HappyRobot's Encompass customer story introduces that FDE engineer Federico and the accessories service team worked together to improve voice model recognition and solve the problem of mixed hearing of characters such as B and 8. The case stated that the process was expanded from Midea-related business pilot to LG. The customer here is Encompass, and it cannot be inferred from the brand name that HappyRobot is delivered directly to the brand manufacturer; the results in the article belong to the supplier's statement.
The model number is a string of identifications. You cannot just understand the general meaning.
The following is ChinaOPC’s independent analysis of the parts query task, and is not an undisclosed design of the customer’s system. In ordinary conversations, similar expressions may refer to the same meaning; but the model number may differ by only one letter, corresponding to different products. If the voice agent uses language conventions to fill in the "most likely" characters, the query results may look reasonable, but they may have deviated from the caller's device.
Small groups can start by keeping the original expressions they heard separate from the edited characters. For uncertain locations, ask the user to repeat it character by character, or use text, pictures, etc. to supplement it when the corresponding conditions are met. Supplementary channels should be designed according to actual business conditions, and customers cannot be promised that the current system supports it just because a method is theoretically feasible.
Candidate record found, accessory matching has not been completed yet
When multiple similar models appear in the catalog, you can first explain that there are multiple candidates, and then check the necessary information to distinguish them. Automatically selecting a model simply because it is more common can allow recognition errors to propagate to inventory, pricing, or purchase portals. The task of AI is not to come up with an answer as quickly as possible, but to help narrow the scope of uncertainty.
Whether the accessories are compatible must be based on the catalog and corresponding information. The model cannot be used to infer the applicable relationship based on model similarity. If there is no corresponding information in the existing directory, it can be clearly stated that it cannot be confirmed and handed over to the appropriate person for processing. The applicable relationship of accessories needs to be checked back in the catalog, and it cannot be directly guaranteed based on the generated answers.
If the query fails, you need to know which step failed.
A voice request did not get a result. The reason may be that the model number was not heard clearly, there is no record in the directory, the compatibility relationship is incomplete, or the inventory interface has not yet returned. By recording all situations as a failure of the robot to answer, it was difficult for the team to determine whether the next step should be to improve the recognition, supplement the information, or handle the system connection.
The different reasons mentioned above can be recorded for a small number of samples approved for use and then verified by business personnel. Samples should cover long and short models, similar alphanumeric characters, pauses, and repeated expressions, rather than just selecting recordings that are clear, quiet, and easily identifiable. When comparing, first check whether the characters are correctly retained, and then check whether the correct record is found later to avoid smooth speech covering upstream errors.
handover to a person is sometimes a question of correct retention
When there is insufficient information, handover to a person can be a necessary treatment, rather than a shortcoming that needs to be eliminated statistically. When handing over, the confirmed characters, still uncertain positions and the results that have been checked should be explained to reduce the user's repetition from scratch. Different types of transfers need to be observed separately, and whether a process is suitable for continued automation cannot be judged based on the overall proportion alone.
The original case also lists return processing as a direction that is still under evaluation and cannot be written as being automated. For a small team to study this type of project, you can first select a product range with relatively complete information, let business personnel check the identification and matching paths, and then discuss expanding to more brands. The expansion of brand scope may bring about new numbering habits and catalog differences, rather than simply copying a prompt word.
This case shows FDE's involvement in improving specific business details. This article does not use call volume, resolution rates or customer sentiment figures as proof of general benefits, nor does it identify large customers or service providers with unverified numbers as OPCs. What we can learn from is to break down a character error that is easily overlooked into problems that can be gradually located.
Extended reading: Equipment manual index service, How to handover AI customer support to a person and How to sample customer service results after going online to compare the different roles of data, handover and daily review.