AI support: how to set everything up

Once you register in the cabinet, you already have your own platform: a support admin panel, a chat for your website and an AI assistant. The bot answers only from your knowledge base, so start there. Everything that rests on your own accounts (domain, mailbox, Telegram bot, website) you connect yourself: we do not ask for access to them. Questions, or something does not work: support@epavai.com.

Where to start

What is ready:

• the support admin panel at the address from the cabinet. The login is your email, the password is in the cabinet, block «Panel password». We keep a copy of it, so change it in the panel: «Teams», «Agents», your own account;
• the AI assistant that answers visitors from the knowledge base;
• the chat for your website and the code to paste into it;
• the team the bot hands conversations to when it cannot answer itself. You are in it.

What to do, in this order:

1. Knowledge base (section 6). Make the first article about your company: what you do, for whom, where and how to reach you. The bot takes answers only from the knowledge base, and it answers «what do you do?» from this article.
2. Chat on your website (section 5).
3. Colleagues who answer when the bot hands a conversation over, and emails to them about handed over conversations (section 7).
4. Language of the panel, the chat and the bot, if you need other than English (section 8).
5. Email and Telegram, if your customers write to you there (sections 2 and 3).
6. Your own support address, if you need one (section 1). If you want your own domain, start with this: DNS records need time.

The model that answers and billing are in sections 9 and 10.

Paths in the panel below are also given as addresses: https://<support address>/admin/... opens the page directly. The panel itself opens with the shield at the bottom of the left icon bar (tooltip «Admin»). «Inboxes» in the paths below is the panel item «Channel», «Inboxes» (/admin/inboxes), not the «Inboxes» icon at the top of the bar: that one holds the conversations.

1. Support address

Right after you register, the panel and the widget live at an address of ours like vega-07.chat.epavai.com. Your company name is deliberately not in it: addresses end up in the public certificate logs, and anyone could read it there. Changing the address is optional, and you do it yourself in the cabinet, the «Support address» block.

Another name under our domain. Type a name, press «Take it». We add the records ourselves, it takes about a minute.

Your own domain, for example support.example.com. Type it, press «Get the records»: the cabinet shows one CNAME and one TXT for verification. Add them wherever you manage the domain's DNS (registrar, Cloudflare, hosting); in Cloudflare both «DNS only» and proxying work. Records propagate from a few minutes to a day. The cabinet checks them itself and switches the address as soon as they appear, so you can close the page: an email will tell you. The check is that https://support.example.com opens the panel login with no certificate warning.

The old address keeps working after the change, so the chat on your site does not break. But the widget code holds the old address: take the new code in the cabinet and replace the old one with it (section 5).

2. Email

This section is for you if customers write to you by email. If they only write in the chat on your site or in Telegram, skip it. The EPAV mailbox does not help here: it only sends emails to your agents (section 7) and does not receive your customers' letters.

What you get:

• letters to your support address (support@example.com) become conversations, the bot and the operators answer them by email from that same address, and the customer's reply lands in the same conversation;
• a visitor who left the site gets the operator's answer by email.

For this you need a mailbox the system reads from (IMAP) and a sending server (SMTP). Use a separate support mailbox, not a personal one: the system reads the whole Inbox and gets full access to the mailbox.

Option A. You have a mailbox with IMAP and SMTP

Google Workspace, Gmail, Zoho, a hosting mailbox and the like.

1. Turn on two-step verification for the mailbox and create an app password (Google: myaccount.google.com/apppasswords, 16 characters). Mail services usually do not accept the normal mailbox password for IMAP and SMTP.
2. Ask your provider for the server addresses. For Gmail and Google Workspace: IMAP imap.gmail.com, port 993, SSL/TLS; SMTP smtp.gmail.com, port 587, STARTTLS.
3. Gmail only: Settings, «Filters and blocked addresses», create a filter «To: support@example.com», «Never send it to spam». The system reads only the Inbox, and a customer letter that Gmail called spam would stay unanswered.
4. Then go to «Entering it into the panel».

