constraints.

New Jobs and Skills: Demand for “personnel who can master AI agents” will increase, with AI literacy and agent supervision/evaluation skills becoming important.

Unclear Effective Utilization Methods: “Not knowing effective utilization methods” is cited as Japanese companies’ primary concern regarding generative AI adoption.

Increasing Costs and Unclear Business Value: According to Gartner’s report, over 40% of AI agent projects may be canceled by the end of 2027 due to “increasing costs” or “unclear business value.”

“Agent Washing”: Movements exist to rebrand existing AI assistants and chatbots as “agent-type AI” without substantial agent capabilities, with many companies facing risks of investing in products without substance.

High Costs and Usage Restrictions: Advanced AI agent functions like OpenAI’s Deep Research are currently limited to Pro plan ($200/month), potentially burdensome for individual users and small businesses. Monthly query limits also exist (Deep Research: 120 times/month).

Handling Diverse Corporate Data: Whether AI can handle companies’ diverse and massive data and ensure security becomes a challenge.

• First, consider incorporating AI agents in narrow scopes or partially, such as chatbots related to company sites and services, and accumulate utilization know-how.
• During know-how accumulation, gradually understand risks arising from AI implementation (legal issues, degree of human intervention and approval, data and model obsolescence).
• Like Klarna’s success case, rather than aiming for perfection from the start, starting with small tasks and gradually expanding through trial and error is important.

• Actively delegate tasks that can be entrusted to AI agents, with humans shifting to more strategic and creative domains.
• Important to have a new management perspective of “managing and nurturing” AI agents rather than treating them as mere tools.
• Designing necessary skill sets in AI-utilizing society and acquiring them through daily work and careers is important for future business people. Abilities to master AI agents and coordination abilities for good human-AI agent collaboration will be needed.
• As new AI tools and services continuously emerge, continuous AI utilization training is essential to improve abilities to select optimal ones for your company and literacy to critically examine AI output and judge its reliability.

• Concerns are rising about information leaks, security risks, copyright infringement possibilities, and inclusion of ethically inappropriate content or bias accompanying AI utilization.
• Companies should strengthen AI governance and create organizational foundations to respond to regulations now.
• Creating internal guidelines clarifying what information can be input to AI enables safe utilization.