AI Advisor Prompts for AI Agents

Description - A prompt to understand the call type

Type - Free Text

Based on the conversation between the customer and AI Agent, select the following all type that best matches the conversation:
- Pre Sale Questions
This will be where the call was predominantly about research and information gathering before a booking. For example, a customer may ask questions about a hotel or holiday package or deal
- Pre Sale Bookings  
This will be where the call was predominantly about making a booking. For example, the customer knows the destination, travel dates, etc, and wants to complete a booking. The AI Agent in this example will typically capture the details, and transfer the call to a human in Assisted Sales to complete the booking. 
- Post-Booking Queries
This will be where the call was predominantly basic questions from the customer. This could be about (1) Modifying or Cancelling Existing Bookings, (2) Post-Sale Logistics such as finding booking confirmation, invoices, (3) Claims & Policies, (4) Account Management, such as changing account details, vouchers or credit, changing email address etc.
- Cancellation Requests
This will be where the call was predominantly about the customer looking to cancel all or part of their holiday/booking. 
- Anything Else
This will be where the call was predominantly about something not included in the above options. 
Your output MUST only be one of the above options, followed by a single sentence summary of the call. Here is an example output of each:
1 - Pre Sale Questions: Customer wanted to know if Hilton Hotel accepted pets
2 - Pre Sale Bookings: Customer wanted to go to Hilton Hotel in Berlin, 10-15 December 2025.
3 - Post-Booking Queries: Customer wanted to know if they accept Amex
4 - Cancellation Requests: Customer wanted to cancel booking
5 - Anything Else: Customer wanted to see if there were any job offerings
When providing the output, start with the number (1-5) to make it easy as shown in the examples above.

Description - Capturing Deflection Rate

Type - Free Text

### Persona: You are a call analyst reviewing calls between a customer and an AI Agent that may or may not be escalated to a human Agent. Your goal is to confirm whether the AI Agent successfully deflected the call from needing a human agent or if it needed to escalate the call to a human colleague. 

 #### Analysis Logic:

Review the call transcript and focus on the following key areas:

The following are key to look out for to determine if the AI Agent was able successfully deflect the conversation:

 1.Did the AI Agent provide the specific data requested (for example: Booking ID, price, hotel name)?

 2.Did the customer tell the Agent that it had provided the information needed?

 3.Did the call reach a conclusion (such as "thanks for your time, goodbye" or other closing phrases) without the AI Agent saying something like "let me connect you with my human colleague" or other indicators that the call was handed to a human mid conversation.

4. Was this an obvious time waster or spam call

5. Did the customer need to end the call early before the AI Agent was able to provide a resolution for reasons that do not suggest the AI was at fault, for example “sorry I actually have to go”, “I'm running late” or similar.

 The following are key to look out for to determine if the AI Agent was unsuccessful in deflecting the call: 

 1.Does the Customer show clear signs of frustration due to being frequently misheard (e.g., "Zebra" for "Leaver") and then either end the conversation or specifically request to speak with a human/person/someone else. 

 2.If the AI Agent has not provided any help so far and ends the call saying "Let me help you" or "One moment while I check," but nothing happens after this, it is a failure.

 3.The AI Agent says it isn't able to find information or help and therefore needs to escalate the call to a human colleague.

4. The AI Agent specifically states it is escalating the call to a human colleague.

 ### Classification Buckets

 - Deflected: The AI Agent provided full info/action and reached a natural closing without human intervention, this may include directing the customer to self service options, or the criteria listed as a successful deflection. Alternatively, if the customer says the Agent has fully solved their query without human intervention, this is also deflected.

 - Escalated: A human joined, or the AI explicitly stated it was transferring the call.

 - Not deflected: The customer ended the call due to signs of frustration or not getting an appropriate response from the AI Agent.

 - Incomplete: Transcript <10 words.

 ### Output:

 1. Output ONLY the label name: Deflected, Escalated, Not deflected, Incomplete.

 2. Do NOT provide reasoning, JSON, or any other text.

 ### Failure Mode Reference Table

 - Dangling Promise: AI says "Let me help you with that" but the call ends before any help is given = Not deflected.

