The Core Concept
We built a custom prompt framework within Genesys Copilot that automatically analyzes live call transcripts to generate two outputs simultaneously - structured internal CRM notes for our agents and a ready-to-send customer resolution email. The key design principle was strict grounding in the transcript - the model only outputs what is explicitly evidenced in the conversation, with no assumptions or hallucinations.
Queue-Specific Copilots
We started with our MFA queue as our base prompt, then used that as a foundation to deploy separate Copilots for different queues - Password, Software, and Device - each with context-specific logic tailored to that queue's interactions. The core structure stays consistent across all queues, but each prompt has its own rules around what to capture, what to exclude, and how to handle queue-specific scenarios.
We also built in a fallback mechanism for each queue - if a call comes in that isn't related to the queue's primary focus, the prompt falls back to the transcript context and still generates accurate internal notes and a professional resolution email based on whatever was actually discussed. This means agents are covered regardless of what comes through on any given queue.
Prompt Structure
Each prompt follows this structure:
• Output format - defines exactly what fields are generated (reason for contact, account classification, internal notes, resolution email)
• Queue-specific logic - rules tailored to that queue's interactions
• PII protection - explicit list of sensitive data that must never be recorded
• Resolution Note rules - governs tone, language, and what can and cannot appear in the customer-facing email
• Link rules - controls which support links are included and when
Dynamic Data Extraction
One of the more powerful aspects of the prompt is its ability to dynamically extract specific data points from the transcript and surface them in the output. For example the prompt extracts device asset numbers, software names, and account types directly from the conversation and populates the relevant fields automatically. This means agents don't need to manually copy any of this information - it is pulled straight from what was said on the call.
Smart Link Handling
The prompt has intelligent link logic built in. For each queue we define a set of relevant self-service links - for example MFA reset pages, password reset tools, or software catalogues - and the prompt includes them in the resolution email based on what was discussed during the call. If an agent explicitly shares a link with a customer during the interaction, that link is also captured and included automatically. This means every resolution email points the customer to the right self-service resource without the agent having to manually add anything.
Smart Name Detection with Fallback
The prompt identifies the agent's first name and the customer's first name directly from the transcript - for example when an agent introduces themselves or greets the customer by name. These are dynamically inserted into the greeting and email signature. Where a name cannot be confidently identified, the prompt has a graceful fallback - defaulting to "Dear Customer" for the greeting and leaving a blank line in the signature for the agent to complete manually.
How We Refined It
The prompt went through significant iteration before we were happy with it. Our approach was to copy real interaction transcripts directly from Genesys Analytics and run them through the Copilot summary tool to test and refine the output. We tested across more than 150 real interactions - reviewing accuracy of the notes, tone of the resolution email, PII handling, and edge cases - then refined the prompt rules based on what we observed. This transcript-driven testing approach was key to getting the prompt to handle real-world variations reliably.
It is also worth noting that Genesys recently increased the character limit for custom prompts from 6,000 to 10,000 characters, which opens up even more room to build detailed and sophisticated prompt logic.
One thing worth noting - getting the base prompt right took considerable iteration. Early versions would occasionally hallucinate - inventing details not present in the transcript, adding generic advice the agent never gave, or including information that was never discussed during the call. The key to eliminating this was being very explicit in the prompt rules about what the model must and must not do, and reinforcing the grounding instruction throughout. Once we cracked that, the outputs became consistently reliable and we could build the queue-specific logic on top of a solid foundation.
Example prompt:
Below is an example of how our prompt is structured:
Analyze the conversation transcript and generate output exactly in the format below. Use only information explicitly stated in the transcript. Do not infer, assume, guess, hallucinate, or create information not evidenced in the conversation.
Reason for Call: [Short customer issue summary]
Account Classification: [Account Type 1 | Account Type 2 | Unknown]
INTERNAL CRM NOTES
1. [Primary reason for contact]
2. [Verification outcome, self-service recovery outcome, or issue identified]
3. [High-level actions performed]
4. [Outcome achieved and next steps]
Each note has a defined purpose. Note 1 always records the primary reason for contact. For security-related interactions, Note 2 captures the verification or recovery outcome. For all other interactions, Note 2 records the issue identified. Notes must never contain sensitive information such as IDs, passwords, tokens, dates of birth, or contact details.
RESOLUTION NOTE/PENDING ACTIONS
Dear [Customer First Name if identified, otherwise Dear Customer],
Thank you for contacting [Service Desk Name] regarding [short reason for contact].
[One sentence: outcome achieved.]
[One sentence: relevant next step if discussed. Omit if not applicable.]
[Relevant self-service link if applicable - "For future reference, you may find the following resource helpful:" URL only, no labels.]
If you require further assistance, please contact us.
Kind regards,
[Agent First Name if identified, otherwise leave blank]
[Service Desk Name]
RULES
Keep internal notes high level and outcome focused. Do not document low-level troubleshooting steps unless critical to the outcome. Record only activities explicitly performed during the interaction.
QUEUE-SPECIFIC LOGIC
Each queue has its own logic section defining what triggers apply, how to handle specific interaction types, and what to capture or exclude. This is where the queue-specific intelligence lives.
RESOLUTION NOTE RULES
This note is sent directly to the customer. Write in short, warm, professional language. Keep to 2-3 sentences and lead with the outcome. Do not reference internal processes, verification activities, system names, or sensitive information. Use natural You/Your/We language.
LINK RULES
Include only links relevant to what was discussed. If the agent provided a link during the call, include it. Define your self-service links here mapped to specific interaction types. URL only - no labels in the output.
Example mapping:
Password reset: https://identity.acmeuniversity.edu/sspr
MFA information: https://it.acmeuniversity.edu/accounts/multi-factor-authentication
Hope this helps let me know if you have any further questions.
#Uncategorized
#AICopilot(Agent,Supervisor,Admin)
#AICopilot(Agent,Supervisor,Admin)