Using AI to power automations is a great way to expand your customer support and leverage all of your existing data meaningfully and efficiently. MagicReply is the AI assistant of Crisp, available directly from your inbox, the overlay, and the Chatbot to deploy scenarios that can help you improve your reactivity and reduce the load on your customer support team.

In this guide, we'll take a closer look at how to properly integrate AI into your flows, the recommendations to follow, and the mistakes to avoid. Ready? 🤖

Introduction — "Why do I need an AI Chatbot for customer service?"

AI is here to help you provide quick and efficient answers to the most common questions your customers have, and more. Crisp AI features allows you to easily integrate AI layers into your flows, directly decreasing your team's workload. This helps resolve conversations quicker, keeps your customers engaged on your platform, and allows your team to better manage their time and focus on critical tasks. Customer service is often a race against the clock, and MagicReply is here as your sidekick, helping to increase customer satisfaction and provide more continuity in your support.

You will find many fitting use-cases for MagicReply's abilities with our AI Chatbot:

As a starter tool to evaluate the first questions your users ask, generate a reply, direct them towards relevant resources, or offer them the option to reach out to an agent.
As a way to provide support during out-of-office hours, helping users resolve their queries and reduce waiting times.
As a fully autonomous interlocutor for your users to interact with regarding your product and features.
And many more possible usages, thanks to the versatility of our AI-powered chat bot platform.

If you are not familiar with our Chatbot yet, don't hesitate to look at our Getting Started guide , full of resources to help you get started on the right foot.

Crisp AI is an addon to the Plus plan and the Chatbot is available on our Essentials plan. Don't hesitate to take a look at this article for more information about the features available on this plan.

I. How to install and train my AI Chatbot?

MagicReply can be installed from the Plugins section. Once installed, you can access it from the Automation menu.



This menu allows you to train your AI on various types of resources:
Answer Snippets : Short Q&As about common themes your customers frequently ask about.
Web Content : Add all your web domains, which will be crawled and used to train MagicReply.
Helpdesk Articles : Your helpdesk articles are automatically fed to the AI and used when answering your users.
Inbox Messages : Conversations between your team and your users.

Note that when using MagicReply in your automations (with the Chatbot), the AI will not use your inbox messages for its training due to privacy reasons. Inbox messages are only incorporated when using MagicReply as an Assistant, when your agents manually generate an AI response from the conversation.

For more information about MagicReply and its training, we have an in-depth guide available on our Helpdesk .

II. How do I embed AI shots in my chatbot scenarios?

MagicReply now trained, you're ready to incorporate it in your Chatbot scenarios. From the same Automation menu, you can access the AI Chatbot and create a new scenario (or pick an existing one) to start adding AI components and get started.

Feel free to visit our Getting Started with the Chatbot plugin if you'd like to know everything about the Chatbot. You will find helpful resources there to learn and master this tool.

Available AI blocks with the chatbot builder

You will find the AI components inside of the Action blocks > AI Actions . Four different blocks are available:

Search Helpdesk : which allows you to use AI to automatically suggest articles from your Helpdesk to your users. We also have a dedicated guide on that popular functionality .
Search Answer : allows you to use AI to interpret the user's question, and automatically provide the most relevant Q&A Snippet that you have created, if any is relevant.
Search Webpages : also uses AI to interpret the question, and suggest the most relevant page of your website to your users, to guide them towards helpful resources.
MagicReply : allows you to generate a response based on your training resources (or specific ones) to reply to the user's question. Unlike previous blocks, MagicReply does not directly share the link to your URL, the page of your website or the literal content of your snippet, but it will use this data to formulate a coherent AI reply to your user in order to answer their question.

Features of the MagicReply block

The MagicReply action block offers various configuration options:

Data sources : You can select which data sources MagicReply will use (or select All Public Data Sources to use all the resources you have trained it on).
Answer Quality : You can also select the quality of the response: High, Balanced, No Qualification. This determines how selective MagicReply will be when attempting to generate a reply. When the quality requirement is higher, MagicReply will only send a reply if it is very confident in the accuracy of its response. With a lower quality, MagicReply will attempt to answer more questions, even when it is not entirely sure.
Auto-respond to greeting messages : If disabled, MagicReply will ignore messages such as "Hi" or "Hello there".
Auto-respond to unclosed response : If disabled, MagicReply will ignore messages which are not explicitly a question (statements, acknowledgments...).
System Prompt inside of the "advanced options": This allows you to give instructions to MagicReply in order to shape its personality. For instance, "You are a very patient instructor. Always provide as much detail as possible, in a fun and interesting way."*

Depending on your scenarios and flows, you can create various MagicReply blocks with different behaviors to be adapted to your users' queries. Don't hesitate to experiment! You can easily try it out from the Draft with a simple scenario, just make sure to save it and test its knowledge and behavior.



