実行だけでは足りない理由

The speaker explains that far more people can now use AI to carry a task through to the end than before. He says that the next difficulty in adopting AI is not simply executing a task, but understanding how each individual company actually does its work.

What does an end-to-end task mean?

Here, an end-to-end task does not mean asking AI to help with only one part. It means executing the whole flow of one piece of work: setting the goal, collecting the needed information, moving through several steps, and checking the result.

The speaker asks the audience who had already been using AI to complete this kind of whole task. Few people raise their hands at first, while many more do so now. This contrast illustrates a change in the audience's experience; it is not a controlled survey result.

For example, having AI revise one paragraph is assistance with a single part. Preparing a proposal for a customer, by contrast, may require researching materials, organizing the content, creating the necessary files, and making everything ready to submit. The latter is an end-to-end flow. Because its steps are connected, making one step convenient does not necessarily change the whole task.

The bottleneck moves from execution to understanding

In the speaker's view, progress has come not only from model intelligence but also from execution harnesses, browser operation, APIs, and MCPs. These connections and control mechanisms let AI do more than answer questions: they let it move through several steps of work. However, this talk does not explain how any of these mechanisms are implemented.

It is important to separate three layers here. The model makes judgments and generates outputs. Harnesses and other tools let the model operate external screens and systems. Operations design determines who does what, in which order, and under which conditions inside the company. Giving people a model and tools does not automatically change this third layer.

The speaker makes the strong claim that execution itself is now possible close to 'near perfection' and that intelligence is no longer the main constraint. This is presented as the speaker's position. It does not mean that the same performance has been confirmed by measurement for every task and environment. The important point is where he places the center of the constraint.

Do not only add AI; redesign the operation

Introducing tools such as Cursor, Claude Code, Codex, or Factory does not by itself make a company's operation AI-centered. The tools may speed up an individual's work, but if roles, approvals, information handoffs, and exception handling stay the same, the organization's overall flow does not change.

The speaker proposes a shift from human-centered operations to AI-centered operations. This does not mean that humans immediately disappear. It means reviewing the workflow first and deciding which parts AI handles, which parts humans check, and which parts humans always handle. This is a proposed future operating model.

Customer-specific context is not easy to extract

Supplementary example: The same sales function differs by company

Healthcare sales and sales at a SaaS company can both be called 'sales.' Yet the customers, internal procedures, required information, and decisions by other people may differ. This contrast is a supplementary example showing that a job title alone cannot describe the actual work. The speaker does not list every difference between these two kinds of sales.

A company's context can be hidden in the experience of its employees and teams. It includes knowledge such as, 'For this customer, use this order,' or, 'Under this condition, check with another owner.' It is difficult to turn this kind of tacit practice directly into the input for a simple model or API call. It also includes undocumented judgments and handoffs across departments.

The FDE's two responsibilities

For this reason, the speaker gives an important role to an FDE (Forward-Deployed Engineer). An FDE goes deeply into a customer's organization and first understands how the current operation actually works. The FDE then rethinks tomorrow's operation around AI. In other words, an FDE both maps the present and designs the map for the future.

These two responsibilities connect to the staffing problem introduced at the start. The deeper the understanding of each customer's operation, the more people the work appears to require for each customer. Even if AI becomes better at execution, relying only on humans to discover context and redesign operations would mean expanding the team in proportion to the number of customers. That is why, in later chapters, the speaker proposes agents and a platform to support FDE work. The idea is not to replace humans immediately, but to extend their ability to understand customers and change operations.