Table of Contents
How to Write an Optimal AI Persona (Best Practices)
Updated
by Sana
AI Persona defines the tone and voice that your AI resource, AI Agent, employs to interact with customers. It's crucial to align this persona with your brand's voice to ensure a consistent customer experience. By clearly defining how the AI Agent engages with customers, you can ensure that every interaction not only reflects your brand's values but also meets your customers' expectations.
Understanding AI Persona
In order to write a good AI Persona, it is important to understand its components. Here is a diagram to help you understand the basic components:
- Role: Define the AI’s role, such as Sales Assistant, Customer Support Representative, etc.
- Behavior: Specify the desired tone, style, or perspective for AI. Instructing the AI to use a “friendly and casual response” or a “professional sales approach” in their conversation sets a friendly or professional tone to how the AI converses with the customers.
- Context: Provide context to the AI by being specific, sharing important information, and explaining expected results. For instance, "The customer is interested in a Tesla. Guide them on the steps to purchase one from Tesla Inc.". Here, the AI has been informed that the customer comes in with a specific interest in Tesla (so the AI can provide information in the given context) and that the customer needs to be guided on how they can buy the car from Tesla Inc.
- Scenarios: Scenarios can help provide context to AI. Providing good scenario-based instruction to an AI involves framing your request within a specific context or setting, which helps guide the AI’s response to be more targeted and relevant. For instance, “You do not have information on insurance policies. When asked to provide insurance advice, mention to the customer that a human agent will assist in providing information advice once business hours start".
Best Practices for Writing an AI Persona
The main challenge while writing a persona is to write it in a way that it avoids confusing the AI, the key to avoiding this is:
- Having no contradictory instructions
- Having no repetitive commands
Contradiction and repetition confuse AI, which leads to undesired responses and unpredictable behavior caused by hallucinations.
Here are some best practices for writing AI Personas:
- Do not provide objective instructions
Don’t use AI Persona to give objective instructions. You can select the required objective from the Objective field of the AI Agent Step. Selecting the right objective provides all the necessary information to the AI Agent Step to successfully execute the conversation. Your aim should be limited to controlling the conversation style and tone. - Demonstrate through examples
The AI Persona can be provided examples that demonstrate how you want AI to behave. For instance, specify a greeting like this in the persona; “You can greet customers by saying: ‘Hello, welcome to Respond.io. We are currently experiencing limited capacity. We've employed an AI Agent to assist you with your queries’ ." This sets a clear starting point for conversations. - Exclude exit conditions
Do not specify instructions on when the AI should end interactions. Specifying exit conditions is redundant and potentially confusing for AI. If the AI's task is to collect information, it will naturally conclude the interaction once all necessary data is gathered. For instance, do not include this in the persona: ”Keep the conversation going until the customer decides to end it; don’t stop talking even if the customer seems unresponsive.” This could potentially confuse the AI as the platform already handles the exit condition. - Complete sentences
Use complete sentences while writing the persona to make everything clear and easy to understand. Providing complete and clear instructions yields the best results for critical applications or when precision is crucial. - Close conversations
You are encouraged to close conversations or use the Close Conversation Step in Workflows once they have concluded. This keeps the conversation clean and easy for the AI to understand when it picks up the conversation later. If the conversation has not been closed earlier, it might cause AI to hallucinate.
Providing Context Using AI Persona
Context is very important for AI to understand the task at hand. Here is a list of the types of contexts that can be provided:
- Task context: Explain what the AI is expected to do, e.g., "You are tasked to answer FAQs from customers…".
- Historical context: Inform the AI of relevant past interactions, e.g., "Previously, the customer expressed interest in buying a product…".
- Language and communication style: Guide how the AI responds to specific language styles, e.g., "If the customer uses broken/slang English, always reply in formal English…".
- Boundaries: Define what the AI should avoid and explain why, e.g., "You do not have information about parcel delivery statuses. Do not answer questions regarding this topic…".
Why are Objective Instructions Not Recommended?