In this example, we use Proxy blocks to create a loop and go back to the "new user message". This allows us to have the same behavior as if we connected it with a wire, but in a more elegant and ergonomic way. You can learn more about proxy blocks in this article .

It is important to create a Fallback branch, in case MagicReply was not able to generate an answer. To handle cases where no pertinent response could be found, you can create a second branch from the parent block. The number on each of these branches indicates the priority in which they are evaluated: First the Chatbot attempts to generate a response 1️⃣ and in case no answer is found it sends our fallback message 2️⃣

III. Building a Chatbot scenario with MagicReply

Let's now take a look at a few application examples of MagicReply to get a better grasp on how to connect it and handle various situations within your scenarios.

Use-case: Handling first questions from your users

This is a very popular usage of MagicReply. It allows you to integrate AI very early during the exchange, diminishing the workload on your support team and quickly providing assistance to your users. In this example, we are using button pickers to route users to different flows depending on their choices, this is another very useful feature of the Bot plugin which is documented in this guide .



In this example, we start by detecting new user messages to trigger our Chatbot scenario. From there, we draw two branches for MagicReply: the first one to send the AI-generated response, and the second one to handle cases where the AI may not have found a relevant answer.

The number 1 and 2 on these wires indicate the priority of these branches. This is the order in which branches will be evaluated by the Chatbot. Here it will first attempt to generate an answer, and fallback to the second branch if it is unsuccessful.

If an answer is generated and sent, we can then offer choices to our users thanks to button pickers action blocks to cover multiple potential intents from our users, and guide them towards a resolution.
In case the user wishes to ask another question, we use a Proxy Block to redirect them back to our initial New User Message block, allowing them to restart the scenario.

⬇️ If you would like to review this scenario directly, you can download our example right here , and import it into your Bot plugin to take a closer look on your end.

Use-case: Build an omnichannel out-of-office AI bot

Now let's dive deeper into another common usage of MagicReply and AI-powered responders: out-of-office scenarios to provide continuity in your support and offer live assistance when your team is unavailable. This usage is often featured in both small and larger businesses, to ensure that your audience is able to receive help and information about your products or services.



We can break this scenario down into three main components:

Here once we first checked if the support team is offline, we then trigger MagicReply to generate a response. We also handle cases where no response could be found in a fallback branch, where we can offer them to provide more details regarding their request for instance. Once again here, we can take advantage of the priority order mechanism: we can detect the term *human* in case the user is now requesting for agent for instance. If this is the case, this branch will be triggered instead of returning back to the initial "New User Message" block.
If a reply is generated, we can therefore continue with our scenario, and offer various choices to our user, to ensure that our flow properly reflects the intent of our users.
In case the user wishes to ask another question, here also we use a Proxy Block to easily redirect them towards a previous step: our initial "New User Message" block, which allows us to loop certain steps seamlessly.

⬇️ If you would like to review this scenario directly, you can download our example right here , and import it into your Bot plugin to take a closer look on your end.

IV. Evaluating the user's intent with AI

One more way to leverage AI capabilities in your flow is by detecting the user's intent. The AI is capable to analyze the user's last messages to determine his intention, the subject of their query, their tone, mood, etc.

This allows you to handle various cases without having to offer them buttons or asking them questions directly. You can passively detect it, and use different conditions blocks to setup branches for each of the cases you want to handle.

For instance, you can evaluate:
The subject of their request (billing question, technical issue, order delivery question...)
Their mood — Does the user want to speak with an agent immediately? Are they getting frustrated?
Tone — Is the user being aggressive/abusive? Is he perhaps spamming?
And many more use-cases, if you are feeling a bit crafty :)

How to setup Intent Evaluation in your scenarios

This is a 2 part setup:

Adding the Action block "Evaluate Intent"
The first thing you'll need to do is use the AI Action block "Evaluate Intent". You will find it the Action blocks, under the "AI" sub-category.

Once this block added to your flow, you can double-click it to start configuring it.
There, you can add new Intent by naming them, and providing a "prompt" to the AI: a description of the intent you're looking to evaluate, the behaviour it should be on the look-out for.
When this block is triggered, the AI will check the user's last messages to determine the user's intent based on the ones you configured, to see if any of them match.

