Smart Email intake using AI models to analyze and route

Smart Email intake using AI models to analyze and route

Recently we have had five different customer engaged at our Microsoft Technology Center targeting similar challenges with a common pattern to enable smart Email intake to analyze, route and response. During a sequence workshops staring with envisioning to identify and validate the business value, a architecture design workshop to design the solution and establish a sandbox environment for the follow-up Hackathon to implement the solution within the customer environment.

Customer challenges

Based on the customer envisioning we can summarize their challenges:

• We’re getting a lot unstructured Emails from our customers and partners triggering a business process to manage existing contracts or services.
• We want to analyze Email automatically and categorize the Email based on the product and customer intent.
• We want to enrich the Email with business context based on the target product and customer intent to increase our agent effectiveness and efficiency.
• We want to assign the analyzed Email to the best skilled and available agent.
  We want to reduce chit chats. A majority of those
• Emails are generating replay or forward chit chats between sender, receivers and fulfiller.
  We need to support multi-languages, as customer or partner Emails are in English, French, Italian or German.
• We want to reduce the TCO of the target solution and enable business analyst to extend and drive on going changes.

Art-of-the possible discovery

As most customers already have an existing Microsoft 365 tenant, we can focus the art of possible discovery to the included Power Platform cloud services and cross-check if there other capabilities to reduce implementation time and TCO. For example:

• Including existing customer application and AI models via API management.
• Use of Power Platform and/or Dynamics 365 Premium services where it makes sense based on business value.

During the discussion we break down Smart Email intake process into six functional domains with different implementation and extension options. With this we can chain the functional domains based on customers objectives and requirement.

Make or buy decision

My principle is use make if we have a competitive advantage, otherwise buy. For example the smart part (AI based) of the solution is make (bring you own model, train model with your data), rest make or buy. With that in mind we used a mix of Dynamics 365 Customer Services and Power Platform functionality. To reduce cost to buy and maintain, and accelerate implementation.

Make (Power Apps) or Buy (Dynamics 365)
20% cost saving using buy

A quick business value justification for an assumed scenario with 500'000 Mails per year, 160 Agents managing responses using Dynamics 365 Team members, 10 Business Analysts managing process & models using Dynamics 365 Customer Services, with 3 Power AI Builder models used, turns out 20% cost benefit against Power Platform Premium licensing.

• Power Apps Pricing - Power Apps Pricing | Microsoft Power Platform
• Dynamics 365 Pricing - Pricing | Microsoft Dynamics 365

Make it smart with AI models

Key component are the AI based models to analyze customer intent and context. Power AI Builder is a perfect starting point and we used the following prebuilt and custom models.

Language detection prebuilt model

The language detection prebuilt model identifies the predominant language of a text document. The model analyzes the text and returns the detected language and a confidence score from 0 through 1.

We used it to identify the language of the Email content to steer the after processing either translate it to a supported language or use language specific pre trained models for category classification or entity extraction.

Text translation prebuilt model

The text translation prebuilt model translates your text data in real time across more than 60 languages. The text translation model can also detect the language of the text data you want to translate.

We used it to translate Emails into supported languages.

Sentiment analysis prebuilt model

The sentiment analysis prebuilt model detects positive or negative sentiment in text data. The scores and labels can be positive, negative, or neutral.

We used it for negative customer intent to escalate it.

Category Classification custom model

Category classification is one of the fundamental natural language processing (NLP) challenges. With category classification, you can identify text entries with tags.

We used it to defined the following tags based on product and customer intent

• Pension fund; Withdrawal; Admission; Wage change;
• Vehicle Insurance; Deductible;

Entity Extraction custom model

AI Builder entity extraction models recognize specific data in text that you target based on your business needs. The model identifies key elements in the text and then classifies them into predefined categories. This can help you transform unstructured data into structured data that's machine-readable.

We used it to identify Vehicle Number Plate, Car insurance, AHV (Swiss Social Number).

Unified Routing engine

Dynamics 365 Unified routing is an intelligent, scalable, and enterprise-grade routing and assignment capability that can direct the incoming work item to the best-suited queue and agent by adhering to work item requirements and matching them with the agent’s capabilities.

Unified routing has two main stages: classification and assignment.

In the classification stage, rules and machine learning (ML) models can be used to add information on the work item, which can be used to find the best-suited agent.

We used it to route cases using Language and Category Classification model, Category Tags for products and customer intent to Agent skills like English, German, Car Insurance, Pension Funds, Withdraw, etc.

Make it real Hackathon

We build this during a 3 to 4 day Hackathon with a customer team with 3-6 (business stakeholder, business analyst and IT/developer) members and deployed the MVP to production.

If you want to discover more, check out the set by step guide providing a starting point to do it by your own.

Sprints to make it real

1. Create a solution “Case Management”
2. Create a Automatically create Case records from Email ruleset
3. Create instant flow “TranslateCasesToTargetLanguage”
4. AI Builder category classification custom model
5. Extend Automatically create Case flow with AI Builder models
6. Configure Unified Routing based on Category Tags to Agents

Conclusion

Within 3 to 5 workshop days we were able to analyze the business value, design a target solution architecture and implement a MVP addressing customer key challenges. Overall it was a great work experience together to make it reality with a mixed customer stakeholder team. If you’ve similar challenges reach out to us for a workshop at the Microsoft Innovation Hub.

• Microsoft Innovation Hub
• Check out our Workshop Offerings and contact your Microsoft Account Manager