Skip to main content

Next Best Message, Search & Autocomplete

The Next Best Message, Search & Autocomplete features in Deepdesk's Agent Assist provide customer service agents with intelligent response suggestions based on historical conversation data. By leveraging AI-powered models, agents can quickly find and insert the most relevant responses during live interactions, enhancing efficiency and customer satisfaction.

How Next Best Message, Search & Autocomplete Works​

The following diagram illustrates the end-to-end process of setting up and using Next Best Message, Search & Autocomplete:

Process Overview​

1. Data Collection​

Deepdesk collects historic chat and messaging conversation data from your CX platform. This data serves as the foundation for training the response recommendation model.

2. Response Identification​

The system analyzes the collected conversations to identify frequently occurring responses. These common responses are automatically extracted and stored in the Recommendation Studio for review.

3. Model Training​

Using the identified responses and conversation patterns, Deepdesk trains a response recommendation AI model. This model learns to predict the most appropriate response based on the current conversation context.

4. Content Redaction​

Before deployment, administrators review the identified responses in the Recommendation Studio:

  • Review automatically identified responses
  • Edit or redact sensitive information
  • Remove inappropriate or outdated responses
  • Approve responses for use in production

5. Model Deployment​

Once responses have been reviewed and approved in the Recommendation Studio, the trained model is deployed to production. The model is now ready to provide real-time suggestions to agents.

6. Real-Time Assistance​

During live conversations, the system provides agents with intelligent response suggestions:

  • Agent receives an incoming customer message
  • AI analyzes the conversation context in real-time
  • Model predicts the best matching responses
  • Autocomplete suggestions are displayed to the agent
  • Agent can select, modify, or ignore suggestions
  • Selected response is sent to the customer

Model Training Pipelines​

When a new account is provisioned in Deepdesk, dedicated model training pipelines are created for each profile in that account. These pipelines are scheduled to run automatically every other week. During each run, the pipelines perform the model training and deployment steps described above—collecting new data, updating response clusters and templates, retraining models, and deploying the latest versions to production. This ensures that response and URL suggestions remain accurate and up-to-date as customer interactions evolve.

Content Types​

Frequently occurring text responses found in historic data are considered the default content type. The default training method uses a clustering algorithm to identify these responses from large volumes of conversation data. Alternatively, another training method uses a matching algorithm, which attempts to match agent responses to pre-existing template responses. When a match is found, the system can recommend these templates to agents during live conversations. To support this, each text response is assigned a content type: either default (for clustered responses) or template (for matched templates).

Template Variables​

In text suggestions you can use template variables that will be filled in according to the context at runtime. The following variables are available:

  • {visitor_name} — name of the current customer
  • {agent_name} — name of the current agent
  • {part_of_day} — part of day based on the current time (e.g., morning, afternoon)
  • {day_of_week} — current day of the week (e.g., Monday)

Example

Good {part_of_day}, {visitor_name}! My name is {agent_name}. Happy {day_of_week}! How can I help you today?

URL Suggestions​

Deepdesk also supports intelligent URL suggestions using a similar process. A separate AI model is trained specifically on URLs found in agent responses, such as links to FAQs or knowledge base articles. During live conversations, this model analyzes the conversation context and suggests the most relevant URLs to agents, making it easy to reference helpful resources and provide customers with direct answers. The workflow for URL suggestions mirrors the steps above: data collection, response identification (focusing on URLs), model training, content review, deployment, and real-time assistance with URL recommendations.

Technical Note

Currently, there is some technical debt in how content types are managed. URLs, which are suggested to agents, should ideally be treated as text suggestions with a content type of URL. However, this refactor has not yet been implemented in the system.