According to the 2025 annual report of Dipu Technology, the company will build its own AI digital employee platform at the end of 2025. FDE engineers will collaborate with four types of AI employees, namely solution, data governance, code and project management, to assist in demand analysis, governance and development. The report clearly indicates that FDE is Forward Deployed Engineer. This is a public statement of how the company works internally and does not mean that the four types of tools have independently completed all customer projects.
Just because the demand report is written completely does not mean that the problem has been decided.
The following is ChinaOPC’s independent analysis of the collaboration process and is not Dipu’s internal operations manual. When organizing interviews into a requirements report, AI easily puts content with different tones into the same conclusion: the arrangements that have been confirmed by the customer, the ideas put forward by a certain user, and the feasible methods speculated by the engineering staff may all be written as the capabilities that the future system will have. The smoother the report, the easier it is for follow-up personnel to miss these differences.
You can first state the source and current status of each item of content. Matters that have been jointly confirmed can be entered into subsequent design; matters that still have disagreements remain as questions and objects to be answered; suggestions that are only for discussion should not automatically become customer commitments. Before the next job begins, the person taking over needs to know whether they are receiving facts, options, or unresolved questions.
Data governance needs to receive business implications
The same word "order" may refer to inquiry records, purchasing records or shipping arrangements in different processes. If the requirements report does not explain it, even if the governance tool organizes the fields neatly, it may still put different business objects together. The team should first find someone familiar with the business to explain what the data represents, and then decide which ones can be connected and which ones need to be kept separately.
The identified data gaps can be passed on to development tasks. For example, if a certain automatic processing requires a field that currently has no source, it cannot be assumed that the field already exists because the model is good at completing it. The examples here are used to explain work dependencies and do not mean that the annual report discloses real problems with an order system.
When generating development tasks, keep why you are doing this
When Code Assistant receives a task with only the final feature name, it's often difficult to understand why a certain approach was previously excluded. A small team can provide the purpose of the task, existing conditions, and unresolved problems at the same time, allowing developers to judge which part can be done now. AI can propose implementation plans, but it cannot quietly turn business choices that have not been made by the upstream into default values.
Several assistants using the same misunderstanding may produce results that are consistent with each other. Just because the scheme, data structures, and code fit together does not automatically prove that they fit the client's work. Rather than just checking whether the text is consistent, it is better to ask the actual user to look back at a specific task: whether the information he provided, the expected actions, and the system-generated results still express the same thing.
The project summary should clearly state the completed and to-do items respectively.
A project management assistant can help summarize the material, but the summary should distinguish between completed functionality, changes being discussed, and tasks still affected by external conditions. Internal organization is completed, but it cannot be written that the customer has accepted it; the demonstration can be run, but it cannot be written directly that it has entered daily business. Keeping these distinctions helps the next collaborator understand the true progress.
For small teams, it is not necessary to set up four sets of independent assistants first. You can start with an approved task and observe which information is most often lost between reporting, data collection and development, and then decide whether specialized tools are needed. When the number of tools increases, it also increases the burden of verification and collaboration, and the supplier's organizational style cannot be directly regarded as the standard configuration of small teams.
Dipu disclosed that it provided research samples on the collaboration between FDE and internal AI assistants. Future platform construction and operation improvement are still expressed at different levels. This article does not use the scale, financial or capability claims in the report to prove customer results, nor does it identify this large enterprise as an OPC. The lesson is to allow each handoff to preserve the true judgment of the previous work.
Extended reading: How to keep unconfirmed items in meeting minutes and How to check the context of other people’s SOPs, check collaboration information with actual working conditions.