AutoCompose Product Guide

Learn more about the features and insights of AutoCompose

Getting Started

This page provides an overview of the features and functionalities in AutoCompose.

After AutoCompose is integrated into your applications, you can use its features to scale up your agent responses.

The following UI descriptions are examples of AutoCompose Integrations with LivePerson and Salesforce. API-based integrations do not include custom UIs.

Suggestions

AutoCompose supports agents throughout the conversation with both complete response suggestions before they type and suggestions while typing to complete their sentence.

The machine learning models powering AutoCompose suggestions use the entire conversation context (not just the last few responses) and personal agent response history to predict the most likely next agent message or phrase in the conversation.

Response Library

AutoCompose suggests responses from a library curated from a wide range of domain-specific conversation topics. The response library is a combination of three lists:

1. Global response list: Messages created and maintained by program administrators available to a designated full agent population.
2. Custom response list: Messages created and maintained directly in AutoCompose by individual agents; only available to the agent that created the message.
3. Organically growing response list: Messages automatically created by ASAPP for each agent based on their most commonly used messages that do not already exist in the global response list or the agent’s curated custom response list.

Agents use custom responses to make their favorite messages readily available for sending quickly: well-honed explanations for difficult processes and concepts, discovery questions, personal anecdotes, and greetings and farewells infused with their personal style.  Agents often curate their custom responses based on global responses, with a bit of their own personal touch.

Agent Interface

AutoCompose provides three complete response suggestions in the drawer above the composer both before typing begins and in the early stages of message composition; phrase completion suggestions are made directly in-line as more of a sentence is typed.

Agents can also search for a response in two places:

1. Composer: As agents type, they can choose to search for their typed text in the global response list to see the full list of related messages with that term. By typing / in the empty composer, agents can also browse their custom response library by using either the message text or title of the custom response as a search term.
   
2. Response panel: In the response panel, agents can browse both the global and custom response lists, either using a folder hierarchy or with the provided search field.
   

The organically growing response list is not available for agents to browse - responses from this list only appear in suggestions.  Agents are encouraged to add these frequently used responses to their custom response list.

Autocomplete

Once the agent begins typing, AutoCompose provides two forms of autocomplete suggestions at different stages in the message composition:

• As the agent begins typing a message, complete response suggestions are available. At this point, the agent is in the early stages of composing their response and several potential complete response options are relevant.
• After several words are typed, a high-confidence phrase completion can be recommended in-line to help agents finish their already well-formed thought.

Phrase completions are generated from common, high-frequency phrases used in each implementation’s production conversations. AutoCompose only makes phrase suggestions when a sufficiently high-confidence phrase is available and only uses language found in the global and custom response library.

Templated Responses

AutoCompose can dynamically insert metadata into designated templated responses in the global response list.

For example, a customer’s first name can be automatically populated into this templated response: “Hi {name}, how can I help you today?”.

By default, AutoCompose supports inserting customer first name, agent first name and the customer’s time of day (morning, afternoon, evening) into templated responses. Time of day can be set to a single zone or be dynamically determined for each conversation.

AutoCompose also supports inserting custom conversation-specific metadata passed to ASAPP. For more information on custom inserts, reach out to your ASAPP account team.

If the needed metadata is unavailable for a particular templated response (e.g. there is no customer name available), that response will not be suggested by AutoCompose.

Fluency and Profanity

Fluency Boosting

AutoCompose automatically corrects commonly misspelled words once the space bar is pressed following a given word.

Corrections are underlined in the composer for agent-awareness and can be undone if needed by hovering over the corrected word.

Profanity Blocking

AutoCompose checks for profanity in messages when the agent attempts to send the message. If any terms match ASAPP’s profanity criteria, the message will not be sent and the agent will be informed.

Customization

Suggestion Model

The AutoCompose suggestion model functions as a custom recommendation service for each agent. The model references the global response list, a library of custom responses created by each agent, and also learns from each agent’s unique production message-sending behaviors to surface the best responses.

Global Response List

Prior to deployment, ASAPP can generate a domain-relevant global response list using representative historical conversations as an input. This is a highly recommended customization to ensure agents receive useful, relevant suggestions as early as possible.

If historical conversation data is unavailable prior to deploying AutoCompose, production conversations after deployment can be used to adapt the response list at a later date.

Option	Description	Requirements
Model-generated	Fully-custom global response list that extracts relevant terminology and sentences from real conversations	200,000 historical transcripts to enable prior to implementation

For more information on sending historical transcript files to ASAPP, see Transmitting Data to ASAPP.

