Transcription
In this video, I will show you how to create a voice assistant that will call and communicate with your clients. Such an assistant can be applied in any business, and it will communicate based on the information we provide it. That is, it can pull data from a CRM system, Google Sheets, or a calendar and conduct a competent and structured dialogue based on this. First, I will show how it works, and then we will create it from scratch, even if you have no experience in this at all. So, we have a Google Calendar where I have entered several events. For example, I just indicated consultations. And here is the event indicated. There is the client's name, there is the reason, the agent's setup, and there is the phone number. I indicated my own phone number as an example. Next, we are testing it now and will see how it conducts the dialogue. It will call my actual phone number. We click workflow. The agent has launched, gathered information from the calendar, and now a call should come in. And we see that it is calling. We answer: >> "Michael, hello. This is Nicole from Eli Media agency." Is it convenient to talk now? >> Yes, >> thank you. I am calling to confirm your consultation on implementing artificial intelligence in your dental clinic. The meeting is scheduled for Saturday, May 10th, at 2:00 PM your time. Do you plan to attend at the scheduled time? >> Yes, I plan to. Listen, are these agents useful at all? Are they needed at all? >> Excellent question. AI agents are very useful for business. They automate routine tasks, such as patient registration, answering frequently asked questions, appointment reminders. This saves your team time and helps attract new clients because patients receive quick answers and convenient service. If you wish, I can tell you more during the consultation. Do you have any other questions or topics that need to be prepared for the meeting? >> No. Super. Yes, let's discuss it at the consultation. >> Okay. Then see you at the consultation on Saturday, May 10th, at 2:00 PM. I will remind you 1 hour before the start. Have a good day. Bye. >> So, this is how it works. That is, the system calls a real phone number and talks. Now, let's see how to create such an AI agent. In this case, we communicated in Russian. Sometimes there are still nuances in Russian, but overall it communicates more than decently and can absolutely be used in real life. In English, the level is even higher, where it is sometimes really difficult to distinguish from a live person. Therefore, this is a very promising development that in the next month, two, three, it may work absolutely perfectly in all languages. But you have already heard that even in Russian, it communicates quite coolly now. How do we create such an AI? We will do it step by step. How do we create such an AI agent? We will do it step by step. We will create it in the N8N service. If you don't know, it is currently one of the most convenient services for creating AIs. The N8N service. In which we register and create our agent here. I will say right away that if you want to delve deeper into the topic of creating AI agents and learn to create them from scratch, you can do so in my Telegram channel, which I created specifically for beginners and those who have been in this field for a long time, where you can find the latest AI news, as well as receive ready-made automations. Come and see for yourself. The link will be in the profile description. And we proceed to the setup. We click plus and select create. We have a trigger, a trigger will be a test for now, but we can create such AI agents on a schedule, for example, every day. If we want it to call, we can specify a schedule here. How will this work? We specify the interval, i.e., how often it will collect information from our calendar and call clients. For example, you can specify once a day. That is, for example, every day at 9:00 AM it goes into our calendar, collects all entries, and calls all our clients. This is so that it works automatically. For testing, a test run is indicated here, which we launch with a button. The next node is Google Calendar. We click the plus sign, select calendar. Here we are interested in the get many events event, i.e., collect all events, collect all events from our calendar. I spoke in more detail about connecting Google Calendar and Google services in general in my previous video. Now I will briefly explain again how to create it. We have a connection. We click here. I already have a connection created. If there is no connection, click Create New Credential and connect. In the cloud version, Google connects very easily. A Sign in with Google button will appear. Confirm the permissions and everything is connected. Next, we select our calendar. Here we specify which events to collect. In this case, it is specified for the period from now to the end interval. For example, a period of one week, i.e., from now for a week, collect all events. You can set any period. The example shows from now to 2 days. A formula is written. You can see the result, i.e., for the next 2 days. Then we perform a test. It is visible that it pulled events for the nearest