Introduction

The jtel Live Agent is an assistant for agents in the call center which provides the following functionality:

• Live transcription of calls (visible in Agent Home and the Supervisor)
• Assistance during calls, such as:
  • Providing suggestions
  • Extracting information from the conversation, such as a customer number or ticket number
  • Extracting sentiment from the conversation
• Assistance after the call, such as:
  • Suggesting which transaction codes could be set for the call by the agent by analyzing the conversation
  • Automatically providing a summary of the call which the agent can copy / paste into a CRM system, or which can automatically be uploaded to a CRM system via API calls
  • Providing an automatic score or satisfaction rating on the conversation 

The live agent makes use of the following technologies in the background:

• Live ASR, currently either
  • Azure Speech Services ASR (Microsoft)
  • Whisper ASR (jtel hosted or on premise)
• and an LLM (Large Language Model) to perform the analysis of the transcribed texts, which can be a cloud based service or an on premise service

An example of the Live Agent in action is shown in the screenshot below (click to see more):

Legal Disclaimer

The live agent transcribes the conversation between a customer and an agent and in real time using an ASR engine.

Technologically:

• It is not necessary to make a call recording to use this technology.
  The audio is processed in real time and not persisted to the system, the audio is not available after the call has been processed, unless, of course, you setup a call recording in parallel.

Legally:

• You must seek legal advice before using this technology. Use of the live-agent tool requires you to obtain all necessary legal consents, for example caller opt-ins, notification that the call will be processed using AI, employee consent and any other legal and data protection aspects which may be applicable to your country or region.
  jtel GmbH disclaims all liability for any damages, losses or any other consequences resulting from your use of this technology in violation of any legal requirements.

Pre-Requisites

Configuration Parameters

The parameters required can either be set in the Client Master Data ... Parameters, or in the ACD Group ... Parameters for a specific ACD Group. 

The following parameters must be set:

Parameter: LiveAgent.Transcribe.Active

This setting defines whether the live agent is active for agents in the whole client or a particular ACD group.

We recommend setting this in ACD Group ... (specific ACD groups) ... Parameters.

Value

• 0 (or parameter not present) - the live agent will not be used
• 1 - the live agent is active and will be used in this ACD Group (or the whole client if configured at the client level)

Parameter: LiveAgent.Transcribe.Provider

This setting defines the ASR engine which will be used for transcription.

• Set to Azure if you are using
  • Azure Speech Services 
• Set to EnderTuring.v2 if you are using
  • the jtel Whisper ASR
  • or Ender Turing ASR
    (which are API compatible).

If you are using one particular ASR engine, we recommend setting this in Client Master Data ... Parameters.

Allowed Values

• Azure
  or
• EnderTuring.v2

Parameter: LiveAgent.AIPipeline.EndPoint

Set to the REST endpoint of the AI Pipeline Service in your installation.

Value

• http://acd-ai-pipeline:4000/api/v1/aipipeline

Parameter: LiveAgent.AIPipeline.Input

This setting defines the AIs which will be used during the conversation. Every time a speech recognition result is produced, the configured AIs will be queried to produce results.

• Set it in Client Master Data ... Parameters if this setting should be applied to all ACD groups which will use the live agent.
• Set it in the ACD Group Parameters for specific ACD Groups if you require different settings depending on the ACD Group used.

Currently two types are supported:

• Suggestions
  This is a tailored bot which will make suggestions to the agent depending on the content of the call. Suggestions can include things like URLs to a customer in the CRM system or product URLs providing information on how to deal with particular enquiries.
• Sentiment
  This is a standard bot which analyses the sentiment of the conversation, indicating areas of the conversation where particularly "good vibes" or "bad vibes" were detected.
  Please note, that sentiment analysis may NOT BE LEGAL in some areas even given caller and employee consent to using AI technology. Please contact your legal advisors before using it.

If you are not using any AIs during the call, then provide an empty pipeline configuration.

Example Values

{
    "endpoints" : [
        {
            "type":"suggestions",
            "url":"http://acd-ai-suggestion-bot:5083/webhooks/rest/webhook",
            "input" : {
                "sender":"%Service.StatisticsPartAID_%",
                "message":"$ai_pipeline_input"
            }
        },
        {
            "type":"sentiment",
            "url":"http://acd-ai-sentiment-bot:4001/api/v1/sentiment",
            "input" : {
                "data":"$ai_pipeline_input"
            }
        }     
    ]
}

{
    "endpoints" : [
    ]
}

Parameter: LiveAgent.AfterCallPipeline.EndPoint

Set to the REST endpoint of the AI Pipeline After Call Service in your installation.

Value

• http://acd-ai-pipeline:4000/api/v1/aftercall

Parameter: LiveAgent.AfterCallPipeline.Input

This defines the after call pipeline for either the tenant as a whole, or for a specific ACD Group. This determines what AIs will be called after the call has completed.

