In the 2026 semi-annual performance briefing materials, iFlytek listed the National Energy Group's recruitment of intelligent agents as an FDE case, and said that the projects have been running stably. The same page also introduces the general FDE path of the expert feedback model, without disclosing how it is implemented within TNB. The complete review sample of the materials has not been made public. This article does not regard the publicity effect as an independent verification result, nor does it regard subsequent construction as completed.
First determine what the expert has changed
The following is an independent analysis surrounding the procurement review task and is not an internal implementation record of the project. An AI judgment may be corrected for completely different reasons: the number in the scan was read incorrectly, the procurement requirements were not correctly mapped, or the attachments provided were insufficient. If these changes are saved as a sample of "the original answer is wrong, the final answer is correct", the same mistake may still be made next time.
For example, if the date recognition is wrong, you should go back to the text and table reading link; if the product range covered by the proof material is inconsistent with the current procurement object, it involves condition matching. Even if the two produce the same modification results, it is not appropriate to use the same correction method. Clearer records should include which judgment the expert changed and where the information needed to make the change came from.
Insufficient materials must be left to be confirmed.
Another situation that is easily confused is that no evidence has been found and that it has been discovered that it does not meet the requirements. Suppose a page of a certain material is missing, the system cannot fill in the unknown into a negative conclusion based solely on the current attachment. Business personnel first need to know what is missing, whether it can be supplemented, and what judgments need to be redone after supplementing. These efforts are distinctly different from directly modifying a conclusion.
Therefore, review aids can help locate the questions to be asked, but they should not allow a smooth explanation to obscure gaps in the information. If the expert changes his judgment based on the newly received material, the new basis needs to be reflected in the record; if the original material is sufficient but the model understanding is wrong, another type of reason should be retained. Mixing the two may mistakenly record changes brought about by new information as improvements in model capabilities.
One revision may not be suitable for all procurement matters
Review requirements often come with applicable objects and conditions. Certification requirements for one item may not apply to another category or procurement document. When collating expert feedback, it’s not the last answer in isolation that’s worth retaining, but the conditions under which it holds true. Replicating experience without conditions may allow the system to learn overly broad judgments.
For the implementation team, they can first find a group of similar items that have been reviewed and observe whether the errors occur in data reading, object correspondence or conditional judgment. Only in this way can we decide whether to change the reading method, supplement the business description, or adjust the model processing link. The classification here is the editorial department’s analytical framework. It does not mean that iFlytek has adopted the same classification, nor can it replace the formal review responsibilities of procurement professionals.
The narrower service scope of a small team can help business departments sort out existing review issues and modification reasons, and clearly explain missing materials and applicable conditions that are easy to communicate repeatedly. Whether access to data is allowed and whether a usable sample can be formed needs to be determined by the client. Value should be discussed in terms of whether certain types of problems are easier to explain and deal with, and avoid using publicity numbers for large projects as quotations or performance promises.
Extended reading: Research on AI classification of Huadian material catalog, Extract of contract elements and original text location, How to fix business examples when comparing models.