11:30 - 13:14
AI, agentic platforms, and the wider customization problem
Video segment: 11:30–13:14. In this part of the talk, Kevin Bai offers a personal hypothesis about why forward-deployed engineering (FDE) may matter to more companies in the AI era. This is not a claim that every AI company should copy Palantir's historical model.
The central idea
AI can make it easier to create code and sophisticated software. But the speaker's point is not only that engineers can build faster. His bigger concern is what happens after a company sells increasingly flexible software to a customer.
More flexible software can solve more kinds of problems. It can also be harder for a customer to understand, choose, configure, and implement well. The customer may own the product but still not know how to turn its capabilities into a useful business result.
That creates a possible role for customer-facing engineers:
Easier creation of advanced software
↓
More customizable, increasingly agentic platforms
↓
More possible ways for each customer to use the platform
↓
A larger understanding and implementation gap for some customers
↓
More value from FDE-like help to deliver an outcome
An implementation gap is the distance between what a platform can technically do and what the customer can actually put into operation. The gap can include technical work, but it can also include deciding which workflow to build and how it should fit the customer's business.
What “agentic” means here
The talk uses agentic platforms as a broad description, not as a detailed technical definition. In this limited context, the important idea is that AI-based platforms are becoming more capable and more open to customer-specific use. Do not read this segment as a specification for what an AI agent must contain.
The important distinction is between configurable and customizable software:
| Term | Plain-language meaning | Why it changes delivery |
|---|---|---|
| Configurable | A customer selects among intended settings or options. | The product has already decided much of the shape of the solution. |
| Customizable | A customer can build or adapt more of the solution around its own needs. | The customer has more power, but also more implementation choices and responsibility. |
The boundary is not absolute. A product can offer both. The speaker's hypothesis is that agentic platforms tend toward the more customizable side, so customers may need more help to realize their value.
Why AI changes the business problem, not just coding speed
A common narrow view is: AI writes code, therefore a company needs fewer engineers. The speaker points to a different question: who will make the software useful for this particular customer?
If a platform is simple and has a clear, fixed workflow, the customer may be able to adopt it with little outside implementation. If the platform can support many workflows, the customer must make more decisions. They may need to connect the platform to their operating context and build the result they actually want.
Teaching example (not a video case)
Imagine a company sells an AI platform that can support many internal workflows. One customer might want an insurance-related process; another might want a legal workflow. Giving both customers the same powerful platform does not decide:
- which workflow is worth building first;
- what customer data and business rules it should use; or
- how people will use the result in daily work.
The platform makes those outcomes possible. It does not automatically deliver them. An FDE-like function can help close that gap by understanding the customer's context and building a solution on the shared platform.
This example illustrates the reasoning. It does not claim that the speaker described a particular insurance or legal implementation.
Expansion makes leaving implementation to the customer riskier
The speaker says it becomes harder to leave product success entirely to customer implementation as a company expands. The three directions he names are easy to mix up:
| Direction | Meaning |
|---|---|
| Upmarket | Sell to larger or more demanding customers. |
| Horizontal expansion | Reach more kinds of customers or use cases across a market. |
| Vertical expansion | Go deeper into a particular industry or domain. |
Each direction can increase variation in customer needs. A larger customer may have a more complex environment. A new industry may use different language, rules, and workflows. A broader set of use cases may expose more ways to configure the product incorrectly or leave useful capabilities unused.
The speaker's warning is therefore about ownership of success. A vendor cannot always assume that a customer will understand a powerful platform well enough to implement it alone. In some cases, sending people who know both the platform and the customer's problem becomes more valuable.
This does not replace the FDE adoption test
Earlier in the talk, FDE was justified by two conditions working together:
- The company sells a technically complex, customizable product to a buyer or customer organization that cannot readily implement it alone.
- The company has, or is willing to invest in, a shared platform with reusable primitives.
AI may make the first condition more common by increasing customization and the resulting implementation gap. It does not establish either condition for every product. A fashionable AI label is not enough. Nor is the fact that an engineer could help a customer.
The second condition still matters because FDE should build customer outcomes from shared foundations rather than create a separate software project for every account. Otherwise, the company risks becoming the kind of custom development shop the earlier chapters warn against.
A useful way to reason about an AI product
When evaluating whether customer-facing engineering is warranted, ask:
- How much customer-specific implementation does realizing value require?
- Can the target customer understand and perform that work on its own?
- Are the solutions built from reusable platform capabilities, rather than from-scratch work per customer?
If the first answer is “a lot,” the second is “not reliably,” and the third is “yes,” the speaker's FDE logic becomes more relevant. If the product is easily adopted by its intended users, or there is no reusable platform beneath the work, an FDE team is not automatically the answer.
What the source does and does not establish
The speaker presents this AI-era account as his personal hypothesis. He broadly suggests that platforms are becoming increasingly agentic and customizable, and that this can widen the customization problem. The segment supports a reasoning framework, not a measured forecast about all AI products or a proof that FDE causes successful adoption.
The chapter ends as the lecture transitions into audience questions and applause. No additional substantive Q&A answer belongs to this segment.