Adding condition blocks for each intents
You can either directly find the condition blocks "Message Intent matches" inside of the Condition blocks menu, or more easily, by selecting the previous configured "Evaluate Intent" action block, and clicking the little lightning icon on the right to automatically pop these condition blocks and insert them for you.



Handling cases where the user's intent doesn't match any of the ones you configured is important. You can actually create an intent "Other" on the Action block, and use a prompt such as "The user's intent doesn't match any of the other categories". An alternative approach also exists: you can simply have one additional branch below your action block "Evaluate Intent" without any particular condition. If this branch is placed in last position (higher priority number) it will be automatically taken by the Bot if the previous conditions did not match :)

Taking a look at a specific use-case

With the setup part out of the way, let's now improve our previous scenario with this new block.
Rather than evaluate the term "human" in their message, maybe we can check what the user's intent is before using MagicReply.

This can actually be implemented very easily. We can add this an extra step between our "New User Message" event block and the MagicReply AI one.
Here we then add a condition block for the "Human" intent, and a secondary branch with no particular condition, which the Bot will take a fallback if the first condition(s) was not met. Quite convenient!



Creativity is key here. This block allows you to handle different types of customers and behaviour very differently. This is a good way to further increase resolution speed, and ensure your users receive a personalized support experience regardless of the circumstances. This block is as versatile as you configure it to be!

⬇️ If you would like to review this scenario directly, you can download our example right here , and import it into your Bot plugin to take a closer look on your end.

Next steps

You now have all the keys in hand to get started and integrate an AI chatbot into your own customer experience to fit your use-cases! There are many more possible usages depending on your needs, so don't hesitate experimenting and leveraging the available features, such as using button pickers to collect feedback from your users, to further tailor the Chatbot and satisfy your audience's requests.

If you would like to keep exploring on the Chatbot's functionalities, don't hesitate visiting our Understand and Master the Chatbot plugin to learn more about the different blocks available. You can for instance add segments after your users click your buttons to directly consult the paths most commonly followed by your users in the analytics, or create filters to review conversations.

MagicReply is an excellent solution to reduce your support team's workload during periods of high traffic and conversation volumes. To learn more about this topic and the other tools designed to address these situations, feel free to check out our dedicated article .

Frequently Asked Questions

Still have some questions about MagicReply and AI usages which were not covered in this article or related guides? Perhaps you'll find the answer here!
Here is a collection of the most frequently asked questions, you'll also find some extra resources to help you better approach AI and its different possibilities with Crisp.

Don't hesitate to visit this in-depth article to learn more about the Bot and master your craftsmanship of scenarios .

If your question is not listed here, share it with us directly by chat, we'll be available!

I would like to be able to verify MagicReply's answer before sending it. Is this possible?

This is possible indeed. MagicReply is not only available from the Chatbot, but also directly from your inbox, in each of your conversations.
Right above the message composition area, you will find a MagicReply button which you can use to call on our AI to generate a response for you, using all your training data, as well as your inbox messages and the context of the ongoing conversation.

The generated message can then be reviewed, edited, and sent by your agents.

MagicReply is not sending any reply. What might be the issue?

MagicReply will not send a reply if it doesn't have a certain level of certainty for its response. You can verify the data you have trained it on or add additional information. Having verbose helpdesk articles is always great for both AI relevancy and SEO purposes!
You can also experiment with the Answer Quality levels you have selected in your MagicReply blocks to adjust them.

Finally, it is good practice to set up a "fallback" so that if MagicReply does not find the proper answer, you can trigger an alternative branch of your flow to be displayed instead, therefore handling all cases efficiently. More information can be found in this section.

Are there more AI tools available with Crisp?

Indeed, you will find more AI functionalities within Crisp!

Overlay is an AI-powered search interface, which you can configure to be displayed to your users before they can contact you. Overlay also uses MagicReply and the resources it was trained on, allowing you to resolve your user's most frequent queries before they even contact you.
We have a dedicated article here for more information.
It is a unique and very efficient way to further reduce your influx of requests, allowing you to quickly provide helpful resources to your users without interrupting their journey and preserving their engagement on your platform.

There are more AI features available as well, such as MagicTranscribe to automatically generate a transcript of the audio messages sent by your users, or MagicSummarize , to offer you a summary of conversations.

Updated on: 16/12/2024