次のボトルネックは顧客を深く理解すること

This chapter centers on the speaker's view that AI's work is shifting from executing tasks to understanding how a particular company works. The talk presents a problem: how can deeply embedded customer work be expanded without continually increasing headcount?

Here, forward-deployed engineering (FDE) means entering a customer's organization, deeply understanding the work and processes carried out there, and building and deploying a technical solution that fits that customer. In other words, the role is not simply to hand over a finished product from a distance. Knowing the customer's working environment is itself an important part of the role.

This kind of bespoke work has a clear benefit. It can uncover rules and ways of working that exist only at that customer's company. At the same time, the deeper the engagement with each customer, the more human time it requires per customer. If the same method is used to add more customers, it appears that the number of FDEs must grow with the amount of work. The speaker uses this tension between the value of deep customer understanding and the constraint of not increasing headcount exponentially as the starting point of the talk.

According to the speaker, AI is approaching a stage where it can execute work from beginning to end, not only provide one-off assistance. For example, it can receive input, look up the needed information, process it, and deliver the result to the expected place as one connected task. The speaker says that current models and execution environments can perform this kind of execution almost perfectly. He also believes that AI's intelligence itself is no longer the main constraint. However, this is not a measurement result for every kind of work, nor does it mean that all execution problems have disappeared.

This execution involves more than the model alone. The model interprets input and decides what to do next. A harness is the execution environment that lets the model carry out the work. Browser-operation and API-calling tools, along with connections such as MCP (Model Context Protocol), are added around it. The speaker explains that combining these parts lets AI carry out a task to the end. The talk does not show the specific implementation methods or safety measures.

Example (a generalization for understanding): Simply telling AI, "Please handle the sales work," does not guarantee the right result at a particular customer's company. In a generic execution, it might extract information from an email, register it in a CRM, and draft the next communication. But one company might prioritize review by a particular employee, while another might prioritize contract size. The person who reviews the work, the exceptions that are allowed, and the names used internally may also differ by company. Without knowing these differences, AI may execute the task without actually understanding that company's work correctly.

This is where the bottleneck moves. If we only give a model tools and improve its ability to execute, we can expand that as a shared capability. But customer-specific judgments, owners, exceptions, and priorities are difficult to compress into one simple prompt or API input. Finding the company-specific context that employees use in their daily work requires an FDE to enter the field and understand how the work actually operates. The speaker therefore argues that, before asking "What should AI do?" the next constraint is knowing "How does this company work?"

Figure 1: The title slide for "The Next Bottleneck." The screen shows "The Next Bottleneck" and "AI Forward Deployed Engineering," with the name Varick Agents also visible at the top. The slide establishes the talk's overall frame: the next constraint in FDE. The relevant section is at 00:57 in the video.

This concern is not merely a title or promotional phrase. If customer-specific understanding is expanded only through people, deeper work remains constrained by headcount. Later, the talk proposes an agent that supports FDE work and a platform that runs agents on top of customers' existing systems as responses to this constraint. The goal is not to remove human conversation or customer understanding, but to extend the human time needed for deep understanding as far as practical.

Note: The opening speaker's name and the company name are not spelled consistently in the transcript. The statements introduced here—that execution is almost perfect and intelligence is no longer the main constraint—should therefore be treated as the speaker's position, not as independently verified facts.

The key point is that as AI becomes able to execute work, the importance of understanding each company's operations and changing them to fit AI increases. The FDE is responsible for finding that customer-specific context. When the number of customers grows, the central challenge is not only the ability to execute one task, but how to scale the ability to understand customers deeply enough.