When sales emails and lead information are spread across mailboxes, CRM systems, and internal files, sales representatives must repeatedly review the details, assess priorities, check past interaction history, and then decide on the next action. As the number of opportunities grows, this verification work alone can consume a significant amount of time.

In addition, when it is unclear who should follow up with which lead and when, responses are more likely to be delayed or missed. This is especially true when a large volume of inquiries arrives within a short period, as important emails can easily get buried among other messages.

When applying AI to sales operations, it is also necessary to consider not only efficiency gains but also risks such as misclassification and inappropriate reply suggestions. For that reason, the system needed to be designed not as a simple automated sending tool, but as a mechanism that supports human judgment.

The goal of this project was to streamline preparatory work from email review through follow-up preparation, enabling sales representatives to focus on making final decisions.

Specifically, the initiative aimed to standardize daily sales operations and improve response speed by summarizing key points from incoming emails, classifying leads, suggesting priority levels, drafting replies, and organizing candidate updates for CRM entry.

Another important focus was validating an operating model in which AI does not make final decisions autonomously, but instead presents recommendations in a format that is easy for people to review.

SMILE was responsible for analyzing the current state of sales email handling; organizing workflows involving Gmail, Odoo CRM, and Google Drive; designing classification rules and AI agent prompts; implementing summarization, priority assessment, and draft generation features; building a review dashboard; and carrying out validation and improvement.

We also designed the workflow so that AI outputs are not used directly for external messages or important updates. Instead, a review step ensures that a person can always verify them, making the system easier to deploy in real operations.

The focus of this PoC was not simply to add AI, but to clarify where AI could be most effective in the sales process and where human judgment should remain in place.

First, instead of processing incoming emails as they are, we organize them into summaries that allow sales representatives to understand the situation at a glance, taking into account the subject line, body text, sender, past interactions, and existing CRM information. Next, based on the inquiry details and level of interest, the system suggests lead classifications and priority levels, making it easier to determine the order of follow-up actions.

In addition, for creating draft replies and suggested CRM updates, we adopted a structure that combines AI with rule-based conditional logic rather than relying on AI alone. This streamlines routine tasks while preserving human review for important decisions and external communications.

One of the challenges was that sales emails vary significantly in content, from brief inquiries and requests involving multiple requirements to replies that rely on the context of previous conversations. Simple keyword-based rules alone could not deliver consistent classification accuracy, while relying solely on an LLM raised concerns about weakly supported judgments and excessive automation.

For this reason, SMILE uses both rule-based logic and AI for classification and priority assessment. For example, rules are used to reinforce checks such as the type of inquiry, whether the sender is an existing customer, and whether the message contains expressions related to response deadlines, while AI is applied to areas where contextual understanding is essential, such as summarization and draft creation.

Rather than automating every action, we made it possible for a person to confirm email sends and important CRM updates on a review screen. This helps balance efficiency with operational safety.

This PoC makes it easier for sales representatives to complete preparatory tasks in less time, such as understanding incoming email content, assessing priorities, preparing replies, and organizing candidate updates for CRM. It is especially expected to help prevent important leads from being overlooked, standardize follow-up criteria, and make initial responses more consistent.

By clearly positioning AI as a tool that supports human judgment, the system is designed to improve efficiency while helping maintain the quality of external communications and operational confidence.

When using AI for sales support, it is important to design more than just a connection to a model. The solution must also incorporate business process understanding, rule definition, review workflows, and continuous improvement cycles. At SMILE, the Japan and Vietnam teams divide responsibilities and operate a structure that allows repeated review of business workflows, output quality checks, prompt tuning, and incorporation of validation results.

This approach makes it easier to identify improvement points in a form close to real-world operations, even at the PoC stage, while also building a stronger basis for considering full-scale deployment in the future.

When improving the efficiency of sales email handling and CRM operations, the priority is not to hand everything over to AI, but to organize information in a way that makes it easier for people to make decisions and take the next step. This PoC demonstrates how an AI agent can be embedded as a support tool for sales activities, helping improve initial sales response by providing integrated support for email summarization, lead classification, priority assessment, and reply drafting.

For companies considering the use of AI in sales and marketing, a phased implementation design that assumes human review can also be a practical approach.