Delivering an AI System to Users

Forward Deployed Engineering, or FDE, is the way this course describes building AI systems alongside a customer. The work is not a single build step. An FDE engagement is organized over time, and this article introduces the frame that most of that organization depends on. The frame has three phases, taken in order: early scoping, validation, and delivery.

Many people picture an AI project as a line with two parts. First the team builds the system, then it checks whether the system works. In that picture, testing is just the tail end of building. The check happens after delivery, it is often framed as a review, and its results are sometimes filed away without action. Teams that hold this picture tend to treat validation as a sub-phase of delivery, or as a post-launch review. The rest of this article shows why that picture causes trouble, and what the three-phase frame does differently.

An FDE engagement, in this frame, runs in a fixed order. Scoping comes first, validation second, and delivery last. Each phase has a distinct purpose, and each one feeds the next. The order is not arbitrary. The later phases depend on decisions made in the earlier ones, so the work of building a deliverable does not begin until the impact area has been chosen and tested for feasibility and potential impact.

Each phase answers a different question. Scoping asks where the work should focus. Validation asks whether that focus is worth building around. Delivery asks how to put a working result into the hands of the people who will use it. Keeping these three questions in mind makes it easier to tell which phase a piece of work belongs to.

The diagram shows the order and the hand-off between phases. Scoping produces an understanding of the problem space, and validation takes the impact area that scoping identified and tests it. Delivery starts only after that test, so the team enters delivery already holding a decision that the impact area is worth pursuing.

Scoping is the initial phase. Its job is to understand where an AI system can have the biggest impact. According to the source, scoping is done by a small team working closely with the customer, so the phase is collaborative from the start. The customer's problems and context shape the search for an impact area from the first day.

Validation is the phase where the feasibility and potential impact of the project are tested. The source describes it as choosing the impact area before committing to the build. This is the discipline of deciding where to focus before spending effort on construction. Validation is also the step that determines whether delivery should be attempted at all, which makes it a decision point rather than a stage of production work.

Delivery is the final phase, where the project is carried through to impact. The source defines it as the point where the team has an impact to show for the work. Because delivery is preceded by validation, the team does not enter delivery with an open question about whether the work matters. It enters delivery already holding a decision that the impact area is worth pursuing, and its job is to carry that validated project through.

This dependency is why the order matters. If validation were skipped or moved after delivery, delivery would be carrying a project whose worth had never been tested, and the decision that validation is meant to inform would arrive too late to change anything.

The source makes validation a separate phase with its own deliverables and decision gate. This matters because validation has a different job from delivery and from a post-launch review. The comparison below places the three side by side. Delivery builds toward impact. A post-launch review looks back at a system that already exists. Validation sits between them and makes a decision that the other two do not make.

Read the columns from left to right and the difference is clear. The validation column is the only one that ends in a decision about whether to proceed. The delivery column begins from that decision. The post-launch review column begins after the work has been finished, which is why its results can be filed away without changing what was built.

Applying the frame means asking which question a piece of work is answering. Work that identifies where an AI system could have the biggest impact belongs to scoping. Work that tests the feasibility and potential impact of a chosen impact area belongs to validation. Work that carries a validated project through to a working result in users' hands belongs to delivery.