Overview

In this episode of *ハイパー起業ラジオ*, hosts Kazuhiro Obara (IT critic, former McKinsey/Google/Rakuten executive) and Kensuke (serial entrepreneur, founder of Al Inc.) argue that the AI industry is undergoing a decisive shift from "universal AI agents" to "workflow-type AI agents"—a transition they believe will reach a tipping point in the coming months. The stakes are high: while universal agents like ChatGPT promise to do everything, they often deliver inconsistent quality, whereas workflow agents (where humans pre-define a sequence of tasks and let AI execute specific steps) offer reliability, reproducibility, and strategic advantage for businesses. The conversation has the feel of two seasoned tech insiders excitedly unpacking breaking news from OpenAI and Google, while grounding their analysis in real-world consulting experience and organizational strategy.

---

2:02The Two Types of AI Agents: Universal vs. Workflow

Obara opens by defining the core distinction that frames the entire episode. AI agents today fall into two broad categories. The first is the universal (or "omnipotent") AI agent—exemplified by ChatGPT's agent mode or Google's Gemini—which can take a vague instruction like "make me a webpage" or "book a restaurant" and figure out the entire process autonomously. The second is the workflow-type AI agent, historically represented by tools like Dify, where the user pre-defines a fixed sequence of steps (a "flow") and deploys AI only within specific, bounded tasks within that flow.

Obara uses a culinary analogy: a universal agent is like a "omakase chef" who, told "make something autumnal," will forage for ingredients, invent a recipe, and cook from scratch. A workflow agent is like a fixed-course menu where the structure is predetermined—appetizer, main, dessert—but the chef can customize each course for the guest (e.g., "use matsutake mushrooms for me," "skip dessert"). The key insight: for businesses, the workflow model is superior because it guarantees reproducibility and process visibility. When a process is visible, it can be transferred across departments, customized per team, and audited for quality.

---

4:45The Breaking News: OpenAI and Google Launch Workflow Tools

The hosts pivot to the week's major announcements. At OpenAI's developer event DevDay, the company unveiled three major features, the second of which was Agent Kit—a tool that lets users build workflow-type agents by visually connecting "blocks" like Lego pieces. For example, a user could create a flow where: (1) a user inputs text, (2) a "research agent" investigates the topic, (3) a "report agent" compiles findings, and (4) a "webpage agent" publishes the result as HTML. Each block is a reusable, standalone AI agent that can be deployed across different workflows.

Simultaneously, Google released Gemini Enterprise, a new enterprise plan that includes workflow capabilities, and quietly updated gOperl (an experimental workflow-building tool) to support Japanese. The hosts emphasize the significance: until now, workflow tools like Dify were powerful but niche—used mainly by technically savvy early adopters. With both OpenAI and Google aggressively pushing workflow features, the hosts argue that the mass-market tipping point has arrived. Kensuke notes that while he hasn't personally tested the new tools yet, early feedback suggests they are significantly easier to use than Dify, especially because the AI itself can now generate a first draft of the workflow from a simple natural-language request.

---

6:57The "No-Code" Leap: From Lego Blocks to Lego Kits

A central theme is the evolution of usability. Kensuke explains that earlier workflow tools like Dify required users to manually select and configure each block, write prompts for each step, and understand input/output relationships—essentially requiring a programmer's mindset. The new generation (OpenAI's Agent Kit and Google's gOperl) changes this fundamentally: now, a user can type something like "create a workflow that researches a topic, writes a report, and turns it into a webpage," and the AI will generate a first-pass workflow automatically, complete with suggested blocks and connections.

Obara likens this to the difference between being given a pile of Lego bricks (Dify) versus being given a Lego kit with a pre-designed model and instructions (Agent Kit). The user still has freedom to customize—swap blocks, adjust parameters, add new steps—but the initial barrier to entry drops dramatically. Kensuke adds that the individual blocks themselves have become more powerful: they can now upload files, connect to external tools via MCP (Model Context Protocol), and handle complex operations that previously required separate, manual integration. This convergence of ease-of-use and power, the hosts argue, will enable non-programmers to become "AI workflow creators" in a way that was previously impossible.

---

10:25The Mindset Shift: From "Using AI" to "Delegating to AI"

Obara introduces a conceptual framework he uses in his consulting and keynote speeches: the difference between the ChatGPT era and the AI agent era. In the ChatGPT era, users had to adapt their own workflows to fit the AI—learning prompt engineering, restructuring tasks to get good outputs. In the AI agent era, the paradigm flips: you delegate tasks to AI agents exactly as you would to a human subordinate. You don't need to change your process; you just need to decide which parts of your existing process to hand off.

The challenge, however, has been: who builds these agents? Until now, building a reliable AI agent required programming skills. But with the new no-code workflow tools, Obara argues that the barrier has dropped to the point where anyone who can describe a task can now build an agent to execute it. Kensuke adds a practical note: the first version generated by AI is rarely perfect—it may have bugs or logical gaps—but the iterative process of debugging and refining (by giving natural-language feedback like "this step should come before that one" or "the output format is wrong") is far faster than learning to code from scratch. He compares it to learning how to give effective instructions to a talented but inexperienced new hire.

---

14:36Real-World Adoption: The "Red Pen Teacher" Experiment

Kensuke shares a concrete case study from his consulting work. He recently ran a project where 50 employees were given accounts on Dify and asked to build workflows over several months. He and his team reviewed every submission—acting as "red pen teachers" for AI workflow design. The results were revealing: most people struggled. The core difficulty was not the tool itself but the conceptual leap of thinking in terms of inputs, outputs, variables, and process flows—a fundamentally programmer-like mental model. Even after multiple rounds of training and feedback, only about 10 out of 50 people became genuinely proficient at building workflows from scratch.

However, Kensuke notes that this is still a viable model: if 10 people in an organization can build workflows, and the other 40 simply *use* those workflows, the company can achieve significant productivity gains. With the new tools from OpenAI and Google, he estimates that the number of people who can become proficient might rise to 20–25 per 50—a meaningful improvement. The key enabler is that the new tools provide templates and examples that help non-programmers understand the abstract concept of "input → transformation → output" by showing concrete instances of their own work.

---

18:11Strategic-Level AI Workflow: From Cost Reduction to 10x Speed

Obara introduces a two-tier framework for thinking about AI workflow adoption in business. Tier 1 (tactical) is about cost and speed advantages: building workflows that let you produce things cheaper or faster than competitors—faster quotes, faster delivery, lower prices. This is valuable but ultimately imitable.

Tier 2 (strategic) is about end-to-end workflow integration—connecting multiple workflows so that the entire business process runs at dramatically higher velocity. Obara illustrates with a sales example: a company could build a workflow that automatically records sales calls (with client permission), generates transcripts, extracts best practices, and compiles a daily "improvement digest." That digest then feeds into another workflow that assigns the best sales approach to the next day's client meetings. The results are tracked in a dashboard showing conversion rate improvements. What traditionally takes a week to compile and a month to roll out can now happen in three days. The result: 10x the learning velocity of competitors.

Kensuke connects this to OpenAI's five-level AI maturity model: Level 1 (chatbots), Level 2 (reasoners), Level 3 (agents), Level 4 (innovators), Level 5 (AI corporations). He argues that end-to-end workflow integration corresponds to Level 4—where AI not only executes tasks but also drives continuous improvement and innovation through rapid experimentation. Obara adds that many of his consulting clients are now asking about Level 5 (fully autonomous AI companies), but in practice, most are still at Level 2 or early Level 3. His role is to help them see the path from tactical efficiency to strategic transformation.