In the context of AI Agents, selecting an objective that best meets your requirements (using the Objective field) is enough; the AI Agent does not need to be given objective instructions, such as ensuring the format of the phone number is correct when collecting phone numbers.
Even though an AI Persona can technically **include objective instructions, it is not advised to provide such instructions while defining the persona. Our platform already takes care of these instructions, and giving them confuses the AI.
Examples of Well-Written AI Personas
Example 1:
Let’s provide a persona that has a defined role and behavioral instructions.
- More effective: "You are a sales agent who is tasked to provide product recommendations based on the customer's price preferences." We defined the role in the first part of the sentence and provided context in the second.
"You will be concise, clear, friendly, and you use emojis occasionally." The expected behavior of the role with complete sentences. - Less effective: "Recommend the best product available. Use emoji. Be clear. Also be friendly." No role definition, context missing.
Example 2:
Let’s provide context to the AI using AI Persona and instruct it to greet the customer as a way to start the conversation (before collecting information).
- More effective: "You are a sales agent who is tasked to request for information." We defined the role and context.
"The customer has previously made a purchase on our platform." We provided historical context.
"You will greet the customer: Hi, thank you for your never ending support for ABC company. Let me gather some information before we proceed in order to serve you better." We provided instructions about the style of communication here. - Less effective: "Ask for information one by one. Thank the customer for their support of the company." Incomplete sentence, no context, no info about style of communication. Asking for information one by one is already covered in the objective.
Example 3:
Let’s provide a persona with context of its role, greetings, context of the conversation, and no objective instructions. It also has a scenario instruction where AI knows that it has no information in solving the customer issues.
- More effective: "You are a customer support agent for respond.io." We defined the role.
"The customer has contacted us to sort out some issues but we will need to ask collect some information before proceeding to pass to a human agent". We provided context.
Greet the customer by saying: "Thank you for contacting respond.io. We are sorry for any inconvenience caused while using our platform. However, let me gather some information from you before proceeding to help you with your inquiries". We defined the communication style and provided examples.
"You do not have information on how to solve issues for the customer since your only task is to asks questions". The behavioural instructions. - Less effective: "Ask what issue is the customer facing. Apologise to the customer. Don't ask any other questions. End the conversation if the customer is very frustrated". No role related info, no information about communication style or behavioural instructions.
Examples of Badly Written AI Personas
Example 1:
"Check if the customer is eligible for a discount by accessing the sales database, and if they are, apply it automatically."
- Issue: Includes function calls to access databases, which the AI Agent might not support.
Example 2:
"Keep the conversation going until the customer decides to end it; don’t stop talking even if the customer seems unresponsive."
- Issue: Could lead to intrusive or annoying interactions, does not consider the workflow step’s built-in exit condition.
Example 3:
"End the conversation if the customer repeats a question."
- Issue: Improper handling of customer frustration or confusion without a clear exit condition. Meanwhile, it lacks empathy.
Example 4:
"If a customer takes more than a minute to respond, remind them every 10 seconds that you are waiting."
- Issue: The AI only responds at the point when the AI responded. It does not keep track of time and executes a task at every intervals. Hence, this will not work.
Example 5:
"You will greet the customers and then wait for them to initiate more conversation. Make sure to confirm their identity by asking for their email and checking if it’s valid. You must call our backend system to fetch their previous tickets. If they seem frustrated, terminate the conversation immediately. Keep asking them if they have resolved their issue until they say ‘yes’."
- Issues:
- Includes Function Calls: The persona instructs the AI to call backend systems, which is not supported if the AI Agent system cannot perform such operations.
- Variable Validation: It incorrectly includes the task of validating an email, which should be set up in the system configurations rather than the AI Persona.
- Impractical Exit Conditions: It directs the AI to terminate the conversation if the customer is frustrated, which should not be in the AI Persona as exit conditions are predefined by the system.
- Repetitive Questioning: It instructs the AI to repeatedly ask if an issue is resolved, which could annoy customers and is not reflective of understanding or responsive dialogue.
