💸
A fee for every resolution
Intercom charges $0.99 per resolved conversation, Help Scout $0.75, HubSpot $0.50, Gorgias $1 on top of its per-ticket fee. The better the bot works, the bigger the bill.
What you need
AI answers your customers.
On a subscription, with no fee per resolution.
An AI chat for your website that does not make things up.
€49.99 per month, no fee per message.
A complete help desk with an AI bot, like the market leaders have. The bot answers from your knowledge base in website chat, WhatsApp, Telegram, email and your mobile app, and hands whatever is not in the knowledge base to an agent. You pay a fixed subscription, and for AI you pay the provider at its own price.
Not a toy chatbot, but the same system as in the support plans: the bot answers only from your knowledge base and does not make things up. You pay a subscription, and for AI you pay the provider at its own price.
Your platform is assigned right after registration, no card needed to try it.
- ✓ €0 per resolved conversation
- ✓ Your own AI provider with your key
- ✓ Server in the EU
Store support
AI answers, human agents on hand
AI answers from the knowledge base
example
Can I return an item without a receipt?
-
Yes, within 14 days. The order number from your confirmation email works instead of a receipt. You can start the return in your account.
Source: article “Returns and exchanges” What if I paid with a gift card?
- No answer in the knowledge base. Conversation handed to an agent.
-
Our knowledge base has no answer to this question. Please ask it differently or ask about something else.
The problem
AI from the market leaders: convenient until the bill arrives
Big help desks and chatbot builders charge for every step: per agent seat, per resolved question, per message. The better your support works, the bigger the bill.
🪑
A fee for every seat
A seat per month on monthly billing: from $39 at Intercom, from $69 at Zendesk Suite, from $100 at HubSpot with the AI agent. Every new team member raises the bill.
🎟️
Credits with multipliers
Chatbot builders count credits, and an answer from a good model costs 2 to 50 credits. You cannot work out in advance what you will pay for the month.
🎲
The model is chosen for you
Intercom, Zendesk, Freshdesk and HubSpot do not let you connect your own model or an AI provider with your own key. The vendor decides what an answer costs and who reads the conversations.
Plans
A subscription for the whole system, not for every seat and resolution
Conversations and resolutions are not part of the price and do not increase the bill. The plans differ only in limits: every feature is in every plan.
Lite
The owner answers personally, the bot handles the rest
€49.99 per month, excl. VAT
- Agents
- 1 + AI assistant
- Knowledge base
- about 1,500 articles
- Storage
- 5 GB
For a small team
Basic
A support team with the bot on the first line
€99.99 per month, excl. VAT
- Agents
- 5 + AI assistant
- Knowledge base
- about 3,000 articles
- Storage
- 20 GB
Ultra
A large support operation and a large knowledge base
€199.99 per month, excl. VAT
- Agents
- 25 + AI assistant
- Knowledge base
- about 7,500 articles
- Storage
- 50 GB
A separate product
AI chat
€49.99 per month, excl. VAT
The website chat works without agents from the start: the bot answers visitors from your knowledge base, and if there is no answer, it says so. Inside is the same system as in the support plans. When you need people, turn on handoff to agents with a single switch.
- Agents
- 1 + AI assistant
- Knowledge base
- about 1,500 articles
- Storage
- 5 GB
In every plan
- ✓ An AI assistant that answers only from your knowledge base
- ✓ Website chat, WhatsApp, Telegram, email and chat in your mobile app
- ✓ MCP knowledge server with no fee per action: your AI maintains the knowledge base itself
- ✓ Files, links, sitemaps and Confluence as knowledge sources
- ✓ Your own support address or domain, your brand and the chat builder
- ✓ “Powered by” in the chat can be turned off for free
- ✓ The EPAV mailbox for emails to agents if you have no mail of your own
- ✓ Reports, SLAs, customer ratings, rules and macros
- ✓ Server and database in the EU, a separate installation per company
And how much does AI cost?
As much as your AI provider charges for tokens: you connect it with your own key and pay it directly, with no markup from us. On an inexpensive model a conversation costs about $0.007, roughly $7 per thousand conversations. You can try it free on our model, but that is a trial option with no availability guarantee.
How to choose a modelCost
Their bill grows with every agent and resolution. Ours barely changes
Enter your team and conversation volume. Competitor prices are taken from their pages as of 30 September 2026, billed monthly, in US dollars.
This share drives the bill of those who charge per resolution. Intercom reports an average of 76% across its customers, Help Scout 73%.
Model at your provider
$0.20 and $1.20 per 1M input and output tokens, about $0.007 per conversation
$15 074
you keep per year compared with Intercom
Monthly bill
Every agent costs extra. With us, plans go up in steps
Monthly bill depending on the number of agents. Conversations per month: 2,000, the bot resolves 60%.
Every resolution costs extra. With us, tokens at the provider's price
Monthly bill depending on the number of conversations. Agents: 5, the bot resolves 60%.
Show as numbers
| Solution | How it is charged | Per month | Per year |
|---|
How we calculated
- Prices from the vendors' pages as of 30 September 2026, in US dollars, excluding taxes, billed monthly. With annual billing they are 15-25% cheaper; we have one price.
- For each vendor we took the lowest plan that has an AI agent: Intercom Essential, Zendesk Suite Team, Freshdesk Omni Growth (AI in chat only in Omni), HubSpot Service Hub Professional, Help Scout Standard, Crisp by number of seats, Chatwoot Startups, Gorgias by number of tickets.
- Intercom, Help Scout, HubSpot and Gorgias charge per resolved conversation, Freshdesk for every AI session, resolved or not. Gorgias also counts every conversation as a ticket. HubSpot charges a one-time $1,500 for mandatory onboarding, which is in the yearly total.
- Zendesk does not publish its price per resolution, so only its seats are counted. Crisp itself estimates an AI conversation at $0.05-0.10; we took the midpoint. At Chatwoot every AI reply costs a credit, with 300 included per month.
- Our conversation, measured in production: 4 customer messages with a knowledge base search, about 29 thousand input and 0.7 thousand output tokens. Token prices from the providers' public price lists as of 30 September 2026. Euros converted to dollars at the ECB rate of 1.1355 on 29 September 2026.
- Our plan is chosen by the number of agents: 1 is Lite, up to 5 is Basic, up to 25 is Ultra. The bill includes both the subscription and the provider's tokens, so we are stricter with ourselves than with competitors.
- Chatbot builders: for each one we chose the cheapest plan for your volume, topped up with credits at their prices. A bot answer is one credit or one message, and a conversation is 4 answers. Credits per answer depend on the model class: 1 for inexpensive models, and from 2 (Chatbase) to 12 (Chatling) for flagship ones. Tidio sells more than 1,000 Lyro conversations only through its sales team.
AI model
You choose the model and pay the provider at its own price
Almost any AI provider will do: you connect it with your own key and switch or combine providers yourself, without us.
Tokens versus per-resolution fees
A thousand conversations per month, the bot resolves 60%. At Intercom that is 600 resolutions at $0.99, so $594. At an AI provider you pay for the tokens of every conversation, both resolved and handed to people:
- Inexpensive$789 times cheaper
- Open, in the EU$1539 times cheaper
- Mid-range$3318 times cheaper
- Flagship$659 times cheaper
Answering from a knowledge base does not need a flagship model: the bot retells the article it found, and an inexpensive model handles that.
A backup instead of downtime
Providers work in order. If the first does not answer, the request goes to the next one and the bot keeps answering.
- 1 Main provideror your own model in the office not answering
- 2 Backup providerfor example, in the EU answering
- 3 Any otherwith your key on standby
A failed provider rests and comes back on its own
- ✓ If there is no backup provider, the admin panel says so plainly
- ✓ Knowledge search has a backup too: if the main search is unavailable, the backup takes over
- ✓ A key expired or the provider account ran out of credit: the bot answers through the next one
Add-on
Your own AI on an office computer
No server GPUs needed. The model behind our own bot runs on an ordinary computer with a single AMD Radeon AI PRO R9700. For a separate fee we set up a gateway, and your bot answers from your own GPU.
- CPU
- AMD Ryzen
- Memory
- 32 GB
- Storage
- SSD, 512 GB or more
- GPU
- AMD Radeon AI PRO R9700, 32 GB
about €2,500
one GPU
about €5,000
two GPUs
Component prices in German shops as of September 2026, excluding VAT; with VAT about €3,050 and €5,900-6,000. Memory prices rise fastest, so we confirm the price before purchase. You buy the computer, we choose the parts and set it up.
Conversations do not go to an AI supplier
The model runs on your side, so neither an AI supplier nor its contractors see the text of your conversations. The system itself and the database still run on our server in the EU.
Electricity instead of tokens
Tokens cost nothing. Electricity in normal support use is about €10-35 per month, and up to €150 with both cards fully loaded around the clock, at the average EU business rate.
One graphics card is enough
Everything runs on a single card. A second one is only a spare: if one breaks, the bot keeps answering.
No server hardware needed
Server GPUs are in short supply, while this card is sold in ordinary shops and fits a desktop computer.
One card handles dozens of conversations per hour
The bot answers several customers at once, without a queue. That is how our own bot runs on the same card.
AI chat
Not a toy chatbot, but the same system as in the support plans
Under the hood of the AI chat is everything the big help desks have: a knowledge base, source checks in code, reports, and agents you can turn on with a single switch.
Answer quality and cost do not depend on the size of the knowledge base
This is an advanced approach: the bot does not reread the whole knowledge base, it finds the few articles that answer the question and writes only from them. A hundred articles or several thousand: the answer is just as accurate and costs the same.
-
1
Understands the question
Searches by meaning: it finds the answer even if the customer asked in their own words or in another language.
-
2
Picks the best
Keeps the articles that answer the question, not the ones that merely look similar.
-
3
Answers from the source
Writes only from the articles it found. No source, no answer.
🧾
When it does not know, it says so
No answer in the knowledge base: the bot says so honestly and asks the customer to rephrase instead of making something up. When asked for a human, it replies that only AI answers here.
🔀
Agents with a single switch
All conversations are visible in the panel, and you can reply in any of them. When you need handoff to people, turn off AI chat mode, and the bot starts passing questions to an agent.
💶
No credits or multipliers
We do not count messages. For AI you pay the provider for tokens at its price, and an expensive model does not eat through a limit 10-50 times faster.
Accuracy
The bot does not make things up: an answer without a source never reaches the customer
The key rules apply to every answer, and they cannot be switched off or bypassed.
- 1Customer questionin any language
- 2Knowledge base searchfinds the articles that answer the question
- ?Source found?checked in code
Yes
An answer based on the article. The customer gets a solution right away.
No
A long answer without a source is never sent. In support the conversation goes to an agent with the reason “not in the knowledge base”, in AI chat the bot says there is no answer.
🧭
A reason for every handoff
The code records why a conversation went to a human: no answer in the knowledge base, the customer asked for an agent, attempts ran out.
📝
Learns from agents' answers
From resolved conversations where a human answered, the system suggests new articles. They enter the knowledge base only after approval.
🛡️
The bot listens only to you
A customer cannot talk the bot into breaking your rules: a request in the conversation does not cancel them.
Knowledge
Train the bot with ordinary files
No markup and no preparation. Upload what you already have, connect Confluence or let your AI assistant do the migration.
Upload as it is
- Markdown
- TXT
- HTML
- Links
- sitemap.xml
- Confluence
- ✓ Files up to 25 MB each, and PDFs with columns and tables keep their text order
- ✓ A sitemap.xml link brings in up to 300 pages in one import
- ✓ Connect Confluence once: the knowledge base picks up changes on its own once a day
- ✓ Uploading again updates the article instead of creating a copy
- ✓ Every article shows whether it made it into search
- ✓ Help center articles in 67 languages, and the bot replies in the customer's language
returns-policy.pdf
In search · 12
help.example.com/sitemap.xml
Indexing
old-price-list.pdf
Not in search
Or move the knowledge base for free with your own AI over MCP
Every plan includes an MCP knowledge server. Connect Claude, ChatGPT, Claude Code or Cursor with the address and an admin sign-in, no keys, and ask it to move your articles. Your assistant converts Word documents to HTML itself.
Your AI assistant, example
You Move all articles from our Zendesk into the knowledge base. Show me the plan first.
AI The dry run plan is ready, nothing in the knowledge base has changed yet:
- +124 to add
- 9 to update
- 31 unchanged
- 2 conflicts: articles edited by hand
AI Two articles were edited in the admin panel. Keep your edits or take the text from Zendesk?
- ✓ Plan first, then write
- ✓ Manual edits are never overwritten silently
- ✓ Any import can be rolled back
- ✓ On any plan, no fee per action
MCP (Model Context Protocol) is an open standard that lets AI assistants work with external systems.
Mobile app
The same chat in your app
A ready Flutter package opens the support chat full screen in your app. The bot, knowledge base and agents are the same as on your website.
- ✓ App users enter the chat without typing an email, and the bot answers their personal questions
- ✓ Attachments from the camera, the gallery and files
- ✓ The language of the chat and the bot's answers comes from the app
- ✓ Signing out of the account clears the chat history
- ✓ The chat's look in the app is set up separately from the website: a change on the website does not alter an app already released
- ✓ Works in native iOS and Android apps as an embedded Flutter module
Intercom, Zendesk, HubSpot and Help Scout have no official Flutter package. We hand over the package when you connect.
Email works even without your own domain
Agents learn about new conversations by email. If you have no mail of your own for sending notifications, turn on our mailbox: there is nothing to set up.
Your own mail
Customer emails become conversations
Emails to your support address land in the same list as chat and messengers. The bot and agents reply from that same address, and the customer's reply goes into the same conversation. Any mailbox with IMAP and SMTP works, even a regular Gmail.
- ✓ A visitor who has left the website gets the agent's reply by email
- ✓ Emails from one address can be merged into one conversation
- ✓ The system skips auto-replies such as “I am on vacation”
EPAV mailbox
Notifications to agents from our address
If you have no domain or mail server of your own, turn on the EPAV mailbox in the cabinet, and your team immediately gets emails about new conversations, mentions and deadlines from [email protected].
- 10 emails a day free
- 200 emails 2 €
- 500 emails 4 €
- 1,000 emails 6 €
Prices exclude VAT, and purchased emails do not expire. The mailbox sends emails to your team; email conversations with customers go through your own mailbox.
What is included
A complete help desk, not just a bot
💬
Five channels in one inbox
Website chat is a small code snippet, WhatsApp and Telegram connect in the admin panel, email arrives in the same conversation list, and a Flutter package puts the chat inside your mobile app. Agents reply from one place. Meta bills WhatsApp messages itself at its own rates, and we charge nothing on top.
👥
Agents and teams
Auto-assignment, SLAs, customer ratings, macros and rules.
⚙️
Actions through your API
Order status, a new request, finding a customer by number.
🎨
Your brand
Name, logo, colors and your own domain. “Powered by” can be turned off.
📊
Reports
Conversations, ratings, deadlines, bot performance and the queue in the supervisor panel.
🔐
Access
Custom roles and permissions, sign-in with Google and Microsoft.
📦
Export any time
The whole knowledge base can be exported, conversations and contacts on request.
📚
Help center
Public articles for customers at your support address, with search and a “Was this article helpful?” vote.
🌐
Languages
The interface comes in 14 languages: English, Russian, Ukrainian and Estonian are fully translated, German, French, Spanish and seven more partly. Help center articles in any of 67 languages. The bot replies in the customer's language.
📵
What we do not have
Calls and a voice bot. Voice messages in the chat and messengers are understood and answered in text.
Data and reliability
A separate installation per company, a server in the EU
Your own installation
Your own database, domain and brand. Not a shared account in someone else's panel: data of different clients is physically separated.
Server in the EU
The system and database run on a server in Amsterdam, on any plan. Intercom offers EU data storage only from the Advanced plan, Zendesk from Professional.
AI disclosure in code
The first message tells the customer they are talking to AI. The code adds this line and it cannot be turned off: Article 50 of the EU AI Act requires this.
Personal data only after verification
Before showing an order status or other personal data, the bot verifies the customer with a one-time code sent by email.
Railway hosting, EU West region (Amsterdam). Railway is a US company and, under its data processing agreement, also processes the data in the US under the EU-U.S. Data Privacy Framework. GDPR applies to any processing of personal data, and the controller's obligations remain yours. Current hosting status: status.railway.com
An honest comparison
EPAV Support and the big help desks
EPAV AI chat and chatbot builders
Only verifiable facts: prices and features from the vendors' public pages.
Scroll the table sideways: the other vendors are to the right.
| What we compare | EPAV Support | Intercom | Zendesk | Freshdesk | HubSpot | Help Scout | Crisp | Chatwoot |
|---|---|---|---|---|---|---|---|---|
| How you pay for AI | Tokens at your provider's price | $0.99 per resolution | Per resolution, price not published | $0.49 per AI session | $0.50 per resolution | $0.75 per resolution | Credits, about $0.05-0.10 per conversation | $0.02 per AI reply beyond those included |
| Agent seat | In the plan: 1, 5 or 25 | $39 per month | $69 per month, Suite Team | $35 per month, Omni Growth | $100 per month | $30 per month | 4-20 seats in the plan, then $10 each | $19 per month |
| 5 agents and 2,000 conversations per month | $127 | $1 383 | $345 plus resolutions | $1 155 | $1 070 | $1 050 | $200 | $249 |
| Your own AI provider with your key | Any OpenAI-compatible one | No | No | No | Only in Enterprise workflows | No | Choice from a list, no own key | Self-hosted version only |
| Your own model in the office | Yes, as an add-on | No | No | No | No | No | No | Self-hosted version only |
| MCP: your AI manages the knowledge base | Free on any plan: plan first, conflicts, rollback | Creating and editing articles | Server announced, not released | Articles, $15 per 1,000 actions beyond those included | Knowledge base not covered | Read only | Read and write, from the Essentials plan | No |
| Telegram out of the box | Yes | In the help center, “coming soon” on the pricing page | Marketplace only | Marketplace only | Marketplace only | No | Yes, from the Mini plan | Yes |
| Flutter package for your app | Yes | Not officially supported | No | Yes, in Freshchat | DIY wrappers only | Third-party only | Yes | Yes |
| Remove “Powered by” | Free, on any plan | From the Advanced plan, $85 per seat and up | In chat, Enterprise only | On paid plans | From the Starter plan | From the Plus plan | Plus only, $295 per month | Enterprise only |
| EU data storage | On any plan, Amsterdam | From the Advanced plan | From the Professional plan | Yes, region chosen at sign-up | Yes, Germany | No, US | Yes | Cloud in the US |
| What we compare | EPAV AI chat | Chatbase | SiteGPT | DocsBot | Chatling | Tidio Lyro | Intercom Fin |
|---|---|---|---|---|---|---|---|
| Entry price per month | €49.99 plus tokens | $40 | $59 | $49 | $25 | $39 for 50 conversations | from $49.50 for 50 resolutions |
| How you pay for AI | Tokens at your provider's price | Credits, 2-6 credits per answer from a good model | Messages, 10 times pricier on the top model | Credits, model multiplier up to 50 | Credits, up to 25 per answer | Per conversation, $0.70-0.78 | $0.99 per resolution |
| 2,000 conversations per month | $70 | $310 | $118 | $98 | $95 | on request | $1 188 |
| Live agent | Panel included, turned on with a switch | From the Standard plan, $150 | Escalation by email | Handoff to third-party help desks | Yes | Yes | Handoff to your help desk |
| Your own AI provider with your key | Any OpenAI-compatible one | No | No | OpenAI only | No | No | No |
| MCP: your AI manages the knowledge base | Free: plan first, conflicts, rollback | Edits and deletions right away, no rollback | Yes, in the plan | Read only | No | No | No |
| Remove “Powered by” | Free | +$99 per month | +$59 per month | The $499 plan | +$35 per month | As an add-on | Not found |
| Telegram out of the box | Yes | Via Zapier | No | Yes | Yes | No | No |
| Refusal instead of making things up | Code blocks answers without a source | Not found | They advise adding it to the instructions | Refuses by default | Docs admit it makes things up | Claimed | Under 1% errors claimed |
available partly or with conditions not found on official pages
Checked against the vendors' official pricing pages and help centers as of 30 September 2026. The bill is calculated with the same formulas as in the calculator above, on an inexpensive model, with the bot resolving 60%. Vendors change their terms, so check their pages before deciding.
Sources
- Intercom pricing seats MCP branding data location Flutter
- Zendesk pricing resolution billing MCP branding data location Telegram
- Freshdesk pricing AI sessions MCP
- HubSpot pricing credits MCP
- Help Scout pricing MCP data location
- Crisp pricing credits MCP
- Chatwoot, Gorgias Chatwoot: pricing Gorgias: pricing
- Chatbots Chatbase SiteGPT DocsBot Chatling Tidio Fin
- Token prices OpenAI Anthropic ECB rate
- Law EU AI Act GDPR Railway DPA
Launch
From registration to the bot's first answer
Everything is done in the cabinet and the admin panel, with no calls and no implementation project.
-
01
Registration
Email, password and company name. Your support platform is assigned right away.
-
02
Knowledge base
Upload files and links, connect Confluence or let your own AI do the migration over MCP.
-
03
Channels
Chat code for your website, WhatsApp, Telegram and email, a package for your mobile app.
-
04
Model
The bot answers on our model right away. For real work, connect your own provider with your key.
Live example
The chat on this page runs on EPAV Support
It answers from the articles on our website. Ask it about the plans, then about something unrelated, and see what it does.
The chat button is in the lower right corner of the page.
Questions and answers
What people usually ask
Lite €49.99, Basic €99.99, Ultra €199.99 and AI chat €49.99 per month, excluding VAT. The plans differ in the number of agents and the size of the knowledge base and storage, and their features are the same. Conversations and resolutions are not billed, and for AI you pay your own provider at its price.
Only by a setting. In AI chat the website chat works without agents from the start: the bot answers on its own and does not promise a human. The limits are the same. When you need agents, turn off this mode in the panel, and the bot starts handing questions to people.
You pay the AI provider directly with your own key, with no markup from us. A conversation on an inexpensive model costs about $0.007, roughly $7 per thousand conversations. Our model is free, but it is for trying things out: we do not guarantee its availability.
Yes, for a separate fee: we set up a gateway to a GPU in your office or on your server. The model runs on your side, and no AI supplier sees the conversations. The system and database still run on our server in the EU.
The request goes to the next provider in the list, and the failed one returns to the queue on its own once it works again. If there is no backup provider, the admin panel warns you in advance, and in the meantime questions go to your agents.
Every company gets a separate installation with its own database on a server in the EU, in Amsterdam, hosted by Railway. Railway is a US company and, under its data processing agreement, also processes the data in the US under the EU-U.S. Data Privacy Framework. The AI model processes requests where the provider you chose operates, or on your side if the model is your own.
The system runs on Railway infrastructure, and its current status is shown at status.railway.com. We do not have a separate availability agreement yet. Our free model runs without guarantees, so for real work connect your own provider and a backup one.
No. The bot understands voice messages in the website chat, WhatsApp and Telegram: it turns them into text and answers as if the question were typed. If a customer asks for a call, the bot passes their number to your agents. There is no voice bot and no telephony.
For email conversations with customers you need your own mailbox, and a regular Gmail will do. For emails to agents your own mail is not required: turn on the EPAV mailbox in the cabinet, 10 emails a day are free.
With your own AI assistant over MCP, and it's free: the knowledge server is included in every plan, with no fee per action. The assistant shows a plan, asks about conflicts and writes only after you confirm. Elsewhere MCP often costs extra, only reads or isn't there yet: Freshdesk charges $15 per 1,000 actions beyond the plan, Help Scout's server is read-only, and Zendesk has only announced its server so far. You can also do it without AI: export the articles to Markdown or HTML and upload them as files.
No. The bot answers from the knowledge base you maintain, and the model is not fine-tuned on it. The system can suggest new articles from resolved conversations, but they enter the knowledge base only after approval.
At 80% of the knowledge base or storage limit we warn you by email and in the cabinet. If you outgrow your plan, move to the next one. There is no plan above Ultra, so on Ultra you buy extra limits: €10 a month per 10 agents, €5 per thousand knowledge base chunks (about 300 articles) and €1 per gigabyte of storage, excluding VAT. The overage invoice arrives at the start of the next month. We do not switch anything off and do not throttle.
Register in the cabinet at my.epavai.com: your platform is assigned right away, no card needed. The bot answers on our model until you connect your own provider.
It still does. GDPR applies to any processing of personal data: contracts with processors, informing customers and their rights remain your responsibility. Your own model settles a different question: no AI supplier receives the conversations.
Still have questions?
Tell us about your support: how many conversations per month, which channels, where your knowledge lives today, whether you need your own model. We will reply within one business day.
- ✉ [email protected]
- 🛡 Data protection: [email protected]
- ⏱ Mon-Fri 09:00-18:00 EET · reply within ~1 business day
Sepapaja tn 6, 15551 Tallinn, Harju Maakond, Estonia