Amazon Connect Customer now allows you to automatically extract any information from conversations using generative AI, making it available for review during After Call Work (ACW).
This page has been translated by machine translation. View original
Introduction
Amazon Connect has introduced the ability to use generative AI to extract arbitrary information from customer conversations.
The official documentation has been updated with instructions for configuring information extraction for voice and chat.
Additionally, "After contact work analysis available," which allows information extraction to be performed during ACW (After Contact Work), has become available as an event source for conversation analytics rules.
The information extraction feature primarily involves configuring the following two items.
- Creating an information extraction definition to define what to extract from conversations
- Creating a conversation analytics rule to define which contacts to run extraction on
In this verification, we configured extraction of customer cancellation reasons from voice calls and confirmed that the following result is displayed during ACW.
Cancellation reason: Because the monthly fee increased
We also confirmed that the extraction results are included in the Contact Lens S3 analysis results file.
This article covers the configuration method for the information extraction feature, the display during ACW, the combination with Step-by-Step Guides, and the contents of the S3 analysis results file.
What Can Now Be Done
Previously, it was possible to review an overview of an entire conversation using Contact Lens generative AI post-contact summaries.
With the new information extraction feature, business-relevant information can be retrieved in key-value format, separately from summaries.
For example, in a cancellation handling workflow, the following items could be considered.
Cancellation reason: Monthly fee increase
Target service: Standard plan
Retention result: Proposal rejected
Competing service: Competitor service
Follow-up: Not required
Summaries and information extraction serve different purposes.
| Feature | Primary use |
|---|---|
| Post-contact summary | Review the overall progress and outcome of a conversation in text form |
| Information extraction | Retrieve specified business items in key-value format |
The official documentation provides examples of extracting information as stated in a conversation, such as the name a customer gave, an invoice number, and a reservation ID.
It also provides examples of extracting information inferred from the conversation, including:
- Reason for inquiry
- Actions taken by the agent
- Next actions communicated to the customer
Extraction results can also be inserted as variables into the following actions added to the same rule.
- Send email
- Create task
- Create case
- Send notification
For example, a configuration where the extracted reason for inquiry is set in an Amazon Connect Customer Cases field is possible.
Verification Environment
The verification environment for this test is as follows.
| Item | Details |
|---|---|
| Region | Tokyo Region (ap-northeast-1) |
| Channel | Voice |
| Conversation analytics | Contact Lens real-time voice analytics |
| Extraction target | Customer cancellation reason |
| Extraction definition name | cancellation_reason |
| Display label | Cancellation reason |
| Rule event | After contact work analysis available |
| Rule condition | Matches specified queue |
The Sample after contact work flow provided by AWS was also used to display an input form during ACW.
However, Step-by-Step Guides are not a required component of the information extraction feature. If you only need to check extraction results in the Contact Control Panel (CCP), the Step-by-Step Guide configuration can be omitted.
Configuration
Adding Permissions to Security Profiles
The following security profile permissions have been added for the information extraction feature.
| Permission in UI | Permission in API | Purpose |
|---|---|---|
| Information extraction - Definitions | InformationExtraction.Definitions.Create, etc. |
Create, view, edit, and delete information extraction definitions |
| Information extraction - Results - View | InformationExtraction.Results.View |
Display extraction results in CCP and contact detail screens |
Grant the "Information extraction - Definitions" permission to users who manage information extraction definitions, based on the operations they need to perform.
Agents need the "Information extraction - Results - View" permission to view extraction results during ACW.
Information extraction permissions added to security profiles
The full list of security profile permissions can be found in the following official documentation.
Creating an Information Extraction Definition
Open Rules in the Amazon Connect management console.
Selecting the Information extraction tab allows you to manage information extraction definitions.
The information extraction tab added to the Rules screen
Select Create information extraction definition and configure the information to extract.
In this case, we initially tried specifying the Japanese name Cancellation reason, but encountered an error when saving.
