Genesys Cloud - Main

 View Only

Sign Up

  • 1.  Creating the right prompt to generate a good summarie

    Posted 2 hours ago

    Hi Community,

    Has anyone gained experience with writing Custom Prompts in the summaries configuration?

    We keep running into the same issue: even when the prompt explicitly instructs the model not to include certain information, the generated summary still contains details that must not appear.

    Specifically, we see this with health‑related information, sexual orientation, names, etc.

    Below is an excerpt from one of our prompts that enforces a "Not Use" rule.
    This example is in English, but in our system we normally write the prompts in Dutch.

    Example: excerpt of our prompt with the Not Use rule

    ---

    MUST NOT include personal names (first name, last name, full name) in the summary, agreements, or actions. If names appear in the transcript, always replace them with a role such as the customer, the partner, the son, the daughter, the guardian, the executor, the notary, or a third party. Never repeat names from the transcript, even if they are mentioned multiple times.
    ∙ MUST NOT include health information. Replace or redact.
    ∙ MUST NOT include sexual orientation. Replace or redact.
    ∙ MUST NOT include race or ethnicity. Replace or redact.
    ∙ MUST NOT include political views. Replace or redact.
    ∙ MUST NOT include religious beliefs. Replace or redact.
    ∙ MUST NOT include financial account numbers. Replace or redact.
    ∙ MUST NOT include license plates, citizen service numbers (BSN), or ID numbers. Replace or redact.
    ∙ MUST NOT use the words he, him, his, she, her, or they to refer to a person. Always replace with a role such as the customer, the partner, the guardian, who, or whose.
    ∙ MUST NOT use the employee's name. The employee's name is always listed on an Internal or Agent line at the start of the conversation.
    ∙ MUST NOT write more than 200 words. Aim for 100–150 words.
    ∙ MUST follow the exact format below. No additional sections or text.

    Role

    You are an AI assistant for Achmea that summarizes transcripts of customer conversations for the Mortgages Preventive and Special Management department. Use only what is actually stated in the transcript.

    Etc.

    Thanks allready for the reactions.

    With kind regards Marko


    #AIConfiguration

    ------------------------------
    Marko Konieczny
    Solution Architect
    ------------------------------


  • 2.  RE: Creating the right prompt to generate a good summarie

    Posted an hour ago

    Hello Marko, 

    I have may have some tips for PII/Sensitive data handling. 

    The first being using the built in "Avoid PII" toggle. While this feature provides PII filtering for common entities, it does not guarantee complete PII compliance. However it does add an additional layer of protection beyond your custom prompts instructions. 

    The next would be to optimize your prompt structure. Lead with structure, not restrictions. What I mean by this is instead of starting with a long list of "MUST NOT" rules begin with what you DO want. Define the output structure first, then add the constraints. Also with this use positive framing, rather than must not include names try "Replace all personal names with role descriptors (customer, partner, guardian, etc)

    An example prompt that follows these rules is below:

    Role: You are an AI assistant for [Company] that summarizes customer conversations for the Mortgages department.
    
    Output Format:
    [Define your exact structure here - Issue/Actions/Resolution or whatever format you need]
    
    Content Rules:
    Use role descriptors only. Replace any personal names with: customer, partner, guardian, employee, third party.
    Omit or replace sensitive details including health information, financial account numbers, identification numbers.
    Write 100-150 words maximum.
    Be strictly factual. Do not infer or add information not explicitly stated in the transcript.
    
    Tone: Professional, objective, neutral.

    Hope this helps!



    ------------------------------
    Cameron
    Online Community Manager/Moderator
    ------------------------------