---

25:32The Human Side: Lower Hiring Bars and the "Autonomous" Future

A surprising benefit of workflow AI, the hosts note, is that it lowers hiring requirements. When AI handles specialized knowledge and routine analysis, humans can focus on "horsepower"—execution, client interaction, and creativity. This means companies can scale their workforce faster because they no longer need to find rare, expensive experts for every role. Kensuke cites companies like Shift and M&A Research Institute as examples of firms that have successfully used this model to grow rapidly.

Looking further ahead, Obara points to the concept of autonomous (self-driving) organizations. At the HR Tech conference in Las Vegas (September 2024), the dominant theme was "from agent to autonomous." The idea is that back-office functions like accounting, HR, and contract management can become 95% autonomous, freeing up human talent for higher-value work. The data generated by these autonomous systems becomes a "treasure mountain" for strategic decision-making.

However, Kensuke raises a sobering counterpoint: this future is exciting for executives and core talent, but terrifying for ordinary employees whose roles may become redundant. Obara acknowledges the tension, noting that the same forces that create opportunity for some will threaten others. The hosts discuss how companies like Hoshino Resorts (led by its founder) have been ahead of the curve for two years, requiring all customer-facing staff to spend 15 hours per week building no-code apps that enhance guest hospitality—turning AI into a tool that amplifies human warmth rather than replacing it.

---

32:48The Organizational Challenge: Building a Sustainable Learning Culture

The episode concludes with a practical discussion of how companies can build internal capability for workflow AI. Kensuke notes that the first step is the hardest: the gap between those who can build workflows and those who cannot remains significant, even with better tools. He shares that CyberAgent, for example, runs a two-week intensive Dify training program continuously, cycling hundreds of employees through it. The key is not a one-time workshop but a sustained, rolling program where early adopters become trainers for the next cohort.

Obara adds that this mirrors the pattern used by Rakuten University (Rakuten's internal training program): identify motivated individuals, train them intensively, and then have them return to their teams to train others. The goal is not to train everyone at once—which is impossible—but to create a self-reinforcing cycle where each wave of graduates produces the next. He emphasizes that companies serious about AI adoption need to commit to this kind of organizational infrastructure, not just buy tools and hope for the best.

---

Conclusion

What stays with the listener is the sense that a major inflection point has quietly arrived. The combination of OpenAI's Agent Kit and Google's Gemini Enterprise means that workflow AI—previously the domain of specialists—is about to become accessible to every business. The hosts' central message is both urgent and practical: the companies that will win are not necessarily those with the best AI models, but those that build the organizational capability to design, deploy, and continuously improve AI workflows at scale. The episode matters because it moves beyond hype to offer a concrete framework for action—from tactical cost savings to strategic transformation—while honestly acknowledging the human and organizational challenges that lie ahead.

---

要点

• AI agents are bifurcating into two types: universal (do-everything) and workflow (pre-defined process with AI at specific steps); for business, workflow agents offer superior reliability, reproducibility, and process visibility.
• OpenAI's Agent Kit and Google's Gemini Enterprise/gOperl, both released in the same week, mark a tipping point where workflow AI becomes accessible to non-programmers through natural-language workflow generation.
• The new tools reduce the barrier to entry from "programmer mindset" to "ability to describe a task"—users can now generate a first-draft workflow by typing a sentence, then refine it iteratively.
• Real-world experiments show that only about 20% of employees can become proficient workflow builders even with training, but that 20% can create tools that the other 80% use, delivering significant organizational impact.
• Strategic AI workflow adoption means connecting multiple workflows end-to-end to achieve 10x velocity in learning and improvement cycles, not just marginal cost savings.
• AI workflows lower hiring requirements by offloading specialized knowledge to AI, allowing companies to scale faster with less expensive talent.
• The long-term trajectory points toward "autonomous organizations" where back-office functions become 95% self-driving, but this creates a sharp divide between opportunity for executives and threat for ordinary employees.
• Successful companies build sustainable internal training programs (like CyberAgent's two-week Dify bootcamp or Rakuten University's peer-training model) rather than relying on one-off workshops.