Microsoft 365 and Outlook have closed password login for IMAP and SMTP; there you need Microsoft sign-in with an app registration. Write to us at support@epavai.com, we will set it up together.

Option B. You have no suitable mailbox

A free scheme when the domain's DNS is in Cloudflare: Cloudflare forwards the letters into a separate Gmail, the system picks them up from there and sends through an SMTP service. Customers never see that Gmail: letters go out from support@example.com.

1. Technical Gmail. A new Google account «for personal use», two-step verification, an app password (myaccount.google.com/apppasswords), a filter «To: support@example.com», «Never send it to spam».

2. Cloudflare Email Routing (Compute, Email Service, Email Routing):

   • Destination addresses: add the technical Gmail and confirm it from Cloudflare's letter;
   • Routing rules, Create routing rule: support, Send to an email, the technical Gmail. If a catch-all rule exists, the new one must stand above it;
   • Settings, Subaddressing: turn it on. This is required: the customer's reply goes to support+...@example.com, and without this setting it is lost silently, without even a bounce.
3. An SMTP service for sending, for example SMTP2GO: an account, domain verification (the service gives you DNS records), an SMTP user and password. The SMTP2GO server for Europe is mail-eu.smtp2go.com, port 2525, STARTTLS.

If the domain is not in Cloudflare, forwarding at your registrar or mail provider will do, as long as it supports addresses like support+anything@example.com. Without that, customer replies would be lost.

Entering it into the panel

Email inbox. Inboxes, «New inbox» (/admin/inboxes/new), «Email inbox», «Other provider»:

Continuing a chat by email. Inboxes, «Live chat», «Email inbox for continuing a conversation», Email.

Emails to your agents from this same address are set up in section 7, «Emails to your agents».

Check it with three letters

This checks the email inbox above. If all you have is the EPAV mailbox for emails to your agents, there is nothing to check here.

1. From any other mailbox write to support@example.com: in a minute or two a conversation appears in the panel.
2. Answer from the panel: the letter arrives from support@example.com.
3. Reply to it from your mail: the reply lands in the same conversation. If it does not and you are on option B, check Subaddressing in Cloudflare.

Auto-replies («I am on holiday») are skipped by the system, so there will be no loop with an autoresponder.

3. Telegram

1. The bot. In Telegram open @BotFather, /newbot, the bot name, the bot address ending in bot. BotFather sends you a token like 123456789:AAH... - that is the bot's password, do not forward it to anyone. Avatar and description: /setuserpic, /setdescription.
2. Is the bot already running in another system? Turn it off there. One bot, one system: a bot connected in two places at once loses messages. You do not need to remove the old system's webhook by hand: saving the inbox sets ours.
3. The panel. Inboxes, «New inbox» (/admin/inboxes/new), Telegram:

The notice «you are being answered by an AI» is always added to the first message and cannot be removed: the EU AI Act requires it. The greeting and this notice come in the language of the person's Telegram. The form will not save a wrong token: Telegram checks it right away.

Leave «Greeting message» in the «Automatic replies» section empty: with the bot, the first message comes from the «First message» field together with the AI notice. The rest of the form is optional:

• «Away message» goes out when a customer writes outside business hours and an agent is handling the conversation. Pick the business hours and the time zone next to it. While the bot handles the conversation, it is not sent: the bot answers around the clock.
• «CSAT Surveys»: when the conversation is resolved, the customer rates it with stars right in Telegram.
• «Reopen window»: if the customer writes within this many hours after the conversation was resolved, the same conversation continues instead of a new one.

Further down the form has two switches:

• «AI chat without agents»: the bot answers on its own and never hands the conversation to operators. When the knowledge base has no answer, it says so and asks for another question. Turn it on when there are no operators, for example when the bot answers employees from internal documentation.
• «Invite link only»: the bot answers only people who opened it through the link in the form (https://t.me/<bot>?start=<code>). Everyone else gets a note that it works by invitation only, and nothing from them appears in the panel. You need this when the knowledge base is internal: anyone can find a Telegram bot through search. After you save, the link appears in the form with «Copy» and «New link» next to it. A new link switches the old one off: everyone who joined through it opens the new one once (someone left the team or the link leaked). To remove one person: their contact, «Block contact».