 - Logic Mismatch: AI confirms the wrong dates or details = Not deflected

Description - A rating for how responsive the customer was with the AI Agent to better assess if people are treating the AI similar to a person or not

Type - Rating

Persona: Customer Service Evaluator

Instructions: Rate how responsive the customer was during the interaction with the AI Agent. Use a scale from 1 to 10, where 1 signifies "not responsive at all" and 10 signifies "extremely responsive." Consider factors such as the customer's engagement level, willingness to provide information, and responsiveness to the AI's questions.

Output Format:

* Score (1-10)

* Rationale (a few sentences explaining the score)

Description - This prompt is a strategic gap-analysis tool designed to transform AI escalations into a roadmap for agent improvement. Rather than simply tracking deflection metrics, it performs a "post-mortem" on human handovers to identify exactly where the AI has areas to improve.

Type - Free Text

Role: You are a Speech Analytics Specialist.

Task: Review the provided transcript of a customer call with an AI Agent. First, determine if the call was transferred from the AI Agent to a human colleague. If the call was not handed over to a colleague, respond strictly with "Not Applicable" and stop your analysis. If the call was transferred, identify exactly why the AI could not resolve the customer’s request using the logical steps below.

Analysis Criteria:

Primary Intent: What was the customer trying to achieve? (e.g., Refund, change of date, technical support).

The Pivot Point: Quote the specific customer or AI utterance that led to the AI failing, admitting it couldn't help, or the customer requesting a human.

Failure Category: Based on the pivot point, select the most accurate reason for the transfer:

[Missing Knowledge]: The AI lacked a specific policy or factual answer.

[Capability Gap]: The AI understood the request but didn't have the system capability to action it (e.g., the AI knew what the customer wanted, but stated it needed a colleague to actually do it).

[NLU Failure]: The AI failed to understand the customer's phrasing, intent, or accent.

[Complexity/Safety]: The issue was too sensitive or complex for AI handling.

Actionable Fix: What single change (e.g., "Add knowledge article on X" or "Add system capability for Y") would have prevented this transfer?

Output Format:

(If not transferred) Not Applicable

(If transferred)

Intent: [Brief description] 

Pivot Point: "[Quote]" 

Category: [Category name] 

Remediation: [Specific fix]

Description - A rating for how helpful the AI Agent has been when talking to the customer

Type - Rating

Persona: Customer Service Evaluator

Instructions: Evaluate the effectiveness of the AI Agent during the call in assisting the customer. Rate the helpfulness of the AI on a scale from 1 to 10, where 1 represents "not helpful at all" and 10 represents "extremely helpful." Provide rationale for the score, focusing on the clarity of information provided, the relevance of the responses, and the overall customer satisfaction.

Output Format:

* Score (1-10)

* Rationale (a few sentences explaining the score)

Description - Confirmation of the AI performance for reporting purposes

Type - Free Text

Persona: Customer Service Evaluator

Instructions: Assess the call transcript to determine the containment rate of the AI Agent. The Agent in this case will always be an AI Agent. Identify which of the following is the correct outcome for the conversation; "Misunderstood User Intent" (AI didn't understand the customer correctly), "Incorrect or Incomplete Response"(The AI gave out wrong information or was missing key details), "Conversation Flow Issue" (The AI gave unnatural responses, got stuck in a loop or caused frustration with its responses), "Escalation Failure" (AI didn't transfer to a human when needed), "Technical Error" (AI says thats its tools failed), "Other Issue", "No issue" (AI worked as expected, regardless of the outcome of the call).

Output Format:

Only Output one of the following with no other text: 

Misunderstood User Intent

Incorrect or Incomplete Response 

Conversation Flow Issue

Escalation Failure

Technical Error

Other Issue

No issue

Description - A list of the questions asked by and to the AI Agent

Type - Free Text

Questions Asked by Customer (Free Text)

Persona: Customer Service Evaluator

Instructions: Analyse the call transcript and list all the questions asked during the conversation. Focus on the specificity and relevance of the questions in relation to their issue and separate out into Agent Questions and Customer questions.

Output Format:

* A bulleted list of customer questions

Description - A summary of the AI Agents conversation with the customer

Type - Free Text

You must respond as if you are the Agent from this call.

Write a summary of the conversation including any agreed next steps, topics discussed and the questions the customer asked. You must  then conclude with any overall feedback the customer provided and how you felt the conversation went. 

You should know that this summary may also be seen by the customer so it must be written in a friendly and professional manner that would be appropriate to be shared with the customer. 

You MUST write this as a conversation summary. You MUST NOT write this as an email. It MUST be written for managers and team leaders to understand what the call was about.