When a Japanese name was specified, an error occurred when saving in this verification environment
Changing the name to cancellation_reason, composed of alphanumeric characters and underscores, allowed creation to succeed.
Within the scope of this verification, no detailed error message was displayed that clearly indicated Japanese as the direct cause of the error. Therefore, it cannot be definitively concluded that Japanese cannot be used in all environments, but using alphanumeric characters for internal names appears to be more practical.
Japanese was specified for Display label - Optional, which is shown to agents.
The configuration is as follows.
| Item | Value |
|---|---|
| Name | cancellation_reason |
| Display label - Optional | Cancellation reason |
| Prompt hint | Please extract the reason why the customer wants to cancel the product or service. |
| If information is not found | Replace with fallback text |
| Fallback text | Could not be confirmed |
Example configuration for a definition that extracts cancellation reasons
The Name is also used as input to the extraction process.
On the other hand, Display label - Optional is used for on-screen display and is not used as input to the extraction process.
The official documentation advises keeping prompt hints short and descriptive rather than overly complex. It is best to limit them to content that makes clear what you want to retrieve.
When information is not found, you can choose from the following options.
| Setting | Behavior |
|---|---|
| Omit | Do not display the item in the extraction results |
| Replace with fallback text | Display the specified fixed value |
During verification, setting a fallback value such as Could not be confirmed makes it easier to identify cases where the extraction process ran but could not detect the target information.
Creating a Conversation Analytics Rule
Simply creating an information extraction definition does not execute the extraction process.
Create a conversation analytics rule from Rules and specify when and for which contacts to run the extraction.
The configuration for this test is as follows.
| Item | Value |
|---|---|
| Event source | After contact work analysis available |
| Condition | Queue matches cm-hirai |
| Contact category | CancellationReasonExtraction |
| Action | Extract information |
| Information extraction definition | cancellation_reason |
First, specify After contact work analysis available as the event source.
Screen with after contact work analysis set as the event source and target queue configured
When After contact work analysis available is selected, information extraction is processed during ACW, and results are displayed in the "Extracted information" widget in CCP and on the contact detail screen.
In this verification, we targeted calls handled by a specific queue.
The official documentation advises using agents or queues as conditions, rather than keywords, when targeting a large number of contacts for extraction.
Next, add Extract information as an action and select the cancellation_reason definition created earlier.
Since the screen in this test also required entering a contact category name, CancellationReasonExtraction was set.
Screen with cancellation_reason specified in the Extract information action
Finally, save the rule.
The main event sources available for information extraction and their processing timing are as follows.
| Event source in UI | Processing timing | Where results are displayed |
|---|---|---|
| After contact work analysis available | During ACW | CCP, contact detail screen |
| After chat work analysis available | During ACW | Contact detail screen |
| Post-call analysis available | After contact ends and ACW is complete | Contact search, contact detail screen |
| Post-chat analysis available | After contact ends and ACW is complete | Contact search, contact detail screen |
Select After contact work analysis available to display results to agents during ACW.
When Post-call analysis available is selected, processing occurs after the contact is closed and ACW is complete, so extraction results will not be displayed in CCP during ACW.
Configuring the Inbound Flow
In the inbound flow, enable Contact Lens conversation analytics in the Set recording and analytics behavior block.
In this voice flow, the following were enabled.
- Recording of customer and agent calls
- Real-time voice analytics
- Japanese
ja-JP - Post-contact summary
Post-contact summary and information extraction are separate features. For this verification, both were enabled to allow comparison later.
The inbound flow for this test was configured as follows.
Inbound flow with conversation analytics and ACW event flow configured
This flow evaluates whether the channel is voice or chat, and then enables the corresponding conversation analytics for each channel.
Afterward, the working queue is set and the contact is transferred to the queue.
The Set event flow block in the screen is the configuration for launching the Step-by-Step Guide for ACW, which is described next.
