Best practices for deploying a conversation bot

Before you begin

Before you get started building your conversation bot, there are some best practices for preparing help center content and planning answers to keep in mind.

Preparing your help center content

You'll want to prepare your help center content to make sure top customer issues are covered and articles and answers are easy for the bot to find. Keep these best practices in mind:

Find common issues to populate your help center. You can review tickets and other resources to review tickets, macros, and other sources to find topics for help center articles.
Optimize your content to make it easier to find using AI. The conversation bot draws information from your help center content, so well-written and formatted articles will lead to better bot performance.

Planning your answers

Before you start creating answers for your conversation bot, consider these planning best practices:

Identify questions that users ask regularly. Look at your top ticket issues, review common help center search terms, and talk to your agents to plan answers you want to create for the bot.
Create answers first for questions that can be resolved on their own and don’t need an agent to take action. Examples of common, easily-answered questions might include:
- Operating hours
- Reset password
- Store locations
Start by answering the most common questions first. It's a good idea to have answers for about 20 of your most common questions, then build up coverage over time. Don't try to address every issue right away.

Creating your standard bot responses

When defining the standard bot responses, keep these best practices in mind:

Encourage users to keep questions short and to the point.
Encourage users to ask a single question at a time. Rather than presenting "I want to cancel but I can't log in," for example, pose two separate questions.
Don’t hide the fact that the user is talking to a bot. If the user thinks they’re talking to a human, they’re likely to write long, conversational messages. The bot might have trouble understanding, and the user might feel that they’ve been misled.
Ask the user to freely ask their question with context if the bot is configured with generative replies, article recommendations, or multiple answers. The user should freely ask the question instead of using single keywords. For example, a single word, such as “refund,” can lead to confusion about the user’s intent, because it's not clear if they are interested in "refund request" or "refund policy."
Pin common answers and clarify if there is an option to speak to an agent in your greeting, clarification, or fallback responses. This can help reduce customer frustration and prevent conversation loops.
Offer the option to speak to an agent. If you can’t offer a real person, let the user know that up front to avoid frustration.

Setting up multilingual bots

If your agents serve a customer base that uses multiple languages, you can turn on automatic translation to facilitate communication. When using auto-translation:

Build your bot in a single language to optimize translation quality.
Use custom translations to manually translate selected bot messages.

Building answers

As you create your answers for your conversation bot, consider these best practices for how to structure your answers for better bot performance.

Engaging the user

When you are thinking about the start of your answer to engage the user, keep these best practices in mind:

Start each answer by echoing the user's issue back to them. This reduces the risk of confusion if the bot matches to the wrong answer. For example, if a user enters "Cancel my account," the bot response should be, "Sorry to hear you want to cancel your account."
Help end-users understand how they should navigate the bot. Depending on how the bot is designed, different end-user interaction styles can impact the bot performance. Make it clear to the user how they should navigate the bot to find answers.
- Ask the user to select from provided options throughout the answer, if the bot is designed to provide a navigation experience (one large answer flow).
- Create a separate agent transfer answer to enable the user to contact support. Link the agent transfer answer throughout the main navigation experience as an alternative option if the presented options aren’t what the user needs.
Create separate answers to handle small talk. For example, you might create a sign off response, such as "Thanks, goodbye."

Finding a solution

As you build the answer to guide the user to a solution, keep these general best practices in mind:

Avoid building out overly complex flows to specific articles. Instead, leverage generative replies to automatically return an answer. This helps minimize answer maintenance for your.
Personalize your customer experience. Create personalized experiences by requiring authentication, including conditional scenarios, or using intents.
Create autonomous actions for the user by making an API call to another system. Doing so, you can automate a majority of user requests, such as making a return, from end-to-end.
Provide a resolution to ensure each answer addresses a question. For example, state the answer to the question in a bot message, provide a link to a help article, or perform a task with an API call.

Closing out the conversation

When you are thinking about the end of your answer and how to close it out, keep these best practices in mind:

Ask for feedback to confirm the user's problem has been resolved. You can ask if question was resolved to ensure the user's issue is revolved. You can also use this feedback to analyze bot effectiveness later.
Provide alternative options to eliminate dead ends.
- If an API call fails, provide an option to transfer to agent or link to related answers. For example, if an order status retrieval fails, provide a link to an article about how users can manually check their order status.
- If you ask for feedback, and the user indicates that the problem has not been solved, provide alternative options. Consider adding an option to transfer to agent or link to related answers that might help.
- If a user contacts you outside of business hours, provide an escalation option to transfer to agent, so the user can create a ticket and an agent can respond later asynchronously.

Improving bot answer-to-question matching

You can improve the chances of the bot suggesting the right answer or article to a user by manually adding training phrases or by assigning intents to answers. To use intents, you must have an intent model assigned.

Using training phrases

You can use training phrases in answers to improve the bot matching performance. If you have an intent model, you should use intents instead of training phrases.

When you use training phrases, keep these best practices in mind:

Make note of how users word common questions and use similar language in training phrases for the answer.
Group common topics together in one answer. For example, include international shipping and domestic shipping in one answer.
Use phrases with similar or related meanings. Zendesk AI utilizes a model that employs semantic matching, where the model considers the overarching meaning of the question. For example, "solar power" and "renewable" are semantically related, and the model can recognize this connection. The model may also suggest two texts as a match if they are likely to co-occur, such as "credit card" and "bank account."
Add a variety of phrases to improve the match rate. However, you don’t need to add every single variant for how the question might be asked. For example, a user might misspell something or phrase it a little differently and still get a match.
Aim for a minimum of 3-5 training phrases for each answer.
Avoid adding single words. Bot training works best with short, multiple-word phrases that provide enough details for context. For example, use “Refund order” instead of “Refund," or use "Renew membership" instead of "Renew."
Avoid using unnecessary words and generic phrases such as "Hi" or “I want to” or "How do I." These can dilute the question's core meaning. For example, instead of "Hi, I want to get a refund," use “Get a refund."
Do not add training phrases in multiple languages. Training phrases are auto-translated, if enabled.

Using pre-trained intents

If you have an intent model, you can assign pre-trained intents to answers instead of manually adding training phrases. When you use intents, keep these best practices in mind:

Assign pre-trained intents to answers to significantly improve the question-answer match performance.
Use generative replies for intents that are frequently asked questions. These common questions can typically be resolved by the bot using information from your help center articles.

After your conversation bot is up and running

As soon as 48 hours after launching your bot, you can start monitoring activity and making updates to improve performance. To do so, keep these best practices in mind:

Monitor bot performance and make improvements. Use the prebuilt Explore reporting dashboard to help you identify how many users received a message from the bot, how many users actively engaged with the bot, and how many bot conversations were transferred to an agent.
Review bot activity and address gaps. Use the Insights dashboard to understand key bot metrics, especially the “Can’t answer” ratio. Try to reduce this rate by improving the bots content coverage, using these best practices:
- Review unresolved bot conversation transcripts to identify issues that aren't being addressed by the bot. You can create help center articles to cover these topics.
- If you have an intent model assigned, review top customer intents that are not assigned to answers. Assign those intents to relevant bot answers to minimize the chances of the bot returning a fallback response, such as “sorry I didn’t get that.”