period, in this case, one. If you set a period for a week and write it down, it will pull all events for the next days. It is visible that there is a consultation with Olga and a consultation with Mark. In the schema markup, you can see that it pulled this event as well. I have a limit of one set, so it only pulled one event. Remove the limit, click return all, and get all events. For testing, I only indicated one. In the working version, we will need to specify to pull all events. If we set a period for a week now, it will pull three events. This means a consultation with Olga, a consultation with Michael, and a consultation with Mark. That is, all events that it sees in the coming week. What do we do next? We want to transfer these events to our AI agent. For those who don't know again, an AI agent is, in essence, our brain that receives information, processes it as we need, and passes it on. What is the difference between a regular GPT or regular neural networks? How do agents differ? The fact that, for example, GPT is just a brain. It cannot perform any actions, but an AI agent can perform these actions. That is, it is capable not only of processing information but, for example, making a call, doing something, entering into a calendar, entering into a CRM system, and so on. To perform actions, it has tools that we connect ourselves. We add a node, click AI, and select AI agent. Then we specify define below. Connect and write the prompt. So, we have a prompt in which we transfer data about our consultations, about our calendar entries. So, you can just take and copy this. How does it work? We specify. I'll show you now. For example, we collect data here, we test. Here they are, these three items. We can specify appointment summary, i.e., general information. And we see that we can delete and drag the fields that are needed. For example, description. In Description, we wrote what the meeting is about, where the data comes from. They are pulled from Google Calendar, from the entry. Here is the summary, i.e., brief information about the event. Then it pulls the start and end times. This data is taken directly from Google Calendar. We specify the start date, start date, and end date. We just drag them so that the agent understands. Important. We add ad option systeme. We write our prompt. The prompt can be copied entirely. The essence is simple. The assistant should generate a structured JSON format. In it, specify the name, phone number, reason for consultation, start time, and end time. Note that the start and end times are transferred in ISO format, but we will change it slightly later. But initially, it is transferred exactly like this. And we tell it to return the output format strictly in JSON so that our voice agent understands all this information. After that, we connect the model. In this case, we connect GPT4.1. It is already available in the AI agents interface. This model is more advanced, it follows the script more accurately. It can be given more description. In short, it is stronger. Next, we need to specify the output parser. What is this? This is the output format. That is, we say that it should process the information and give it to us strictly in the format we specify. To do this, we turn on the slider here. Then we need to connect the parser itself. To do this, we click the plus sign and select sam parser. Here we select define below and copy the format. What does it mean? Simply that the result should be given in a special JSON format with specific fields. Properties, name, phone number, reason are strings. Start time and end time. All fields are mandatory. Done. The data comes to our AI agent. We test. We click test and wait. It processes. We look. The data is transferred in a beautiful format. Then the output format. Name, phone number, reason, start time, end time. That is, it took them from the description and turned them into structured strings so that the voice agent clearly understands. Now we need to create our voice agent. We will do it in the Retail AI service. Type Retail AI in Google and go to their website. Their website will open. Registration is simple, basic steps. You will manage. After logging in, we get to the Retail AI service dashboard. We need the AI agents tab. By the way, upon registration, they give a $10 bonus. You can register and test for some time without payment. Go to the AI agents section. Here we create an AI agent. Before that, we link the service with the output data format for the voice assistant. We create an HTTP request element. We select the POST method because we are sending data, we specify the URL for the create call element - this is a real phone call. Then authentication. We select generic credential type. Then custom Authentication. If the credential is missing, click Create New Credential. Paste the JSON Headers, where we write our API key. The header format is Bearer key. Where to get the key? In the Retail service, open API keys. Copy your key and paste it after Bearer. Save. Authentication check passes. Then in the body, select the JSON that Retail AI expects. From number field. The number that will be linked in Retail AI later will go