If Step-by-Step Guides are not used, the Set event flow block can be removed. The conversation analytics settings required for information extraction, as well as the queue settings and queue transfer for connecting to agents, should remain.
Configuring the Step-by-Step Guide for ACW
In this verification, the Step-by-Step Guide for ACW provided by AWS was also configured.
Step-by-Step Guides are a feature that displays a screen in the agent workspace for agents to enter items such as contact classification, summary, follow-up, and notes when they enter the ACW state.
This configuration is not a required setting for the information extraction feature. If you only use the extraction results displayed in CCP, the configuration in this section can be omitted.
Importing the Sample after contact work flow
Copy the flow JSON from the following official documentation and import it into Amazon Connect.
The imported flow includes a Show view block that displays the AWS-managed After Contact Work view.
The imported Sample after contact work flow
After importing, open the Show view block and reselect the AWS-managed After Contact Work from the view dropdown.
The official documentation also advises explicitly selecting from the dropdown even if the view name appears to be displayed at the time of import.
After configuring, save and publish the flow.
Referencing the ACW Flow from the Inbound Flow
Add a Set event flow block to the inbound flow and specify the Sample after contact work flow imported earlier.
In this flow, the Set event flow block was placed before transferring the contact to the queue.
Set recording and analytics behavior
↓
Set event flow
↓
Play prompt
↓
Set working queue
↓
Transfer to queue
This ensures that for contacts processed through this inbound flow, the After Contact Work view is automatically launched when an agent enters the ACW state.
Note that the sample flow contains the following fixed values.
Jane Doe
(880) 953-0129
Support
Billing inquiries
Very Satisfied
During this contact, the customer...
When using in production, these fixed values must be removed or replaced with actual contact attributes or similar data.
Operation Verification
A test call was conducted with the following content.
Customer:
I would like to cancel my service because the monthly fee has gone up.
Agent:
Understood. I will process the cancellation. Thank you very much.
After the call ended and the agent entered the ACW state, the Contact Lens summary and extracted information were displayed.
Summary
The customer requested cancellation of the service due to an increase in the monthly fee.
The representative acknowledged the cancellation request and processed it.
Extracted information
Cancellation reason: Because the monthly fee increased
Screen displaying the generative AI summary and cancellation reason during ACW
The summary is the post-contact summary generated by Contact Lens generative AI.
On the other hand, Cancellation reason: Because the monthly fee increased is the result of the cancellation_reason information extraction created in this test.
| Displayed content | Configuration |
|---|---|
| Summary | Post-contact summary enabled in the Set recording and analytics behavior block |
| Extracted information | Information extraction definition and conversation analytics rule created |
Both use generative AI, but their configurations and purposes differ.
After the contact ended, the same extraction results were also viewable on the contact detail screen.
Cancellation reason displayed on the contact detail screen
Extraction results can be reviewed not only in CCP during ACW, but also after the contact ends.
Reviewing Content Entered in the Step-by-Step Guide
The After Contact Work view allows input of items such as the following.
- Contact classification
- Reason for inquiry
- Customer satisfaction
- Whether follow-up is needed
- Resolution status
- Modified summary
- Contact notes
In the Sample after contact work flow used in this test, the submitted content is saved to the following contact attributes.
Category
Driver
Satisfaction
FollowUp
Resolved
ModifiedSummary
ContactNotes
On the contact detail screen, these were viewable as user-defined attributes.
Result of content entered in the After Contact Work view being saved as contact attributes
In the Sample after contact work flow, the values returned from the Show view block are saved to the current contact and associated contacts using the Set contact attributes block.
It is also possible to review the extraction results displayed in CCP and enter them into the modified summary or contact notes fields of the Step-by-Step Guide.
Cancellation reason: Because the monthly fee increased
With this approach, agents can review the values extracted by generative AI before saving them as contact attributes.