Share the link where only your people see it: your team chat or the sidebar of your Confluence space. In Confluence that is Space tools → Configure sidebar → Add link; it needs space administrator rights, not rights over the whole wiki.

4. Check. Write to the bot (with the invite link on, open it through the link): the greeting and an answer arrive.

The bot answers in private messages; it does not read group messages.

Token leaked. @BotFather, /mybots, your bot, API Token, Revoke current token. Paste the new token in the panel: Inboxes, Telegram, «Bot token», Save. An empty field on save keeps the old token.

4. The bot on a channel added after launch

Channels that existed at launch are already served by the bot. A channel added later gets the bot through a rule:

Automations (/admin/automations), the «New conversation» tab, «New rule»:

• name: Telegram to assistant (or Email to assistant);
• conditions: «Inbox», equals, your new inbox;
• actions: «Assign agent», your AI assistant (it sits in the list next to the operators);
• save, the rule is on.

Or write to us at support@epavai.com: the rule will appear by itself at the next system update.

To take the bot off a channel, switch that rule off. To hand the bot a single conversation an operator was running: the conversation card, «Details», «Actions», assignee, the assistant.

5. The widget on your site

1. Inboxes (/admin/inboxes), «Live chat», the «Installation» tab: copy the code.
2. Your developer pastes it before the closing </body> tag on every page where you want the chat.
3. The «Security» tab, «Domain list»: the addresses of your sites, one per line. *.example.com covers only subdomains, so add example.com separately. An empty list means the widget can be embedded on any site.
4. Visitors who are signed in on your site can enter the chat under their own account: the «Installation» tab, the «Identity verification» section.

If the support address changes, the code on the site has to be updated.

How the chat looks. Inboxes, «Live chat», the «Designer» tab: colors, font, window and button sizes. On top there are two tabs: «Website» is the chat on your site, «App» is the chat in your mobile app, if you connected one. Website changes are stored by the «Save» button of the settings page. In the «Widget type» group you can pick «AI chat»: the chat opens straight into the conversation, without the home screen and the list of conversations, and the «New conversation» button in the header starts over. Who answers, the bot or agents, does not depend on this type: that is the setting below.

AI chat without agents. Inboxes (/admin/inboxes), «Live chat», the «Features» tab, «AI chat without agents», «Save». The bot answers on its own and never hands the conversation to agents. When the knowledge base has no answer, it says so and asks for another question, and the conversation goes on. Office hours are not shown in the chat, and there is no reply limit. The conversations are in the common list, and an agent can answer any of them by assigning it to themselves. If you need both an AI chat and a chat with agents, create a second «Live chat» inbox, turn this switch on there and route it to the assistant with a rule (section 4).

Voice messages. A customer can say the question instead of typing it: in Telegram and WhatsApp as usual, in the chat on your site and in the app with the microphone button next to the paperclip. On the first tap the browser or the phone asks for permission to use the microphone. The bot answers a voice message in text, as if the question were typed, and the agent sees the text under the recording, marked «Voice message transcript». This works once speech recognition is set up (section 9). To remove the microphone button: «Designer», the «Voice messages» switch. In a mobile app the microphone needs one Android permission line, which the app developer adds (see the chat package guide).

6. The knowledge base

The bot answers from the knowledge base and nothing else. You keep it going: «Knowledge base» (/admin/ai/snippets), «Add knowledge» - files (PDF, Markdown, text, HTML) and links, including your site's sitemap.xml. The state of every article is visible in the list: «In search» means the bot answers from it.

A Confluence wiki is connected in the same place, on the «Confluence» tab: the address, the email and an access token, then tick the spaces. From then on the knowledge base picks up new and changed pages once a day and removes deleted ones (right away with «Update now»), and these articles are edited in Confluence. The wiki must open from the internet: Confluence Cloud (an address like acme.atlassian.net) or your own server with a public address. The bot reads Confluence tables row by row with their column names, and when asked for a link it sends the address of the page in Confluence.

The bot answers everyone who sees the chat. If the chat is on an open website, connect only spaces your customers may read. The chat can also live inside Confluence itself, for employees: the code is in the cabinet, step «Widget on the site», tab «In Confluence» (and in the panel: the web chat inbox, tab «Installation»). Paste it into General Configuration → Look and Feel → Custom HTML → At end of the BODY, and only signed-in employees see the chat. Only a Confluence system administrator can paste it, and the chat then appears on every page of the whole wiki, for every team. If the administrator says no (common in large companies), set up a Telegram bot with «Invite link only» switched on (section 3) and put the link in the sidebar of your space: space administrator rights are enough for that.

Make the first article about your company: what you do, for whom, contacts. While our model answers, the assistant description is closed for editing (section 9), and the bot answers general questions about you from this article.

You can delegate this to your own AI (Claude, ChatGPT, Cursor): «AI, Connect your AI» (/admin/ai/mcp), the step by step instructions are there too. Allowing a connection takes a staff member with the «Manage AI features» permission, and any connection is switched off in the same place with one button.

What matters more than volume: one article answers one question, and it is written the way a customer asks it. Three hundred short precise articles work better than thirty long ones.

7. Agents and handing conversations over

The bot hands a conversation to people when it cannot answer or the visitor asks for a human. The conversation goes to the default team (a new platform has one: «First line»). You are already in it.

How agents learn about a handed over conversation:

• whoever is in the panel with the «Online» status gets it assigned automatically;
• the whole team gets a notification in the panel and an email «A conversation is waiting for an operator». Emails go out once emails to your agents are set up, with your own mail or the EPAV mailbox (below); without them, only the panel notification remains.

The team email comes from the rule «Handoff to a team → notify the team» in «Automations» (/admin/automations): you can turn it off or change the text.

If you have no agents at all, switch the chat to «AI chat without agents» (section 5): the bot will not promise a human.

«Call me back». When a customer asks to be called, the bot asks for the number and a convenient time and hands the conversation to your agents. A number already known from the app or WhatsApp is not asked again. The conversation gets a note «Asks to be called back: number. Convenient time: ...», and the customer is told they will get a call back. Call from your own phone: the system has no calls. In the AI chat without agents the bot does not offer a call.

Emails to your agents

The panel emails your agents about conversations handed to them, invites new ones and sends password reset links. These emails can go out in two ways.

The EPAV mailbox, the simplest: in the cabinet (my.epavai.com), card «Emails to your agents», «Use the EPAV mailbox». Emails then come from helpdesk@epavai.com, nothing to set up in the panel. The mailbox rules:

• it works on a paid subscription;
• emails go to your sign-up email and to colleagues' addresses confirmed with a code: enter the address, press «Send a code», your colleague tells you the code, enter it in the cabinet;
• each address has a «receives emails» tick: untick it, and emails to your agents stop going there, your sign-up email included. Cabinet emails about payment and limits still go to your sign-up email;
• the card shows how many emails a day are free; after that, packages of 200, 500 and 1000 emails, and bought emails do not expire;
• the mailbox does not send the verification code to a chat visitor (identity checks in the bot's actions): that needs your own mail.

It is switched off in the same place with «Back to my own mail»: the panel then sends emails with the mail from Notifications again.

Your own mail. Notifications (/admin/notification), Email: on, the host, port, user, password and TLS of a sending server (SMTP), the sender address, authentication Login. The SMTP of your mailbox or of an SMTP service will do. If you connected email as in section 2, enter the same details as the SMTP of the email inbox; section 2 also says where to get an app password and the server addresses. Emails start going out as soon as you save.

Colleagues

«Teams», «Agents» (/admin/teams/agents), «New agent»: name, email, the Agent role, the team. Tick «Send welcome email?», it is unticked by default: your colleague gets a link and sets a password. The email goes out only if emails to your agents are set up (above), and with the EPAV mailbox only to a confirmed address: confirm your colleague's address in the cabinet before you add them in the panel. Section 10 says how many agents your plan includes: on Lite and AI chat it is one, yourself.

No email arrived: confirm the address, and ask your colleague to press «Forgot password?» on the sign-in page. You can also do without email: open the saved agent, enter a password in «Set password» and pass it on to your colleague.

8. Language

A new platform is in English. What to switch to another language:

• Platform language: «General» (/admin/general), «Language». It sets the panel labels and the system messages visitors get. System messages switch at the next system update; if you need it at once, write to us at support@epavai.com.
• Chat texts (greeting, button labels): Inboxes, «Live chat». A text in several languages is written as lines ru-RU: Здравствуйте!, en: Hello!, and the visitor sees their own language.
• Bot language: «AI», «Assistants» (/admin/ai/assistants), «Languages». Empty means the bot answers in the visitor's language, whatever it is. A list means it answers in the visitor's language if it is on the list, otherwise in the first one.

9. The model that answers

By default our model answers, and you do not pay for it. The capacity is ours and free, so we do not promise one hundred percent availability: at peak times an answer takes longer, and now and then we switch the capacity off for updates. When our model is unavailable, the bot does not go silent: the visitor gets a message that a colleague will answer, and the conversation goes to your agents. The state of the model is shown in the cabinet, block «Model».

While our model answers, the assistant description, instructions and guardrails are closed for editing: they are tuned by measurements, and editing them at random quietly makes answers worse. Name, tone, answer length and languages stay open.

Your own provider (OpenAI, Anthropic, OpenRouter and others): «AI», «Providers» (/admin/ai/providers). It takes over when our capacity is busy or off, and put first it always answers; then you pay for the model, and the assistant description opens. The backup knowledge base search, for when our search model is unavailable, is set up there too.

Speech recognition. A separate model turns voice messages into text. By default it is ours, free, with the same note about availability. Your own provider: «AI», «Providers», the «Speech recognition» list, ready options OpenAI and Groq; a provider charges less than a cent per minute of speech. If recognition does not answer, the voice message still arrives: the agent listens to the recording, and the bot asks the customer to type the question.

10. Billing and limits

Plans, prices per month excluding VAT:

All plans have the same features and differ only in limits. AI chat differs from Lite in one thing: the website chat works in the «AI chat without agents» mode (section 5). The cabinet switches this mode on by itself when you sign up from the AI chat button on our website or pay for AI chat. You can switch it off in the panel, and the cabinet does not switch it back on.

Until the subscription is paid, the cabinet shows «Trial mode, not paid», and the platform works. At the top of the cabinet you see the plan you picked on the website, or the one you already fit into, and the «Pay» button. Under the usage bars in the block «What is included and what you have used» there is the «Plans» list: you can pay for any of them. To change the payment method or cancel, use «Manage billing».

Moving to a higher plan: the «Switch to ...» button next to the plan in the same list. The limits grow right away, the difference for the rest of the month is charged to the same card at once, and from next month you pay the full price of the new plan. Coming from AI chat, the mode without agents stays on: to let the bot pass conversations to people, switch it off in the panel. There is no button for a lower plan: cancel the subscription in «Manage billing», and when it ends, pay for the plan you need.

Over a limit, nothing is switched off or cut down. On Lite, Basic and AI chat you cannot buy extra: under the bar the cabinet shows the plan you fit into, with a button to switch. On Ultra you buy extra: we bill the overage for the next month, 10 € per 10 agents, 5 € per thousand knowledge base pieces (about 300 articles), 1 € per gigabyte of storage over the limit, excluding VAT.

Emails about limits go to your sign-up email while the subscription is paid. For the knowledge base and storage we write early, once more than 80% is used; for agents, once there are more of them than the plan includes. Over a limit, on Ultra we write about the invoice, on the other plans we name the plan to move to.