Implemented a new function, "AI Multi-Stage Inference," that utilizes multiple generative AIs to enhance processing.

AmiVoice Communication Suite now features an "AI multi-stage inference" function that uses multiple generation AIs depending on the content, purpose, and use of the call, and divides the processing load in stages, thereby improving accuracy, efficiency, and cost optimization.

"AmiVoice Communication Suite" has the No. 1 domestic market share^{On September 1, XNUMX,} It is an AI speech recognition solution for contact centers. In August 2024, we began offering "AOI LLM for AmiVoice Communication Suite (AOI LLM)," an optional generation AI that can be used securely in a local environment, and in February 2025, we released an external generation AI integration function.

This function was developed in response to the needs of companies using "AmiVoice Communication Suite" and "AOI LLM" who wanted to "introduce a system that automatically switches to the most appropriate prompt depending on the content of the call." In calls at contact centers, the necessary summarization and information extraction points vary depending on the inquiry content and purpose, such as "I would like to change the delivery date" or "The staff member's response was poor." Therefore, it was difficult to respond adequately with processing using a single conventional generation AI, and there were issues with accuracy and operational efficiency.

In this context, AmiVoice Communication Suite has implemented a new function called "AI Multi-Stage Inference," which enables more accurate and flexible processing by combining and gradually utilizing multiple generation AIs according to the content, purpose, and use of the call.
The type of processing to be performed at each stage of "AI multi-stage inference" can be flexibly configured for each customer. By selecting and switching between the optimal generation AI and prompt depending on the purpose and application, the output quality of the generation AI can be improved, enabling more advanced and flexible processing. This function makes it possible to organize complex topics and process multifaceted information, which were difficult to handle with conventional processing using a single generation AI. In addition, by switching between low-cost and high-performance generation AI depending on the difficulty of the task, it also contributes to cost optimization.
This function also allows you to set the AI ​​and prompts to be used according to keywords included in the call. Furthermore, if you are using CTI integration, you can automatically change the AI ​​and prompts to be used according to the skills defined in the system.

■Examples of using "AI multi-stage inference"

① Classify the inquiry content and summarize it according to the call reason

The first-stage generation AI analyzes the content of the call and categorizes it into categories such as "address change," "lost card," and "confirmation of usage details." Based on the classification results, the second-stage generation AI performs summarization using prompts optimized for each category. By automating summarization according to the call reason, it significantly reduces the workload of operators and contributes to shortening post-processing time.

② Utilizing AI generation based on call attribute information through CTI integration

For inbound calls, the generation AI summarizes and analyzes the content of the call, and for outbound calls, it proposes scripts, making it possible to utilize AI according to the attributes of the call through CTI integration.In addition, by setting the processing content of the generation AI for each skill defined in the system, such as "updating contract information" and "handling cancellations," more flexible and accurate support can be provided, achieving high-quality responses that are not dependent on operator skill.

3) Combining local and cloud-based generation AI to achieve secure operation

In the first stage, AOI LLM, a secure generation AI that can be used in a local environment, is used to mask personal information such as names, addresses, and phone numbers. Next, the masked text data is passed to a cloud-based generation AI (such as ChatGPT), which performs advanced processing such as summarization, classification, and analysis. By passing only data that does not contain personal information to the cloud-based generation AI, it can be used safely even in industries with strict security policies.

④ Utilizing multiple generation AI to suppress hallucination

Combining different generation AIs can also be expected to suppress hallucinations that are dependent on specific models. For example, by re-presenting the summary results output by the first-stage generation AI along with the call data to the second-stage generation AI, the consistency and validity of the content can be verified and supplemented. By utilizing multiple generation AIs, it is possible to prevent the generation of incorrect information and obtain more reliable output.

■ Features of "AI multi-stage inference"

1. Combining multiple generative AIs for different purposes to improve accuracy and efficiency

By combining multiple generation AIs according to the content of the call, purpose, and application, more accurate and flexible processing can be achieved, helping to improve the accuracy of output results from generation AI and significantly improve work efficiency. The generation AI, prompts, and processing content to be used can be configured for each customer.

2. Optimize costs by selecting generation AI according to task difficulty

By using either low-cost or high-performance generation AI depending on the difficulty of the processing to be performed by the generation AI, operational costs can be optimized, achieving both operational efficiency and cost performance.

3. Versatility that allows flexible use in a variety of tasks

It can be used for a variety of purposes in contact centers, including classifying and analyzing call content, summarizing, extracting information, arranging text, masking personal information, and hallucination checks.It can be used in a wide range of industries, from consumer inquiry desks to insurance and financial contact centers that handle personal information.

*1 Source: ITR "ITR Market View: Image and Voice Recognition Market 2024" Voice recognition market - Sales share by vendor for contact center operations (FY2024 forecast)