The two examples shown below include:

• Summarization, Transaction Code Suggestions and Satisfaction
• Summarization only

Values

{
  "endpoints" : [
      {
          "type" : "summary",
          "url" : "http://acd-ai-summary-bot:4002/api/v1/summary",
          "input" : $ai_pipeline_input
      },
      {
          "type" : "tacs",
          "url" : "http://acd-ai-tac-bot:4002/api/v1/tacs",
          "input" : $ai_pipeline_input
      },
      {
          "type" : "satisfaction",
          "url" : "http://acd-ai-tac-bot:4002/api/v1/satisfaction",
          "input" : $ai_pipeline_input
      }
  ],
  "respondTo" : {
      "host" : "acd-tel1",
      "port" : %ACD.UDP.Daemon.Port%,
      "message" : "AIDATA;AI_AFTERCALL_PIPELINE_RESULT;%Service.ClientsID_%;%QueueCheck.AgentDataID%;%CallTransfer.params.AcdConfigurationGroupsID%;%Service.StatisticsPartAID_%;%varCallData.ID%;AI_PIPELINE_OUTPUT=$ai_pipeline_output"
  }
}

{
  "endpoints" : [
      {
          "type" : "summary",
          "url" : "http://acd-ai-summary-bot:4002/api/v1/summary",
          "input" : $ai_pipeline_input
      }
  ],
  "respondTo" : {
      "host" : "acd-tel1",
      "port" : %ACD.UDP.Daemon.Port%,
      "message" : "AIDATA;AI_AFTERCALL_PIPELINE_RESULT;%Service.ClientsID_%;%QueueCheck.AgentDataID%;%CallTransfer.params.AcdConfigurationGroupsID%;%Service.StatisticsPartAID_%;%varCallData.ID%;AI_PIPELINE_OUTPUT=$ai_pipeline_output"
  }
}

Parameter: LiveAgent.AfterCallPipeline.Prompts.SatisfactionBot

This is the prompt which will be used by the LLM to generate a customer satisfaction score. 

You should tune this prompt to your specific needs, as the answers you will get from the LLM will totally depend on the question you ask the LLM.

Note the script at the bottom. This loops over the complete transcription of the conversation and passes it all to the LLM for processing.

Value

# Instructions
Please give me the most likely score from 0 to 10 the caller would give to the net promoter score question regarding his satisfaction in the following conversation. 
Answer with a JSON structure like this:
{
  "score": score
  "reason":"your reason for selecting this score"
}

The score should be a number from 0 to 10 just like the caller might answer the net promoter question.
In the "reason" field please state clearly why you selected this score, including factors or statements in the conversation which led to this conclusion.

VERY IMPORTANT: Provide all answers in the German language.
VERY IMPORTANT: DO NOT RETURN TEXT. Only return a JSON structure as described above and DO NOT return any formatting or markup. Just the JSON structure as text.

# Conversation
{% for entry in conversationData.conversationEntries %}
  {% if entry.source == 'agent' %}
Agent: {{ entry.transcription }}
  {% else %}
Caller: {{ entry.transcription }}
  {% endif %}
{% endfor %}

Parameter: LiveAgent.AfterCallPipeline.Prompts.SummaryBot.DetectLanguage

This is the prompt which will be used by the LLM to detect the language of the conversation.  

Note the script at the bottom. This passes the first 4 conversation entries to the LLM for processing This should usually be enough to reliably detect the language, but you can vary the counter if you want.

Value

# Instructions

Tell me what language this conversation was held in.
Please provide the answer as one word only - just the language name.

# Conversation
{% for entry in conversationData.conversationEntries %}
  {# Use up to 4 conversation entries for the language detection #}
  {% if loop.index <= 4 %}
    {% if entry.source == 'agent' %}
Agent: {{ entry.transcription }}
    {% else %}
Caller: {{ entry.transcription }}
    {% endif %}
  {% endif %}
{% endfor %}

Parameter: LiveAgent.AfterCallPipeline.Prompts.SummaryBot.Summarize

This is the prompt which will be used by the LLM to detect the language of the conversation. 

You should most certainly modify this prompt to your particular needs. If you have different needs for each ACD Group then setup this parameter in each ACD Group as required.

Note the script at the bottom. This loops over the complete transcription of the conversation and passes it all to the LLM for processing.

Value

# Instructions

Please summarize only the most important information from this conversation using bullet points in {{ detected_language }}.
Use as much space as you need but only up to 512 characters.
If you need to leave anything out, make sure the most important information is included.
The most important information includes:
- customer numbers
- ticket numbers
- problems and resolutions

# Conversation
{% for entry in conversationData.conversationEntries %}
  {% if entry.source == 'agent' %}
Agent: {{ entry.transcription }}
  {% else %}
Caller: {{ entry.transcription }}
  {% endif %}
{% endfor %}

Parameter: LiveAgent.AfterCallPipeline.Prompts.TACBot

This is the prompt which will be used by the LLM to suggest transaction codes.

You should modify this prompt to your particular needs. If you have different needs for each ACD Group then setup this parameter in each ACD Group as required.

Note the scripts at the bottom.

The first part loops over all transaction codes which will be the specific transaction codes which apply to conversations in the current ACD group.

The second part loops over the complete transcription of the conversation and passes it all to the LLM for processing.

Value

# Instructions
Please find the transaction codes (reasons for conversation) which apply to the following conversation.
Please only report transaction codes which actually happened, if nothing applies, then simply report an empty result.
Answer as a JSON array of objects with this structure:
{
   "ID":ID,
   "ExportKey":ExportKey,
   "reason":"your reason for selecting this transaction code"
}

In the "reason" field please state clearly why you selected this transaction code, including the questions or statements in the conversation which led to this conclusion.

VERY IMPORTANT: Provide the reason in German.
VERY IMPORTANT: DO NOT RETURN TEXT. Only return a JSON array as described above and DO NOT return any formatting or markup. Just the JSON as text.

# Transaction Codes
{% if transactionCodes %}
{% for entry in transactionCodes %}

## Transaction Code
ID: {{ entry.ID }}
ExportKey: {{ entry.ExportKey }}
Instructions: {{ entry.LLMPrompt }}
{% endfor %}
{% endif %}

# Conversation
{% for entry in conversationData.conversationEntries %}
  {% if entry.source == 'agent' %}
Agent: {{ entry.transcription }}
  {% else %}
Caller: {{ entry.transcription }}
  {% endif %}
{% endfor %}

Resources / Rights

The following resources / rights can be set per security group to allow access to the live agent: