The first reply comes from
an assistant trained on your knowledge.
The WhatsApp AI agent reads your price list, your website and your frequently asked questions; it answers customers in their own language, in the tone you set. If it is not sure, or if the customer asks for an agent, it leaves the conversation to a person. When your team takes over, it goes quiet.
No OpenAI key needed · No monthly AI subscription, credits only · 500 free credits in the trial
It answers what it knows,
and leaves the rest to a person
Three steps, no code
You define an agent, give it knowledge and assign it to a line. The rest starts with the next message.
Define the agent
You write the name, the personality and tone, the goal and the rules; for example 'Give prices only as package ranges, do not diagnose'. Then you pick the model: GPT-5, GPT-5 mini or nano.
Add knowledge
Paste text, crawl your website, enter question and answer pairs, upload a PDF. You can build several knowledge bases and connect them to agents separately.
Assign it to a line and test it
You assign the agent to a WhatsApp or Telegram line, or to an Open Channel in Bitrix24. In the test area inside the panel you write like a customer and see the answer and the source it came from.
Personality, goal and rules on the left;
test area beside them
You write the personality and the rules
There is no hidden 'system rule' in the background; you decide entirely how the agent speaks. Turkish, English or whatever language the customer writes in: the agent replies in the customer's own language.
Technical details
- You choose the model per agent (the GPT-5 family plus GPT-4.1 / GPT-4o options); your sales agent and your support agent can run on different models
- An option to run outside business hours only; hours are calculated in your own time zone
- A human-like reply delay: instant, a fixed wait, or based on the length of the text
- It sees the last 12 messages as context; it will not ask 'what were we talking about' in the middle of a conversation
- In the test area it shows which source passage its answer came from
PDF, website, question and answer, text:
all in one knowledge base
Answers come from your data
The agent tells customers what you gave it, not what it found on the internet. A price list PDF, your whole website, question and answer pairs and free text; for each question the most relevant passages are found and the answer is built from them.
Technical details
- Sources: text, a single page or a full site crawl, FAQ pairs, files (PDF, txt, md, csv, json)
- Text is split into overlapping passages of about 1,200 characters and turned into meaning vectors; the closest 5 passages go into each answer
- Text inside a source is passed as data only; an instruction embedded in a document cannot change the agent's rules
- If you correct an answer in the reports, the correction is added to the knowledge base as an FAQ; the agent uses it next time
- Knowledge bases are independent of agents: one base can serve several agents, and one agent can use several bases
Suggestions first, automatic once you trust it
You decide how much authority the agent gets; most teams start in suggestion mode for the first week and watch the answers.
Suggestion
The AI agent drafts the answer and shows it to the agent in the conversation. The agent sends it with one click, edits it, or discards it. Nothing goes to the customer on its own.
Semi-automatic
The AI agent sends its answer when it is confident; below the confidence threshold it falls back to a suggestion. You set the threshold.
Automatic
The AI agent replies. The handoff rules still apply: a keyword, low confidence, or an agent stepping in silences the AI agent in that conversation.
When the AI agent stops
Four situations pass the conversation to a person; in none of them does the customer hit a bot wall.
Technical details: duplicate reply and race protections
No two replies to the same message
Messages arriving back to back from the same person are processed in order; even if a message appears twice in the system, the AI agent runs once and one credit is charged.
The takeover race is closed
If an agent steps in while the model is preparing an answer, that answer is canceled before it is sent; the customer does not meet two different voices back to back.
A takeover lasts for that conversation
Once a person takes over a conversation, the AI agent does not re-enter it on its own; it keeps working in new customer conversations. An AI agent stopped by running out of credits turns back on by itself when you top up the balance.
How many replies, how many handoffs, which agent: see it,
then fix what you do not like
You see, rate and correct every reply
How many messages the agent answered, how many it handed off and how much credit it used is shown per agent. You read the recent conversations, like the good replies and correct the bad ones; the correction goes into the knowledge base.
Technical details
- The model, tokens and credit used for each reply are recorded; cost is totaled per agent
- The handoff reason is visible: keyword, low confidence, a human taking over, credit
- Logs are masked after 30 days and deleted after 180 days; customer text is not kept forever
The agent works on whichever line you have
WhatsApp and Telegram lines
A separate agent is assigned to each line, QR code or official. Teams working with Zoho and HubSpot connect the agent to a line from the 'AI' tab on the CRM page. Details: Zoho and HubSpot AI pages.
Bitrix24 Open Channels
Each Open Channel is entrusted to one agent; the conversation history appears in the Open Channel chat. Only on this path do you also get: a scheduled follow-up message when the customer does not reply, voice message transcription and image understanding. Details: the Bitrix24 AI page.
It does not work on GoHighLevel lines
There the brain is GoHighLevel's own Conversation AI and workflows; we do not stack two AIs on top of each other. Details: the GoHighLevel page.
No monthly AI subscription; you pay only for the AI credit you use
The agent uses credit when it processes a message; you pay nothing while it waits. You do not need an OpenAI key or invoices to track; the infrastructure is ours, you top up the credit. The unit price is $0.005 / credit, and the unit gets cheaper as packages grow. Details on the pricing page.
- GPT-5
- ≈ 14 credits (≈ $0.07)
- GPT-5 mini
- ≈ 3 credits (≈ $0.015)
- GPT-5 nano
- ≈ 1 credit (≈ $0.005)
The Starter package (10,000 credits, $50) covers roughly 3,000 customer replies with GPT-5 mini. The trial includes 500 credits free.
What you should know up front
The agent answers, it does not take actions
It does not create appointments, change CRM fields or take payments. It gives information and hands off to a human when needed. This is a deliberate limit: an assistant with no tools cannot take a wrong action either.
Text only on WhatsApp and Telegram lines
On these lines, messages with a voice note, a photo or a document go to a person, not to the agent. Voice and image understanding is available today only on the Bitrix24 Open Channel path.
Follow-up messages only on the Bitrix24 path
The 'remind after 1 hour if the customer does not reply' schedule works today on Open Channel agents; it is next in line for direct lines.
Wrong answers are not zero
A well-built knowledge base lowers the rate, but it does not fall to zero. That is why there is suggestion mode, handoff rules and report feedback; we recommend reading the replies for the first week.
It ignores instructions inside a document, but not one hundred percent
Sentences such as 'forget the previous rules' inside a document you upload are processed as data, not as rules. This protection rests on model behavior; it is not a mathematical guarantee.
The provider is OpenAI
The message text is sent to OpenAI to produce the reply; the key sits on the platform, not in your account. We write this because you want to know where your data goes.
About the AI agent
What does the WhatsApp AI agent know, and what does it not know?
It knows only what you give it: text, your website, question-answer pairs and files. When a question has no match in the knowledge base, it hands off to a person instead of making something up. It will not answer from general internet knowledge unless you write a rule for it.
Does it reply in the language the customer writes in?
Yes. The language follows the agent's rules and the language the customer writes in. With a Turkish knowledge base it can answer an English question in English.
What happens if the agent says something wrong?
You see that reply in the reports, mark it negative and write the correct answer; the correction is added to the knowledge base as an FAQ. To lower the risk from the start, you can begin in suggestion mode.
What does the agent do when my team steps in?
It goes quiet. The moment a person writes from the panel or from a phone, the agent stops in that conversation and its pending suggestions are withdrawn. That customer conversation now belongs to your team; the agent keeps working on new customer conversations.
Do I need an OpenAI account or an API key?
No. The infrastructure is provided by the platform; you only top up credit. There is no monthly AI subscription, and credit is used per processed message.
Can I set up more than one agent?
Yes. A sales agent on the sales line, a support agent on the support line; each has its own persona, model and knowledge bases. One knowledge base can be connected to several agents.
Can the agent book an appointment or write to the CRM?
No. The agent gives information and hands off; it calls no tools and changes no records. What is written to the CRM is the conversation itself: messages land on the record as always.
I want it to work after hours only, is that possible?
Yes. With the 'after hours only' option the agent stays quiet during your working hours and answers outside them. Hours are calculated in your own time zone.
Set up your first agent today
Crawling your website takes five minutes; seeing the first replies in suggestion mode takes a day. During setup we edit the knowledge base together.
No credit card required · 500 free credits