However, in this verification, the extraction results displayed in CCP were not automatically populated into the After Contact Work view. The configuration requires agents to review the extraction results and manually transcribe them into the appropriate input fields.
Reviewing the S3 Analysis Results File
Upon reviewing the Contact Lens S3 analysis results file, the extraction results were found in the following location.
ConversationCharacteristics
.ExtractedInformation
.ExtractedInformation[]
Extracting only the relevant portion, the content is as follows. IDs have been replaced with examples.
{
"ConversationCharacteristics": {
"ExtractedInformation": {
"ExtractedInformation": [
{
"AnalyticsMode": "RealTime",
"ExtractionDefinitionDisplayLabel": "Cancellation reason",
"ExtractionDefinitionId": "xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx",
"ExtractionDefinitionName": "cancellation_reason",
"Values": [
{
"Content": "Because the monthly fee increased",
"PointsOfInterest": [
{
"BeginOffsetMillis": 2767,
"EndOffsetMillis": 6047
}
]
}
]
}
]
}
}
}
The content of each item is as follows.
| JSON key | Content |
|---|---|
AnalyticsMode |
The analytics mode in which information extraction was executed |
ExtractionDefinitionDisplayLabel |
The label displayed on the agent screen |
ExtractionDefinitionId |
The ID that uniquely identifies the information extraction definition |
ExtractionDefinitionName |
The internal name of the information extraction definition |
Values[].Content |
The extracted value |
PointsOfInterest |
The position of the utterance that served as the basis for extraction |
In this example, the following values can be retrieved.
ExtractionDefinitionName: cancellation_reason
Values[].Content: Because the monthly fee increased
When integrating with external systems such as CRM, it is more practical to use ExtractionDefinitionName as the integration key rather than the display label.
Display labels may change during operation, but keeping the internal name fixed makes it easier to maintain the mapping with CRM fields.
Note that Values is an array. When integrating with external systems, implementations should account for cases where a single extraction definition contains multiple values.
Can Also Be Used for CRM Integration
By receiving the COMPLETED event from Amazon EventBridge after ACW ends and using AWS Lambda to retrieve the Contact Lens S3 analysis results file, extraction results can be transferred to a CRM.
ACW ends
↓
Amazon EventBridge
↓
AWS Lambda
↓
Retrieve Contact Lens S3 analysis results file
↓
Read ExtractedInformation
↓
Update CRM
The COMPLETED event indicates that a contact has fully ended, including ACW where applicable.
If you have previously been sending transcripts from Lambda to Amazon Bedrock to generate CRM fields, there is a possibility of replacing the Bedrock inference portion with Amazon Connect's information extraction feature.
This eliminates the need to call Amazon Bedrock models from Lambda, which can reduce the model inference costs that were incurred from those Bedrock calls. It can also simplify implementation around prompt management, response parsing, and retry logic on errors.
In this verification, the information extraction results were not included in the response from the DescribeContact API. Therefore, in the configuration confirmed in this test, extraction results are retrieved from the Contact Lens S3 analysis results file.
On the other hand, content entered by agents via Step-by-Step Guides is saved as contact attributes, so it was retrievable from Contact.Attributes in the DescribeContact API response.
Summary
Amazon Connect has introduced the ability to use generative AI to extract arbitrary information from customer conversations.
In this verification, we extracted cancellation reasons from customer utterances and confirmed they could be viewed in CCP during ACW and on the contact detail screen. We also confirmed that extraction results were included in the Contact Lens S3 analysis results file.
When combined with Step-by-Step Guides, agents can review extraction results and save them as contact attributes alongside summaries and notes.
If you have previously been using a combination of Amazon EventBridge, AWS Lambda, and Amazon Bedrock to generate CRM fields, leveraging Amazon Connect's information extraction results may allow you to reduce the processing that calls Amazon Bedrock from Lambda. This can simplify the integration process while also reducing the model inference costs that were incurred from those Bedrock calls.