here. The field for the number is the client's phone number. These values are substituted dynamically so that each call goes to a new client. We take the feed Phone Number that our AI agent passed and substitute it, similarly substituting the other dynamic fields: name, client's phone number for clarity, reason for consultation, AI agent setup from the calendar, i.e., our service field, which is important to transfer. We adapt this JSON template for our tasks. We also specify the start time, end time, and agent ID - this is the identifier of your AI agent. We have inserted all the fields. No additional actions are required here. We move to Retail AI, configure the AI agent. API key added. Create an AI agent. Click agents. Then Create agent. Select single prompt. There are different options, for example, conversational. But now we take single prompt, click create from scratch. What needs to be filled in? Insert the prompt, i.e., the description and dialogue script, select the model. The prompt is already prepared. Copy from the ready-made agent and paste. The prompt states: "Speak friendly, confidently, humanly, clearly, and concisely, like a caring administrator." Role: Voice assistant for the agency for implementing artificial intelligence in business Eliamedia. Assistant's name. Specify. Goal. Call the client to confirm the consultation. Structure, role, call objective, context. What data is already available? Pay attention to dynamic data. Reason, i.e., the reason, is pulled from our fields. The assistant understands the context and can immediately say that the consultation is about setting up an agent for a dental clinic. Then follows a step-by-step conversation scenario. The key moment in voice and dialogue AI agents is that the quality of interaction depends 80% on the prompt. The prompt should be tested and refined. The example from this lesson can be adapted to your goals. I recommend using a step-by-step scenario, i.e., an example of a quality dialogue with specific wording and with various branches. For example, if the answer is yes, then confirmation of participation. We write a script. Essentially, we train the agent just like an employee. Where we give a script to a person, our manager, we give the same script to the agent. If the answer is no, not sure, let's pick another slot. Is option one or option two more convenient for you? Write a clear script. Then follow additional instructions. Be sure to clarify them. For example, if the agent did not hear or understand the answer, it should politely clarify. Excuse me, I didn't hear, please repeat. You can also set a rule that if the client asks about the usefulness of agents, the agent should briefly explain how they help businesses. It is very important to always convert ISO date and time format into a understandable form in Russian so that the agent also substitutes correct values for each client. The coolest feature is that dynamic values are substituted. It calls and says, "I am calling to confirm the consultation at a specific time," and takes this time directly from the calendar that was transferred to the voice agent. Initially, the date comes in ISO format, for example, 20250514. If you leave it like this, the voice assistant will read this nonsense. We say: "Convert it into human form." For example, Wednesday, May 14th, at 3:00 PM. Substitute the actual date and time values. I separately specified that instead of 3:00 PM, say 15 hours. It needs to communicate in an understandable human style. Everything, the script is written. Then you can specify the agent's name. Now we select the model it will work with. There are different models. The most expensive GPT O Real Time costs $0.50 per minute, but you can choose GPT 4.1. It costs only $0.045 per minute and works great. It's very inexpensive. After selecting the model, settings like temperature. We usually don't touch it, leave it as is. Now we choose the voice. By default, all voices in the system are in English, we don't use them. Where to get Russian ones? Go to add custom voices. There are two options. The first is to clone your voice or an employee's voice. You need to upload an audio or video file up to 10 MB with a high-quality speech recording. The system will be able to clone a real person's voice. The second option is to use the 11 Labs service. In the Community Voices section, click Open Voices. The library opens. This is the most advanced service for voice agents. In the search, type Russian, select the voice you like, click View, copy the voice ID, and paste it into the system. Save. And we get a custom voice that can be selected. Then, be sure to specify the language as Russian. In the settings, you can play with the parameters. First, speech speed. I recommend increasing it slightly so that it sounds better on the phone. Second, temperature, how calm or emotional the voice is. Third, volume, make it quieter or louder. These parameters can be adjusted to your liking. Save. Everything, the agent is written. Then specify the welcome message. The advantage of this is that the agent can initiate the dialogue itself. We have three options. The first option: the agent remains silent