Queue/Skill Response List Filtering

AutoCompose can filter the global response list by agent queue/skill for a given conversation. For example, a subset of responses appropriate only for sales conversations can be labeled to be removed from technical troubleshooting conversations.

Responses are labeled with applicable queue(s)/skill(s) and are unavailable for suggestion if their labels do not match the conversation.

Option	Description	Requirements
Global Response List with filter attributes	Global responses are labeled with optional attributes for skills for which they are exclusively appropriate.	Review and labeling of specific responses

For technical information about implementing this service, refer to the deployment guides for AutoCompose:

• AutoCompose API
• AutoCompose for LivePerson
• AutoCompose for Salesforce

Use Cases

For Improved Agent Productivity

Challenge

Agents spend a lot of time manually crafting responses to similar customer problems.  Using scripts can make the conversation sound robotic, so agents who do use canned responses spend a lot of time adjusting the language to sound more like them or use them too rarely to impact their response time or ability to handle multiple conversations concurrently.

Average response crafting time and messaging concurrency, even with canned response library usage, remains high for most digital messaging programs.

Using AutoCompose

AutoCompose drastically reduces crafting time by not only serving up response suggestions from a much more exhaustive set of responses, but it also addresses the problem of canned responses sounding overly generic by empowering agents to craft messaging in their personal style.

This is why adoption, and therefore efficiency gains overall, are so impressive when AutoCompose is deployed.

For CX Quality and Consistency

Challenge

Agents have a lot of information to learn to become domain experts and are often handling issues with which they have limited experience or that they have trouble recalling the best way to handle.

Many companies use a variety of resources that agents have to search through to find answers on how to best handle a customer’s problem.  This can be difficult for an agent to juggle while in a live conversation, especially if they are unsure where to begin their search.

Using AutoCompose

AutoCompose learns from the global population of agents over time, which is incredibly useful to newer agents or agents who are beginning to handle conversation in a newer domain.

While the model may not initially have much indication on language that a particular agent likes to use, it naturally adapts to this by surfacing up suggestions from the global response list and global history of how similar conversations have been handled in the past.  This helps ensure that agents follow consistent behaviors when handling issues that they are less certain about.

FAQs

Model Improvement

How does the suggestion model improve over time?

The model is automatically trained weekly on the latest historical data, informed by agent interactions with AutoCompose at given moments of conversations. As more situational agent behaviors are observed, better response suggestions are surfaced at more relevant points in the conversation.

Does the model adapt to topics not seen before?

As a baseline, models are able to make inferences about what existing responses are most relevant even if the topic is new.

Do new topics require new entries to be added to the global response list?

Major new topics are best handled by updating the global response list with appropriate responses. If, for example, you want to prepare for a new product launch, our recommendation is to make additions and edits to the global response list in advance, then upload on the day of the launch.

Our models will immediately start making suggestions from the updated response list. As more agents use the system, the models will become even smarter at identifying when these new responses should be suggested.4.

Response Lists

How is the global response list updated?

AI-Console gives program administrators full control to make targeted or bulk updates to the global response list and manage deployments of those changes. Once deployed to production, response list changes are immediately available to agents.

For more information, refer to the AutoCompose Tooling Guide.

Does the global response list change automatically?

The global response list does not automatically update. It is managed exclusively by product owners for a given implementation.

The organically growing list of commonly used responses updates automatically without need for manual updates.

What is the difference between the global and custom response list?

The global response list is managed by center leadership and contains a comprehensive list of responses available across the agent population. This list is intended to be the recommended wording for responding to customers.

The custom response list is managed by each agent and is exclusively accessible to them. Responses in the custom response list are suggested by AutoCompose alongside entries in the global response list. Like the global response list, the custom response list also supports a folder structure and can be manually searched by the agent.

Does the suggestion model act differently from one agent to the next?

The suggestion model uses the agent’s live conversation context and uses agent-specific response from both the custom response list and the organically-growing response list. AutoCompose suggestion models are not unique to each agent, but have different inputs and potential responses that personalize their experience.

Can the global response list be customized by queue/skill?

Yes. When the global response list is being created or edited, ASAPP can add metadata to specific responses that make them eligible for specific queues or skills, and ineligible for suggestions for all others. For example, a set of 40 responses may only be applicable for an escalation queue, and be tagged such that they don’t appear as suggestions in any conversation that appears in another queue.