until the person speaks. This is if we are working with incoming calls. Voice agents can be used for both incoming and outgoing calls. If it's an incoming call, we can set it so that the agent waits for the client's question. But most often for incoming calls, it's also good to have it introduce itself first. Therefore, we most often choose the second option, especially for outgoing calls. In this case, the AI agent initiates a dynamic first message. There is a third option, when it initiates a strictly defined welcome message. This may also be necessary if you need a specific scenario. If you want a dynamic greeting, use the second option. After selecting the greeting scenario, we have a whole list of additional voice agent settings. We won't go into detail about this now. There are functions that can be connected. For example, checking data in the calendar. You can add custom functions and link the agent to any services. Write end call. Call termination. Nothing needs to be changed here. Write call transfer. This is in case the person says they want to talk to a live employee. The agent replies: "Okay, no problem. I'll transfer you to a person now." And transfers the call to an operator or manager. We specify the company number to which it redirects. Thus, we do not lose the client. Next, Knowledge Base settings. You can add information about the company, about the business, so that the agent has context. Click Add Knowledge Base. We go to the menu, select plus. You can upload documents in PDF or text format, or add web pages. If you don't want to prepare separate regulations, just add URLs. The agent will read and understand information about the company itself. This is convenient when the client asks clarifying questions. Now about voice settings. Here you can adjust it to your liking, for example, add background sound, coffee shop noise, or static background, but it's better not to touch it. The rest is easy to understand even with Google Translate. Responsiveness, sensitivity settings, everything can be adapted to your tasks. There are also call settings, for example, timeout. After what time to end if there is no answer. Maximum call duration can also be set, but for basic purposes, you can leave it as default. There are webhooks. For advanced users, you can connect them, but beginners don't need this. Everything, click publish. Now we connect the outgoing number. You can connect an incoming or outgoing number. We choose outgoing and publish. To actually call people, go to phone numbers and add the number from which we will call. It's simple, click plus. You can buy a new number through the Twilio service. It works in America and many European countries, about 100 countries. If you are in CIS countries such as Ukraine or Kazakhstan, then Twilio will likely not work. In that case, there are other connection options. How to connect it? It's very simple for you. Click plus, connect your number via SIP. You can connect any numbers here. For example, there is the Zadarma service, which works with Russian-speaking numbers, or others depending on your country of residence. You can connect your numbers from any country because SIP numbers are available in every country. Connect the number you need, and thus you can call from any country. It absolutely doesn't matter from where. If you connect through Twilio, then specify the country code, confirm your identity online, upload a photo and document, and everything works. When the number is added, you need to check it. Here we specify our agent. Select the desired agent. If you have multiple versions, then simply select the latest published one. Add the current version. And now this number is linked to the new agent. Next, specify the new AI data. This is how we created an agent that actually calls. And this is just the beginning. Such voice agents can be connected for various tasks. Dialing from CRM, working with a cold lead base, extracting numbers through Google Maps, mass calls to hundreds and thousands of contacts. All this can be done quickly and according to a script. It is cheaper than using people for cold calls. We increase business efficiency by reaching more clients without increasing staff. At the same time, to avoid losing leads, we always leave the option to transfer the call to a person. In the prompt, you can specify: if the client is not engaging in dialogue or asks for a live employee, always redirect to a person. In the call history, we see call recordings. There is a transcription, meaning each call is converted into text, you can see the details. With the transcription, you can work further, transfer it to another agent so that they analyze where clients are lost, how to improve scripts and dialogues, or use it for other business insights. There are many possibilities. If you want to understand how to implement AI agents, advance your business to a new level, or implement this for clients, I invite you to my Telegram channel, where I constantly publish the latest AI news and ready-made automations. The link will be in the profile. Subscribe and join Telegram so as not to miss the opportunity to be one step ahead. See you in the next video.