Add your text, website, FAQs and files; the AI agent replies only from your knowledge (RAG). It runs in suggestion, semi-automatic or fully automatic mode, and hands off to a real agent when unsure. It works directly on WhatsApp and Telegram, and through Bitrix24 Open Channels on Instagram, Facebook and web chat too.
✓ Answers from your own data · ✓ No making things up · ✓ Hands off to a human when unsure
Not a generic chatbot; a trainable customer agent that answers only from your data.
An ordinary AI chatbot talks from generic knowledge it learned off the internet, and happily makes things up when it doesn't know. The FlexInbox AI agent works on a RAG (Retrieval-Augmented Generation) architecture instead: before producing a reply it retrieves the relevant pieces from your knowledge base and grounds the answer only in those sources. So the agent conveys your products, prices, shipping and return rules from the text you wrote, and when no source exists it hands off to a real agent rather than inventing an answer.
Feeding the knowledge base is entirely in your hands and needs no technical skill. You can paste plain text, crawl your website (a single page or the whole site), enter FAQs as question-answer pairs, or upload files (PDF, DOCX, CSV). When you update a source the agent instantly starts replying with the current information; when a campaign, price or policy changes, all you do is edit the source.
You define not only what the agent knows but also how it speaks. You set its personality, goal, tone and the rules for when it should hand off. Before going live you try the agent in the playground, see its answers, and fine-tune the knowledge base or rules if needed.
It doesn't just read text; it understands images and voice, follows up, measures, and gets better over time.
Fed by text, website crawl, FAQs and files (PDF/DOCX/CSV). It blends multiple sources in one agent, and answers update as your content updates.
It interprets photos a customer sends (a product, receipt or screenshot) and transcribes voice messages, replying in the same flow.
With a reply delay setting, the agent answers at the rhythm of a real agent rather than instantly like a robot; the conversation feels natural.
It sends automatic follow-ups to customers who didn't reply or dropped off, reopening cooling conversations on its own.
See message volume, tokens spent, estimated cost, and AI reply and human-handover rates broken down in the panel.
Define several agents per organization and assign each line (number/Open Channel) to its own agent and knowledge base, so different brands and teams never mix.
You decide how much authority the agent gets; let it draft, reply by confidence, or run end to end.
The agent drafts a reply for every message but doesn't send it. The operator sees the draft and sends it as is, edits it, or discards it. It's the most controlled way to introduce AI to your team and measure reply quality with confidence.
The agent evaluates how confident it is in the reply it produced. Replies above the confidence threshold you set are sent automatically; anything below, the uncertain cases, are left to the operator as a suggestion. It answers routine questions without losing speed and calls a human for the hard ones.
The agent runs the conversation from start to finish, serving uninterrupted even after hours. When it's unsure, on a triggering keyword, or when the customer explicitly asks for an agent, it automatically hands off to a real person. A 24/7 front line, with a safe handover to a human when needed.
The AI and your team work in the same chat without overlapping; control is yours at every moment.
When the agent is unsure, on the keywords you set, or when the customer says "I want a human", it automatically transfers the conversation to a real agent.
The moment the operator types into that chat from the panel or the phone, the AI goes quiet; two voices never clash in one conversation and the customer sees a single flow.
Give replies a 👍/👎 or write the correct answer; corrections are added to the knowledge base as FAQs and the agent gets more accurate over time.
AI usage is metered with a separate AI Assistant Package that runs independently of your lines. Each reply uses your package allowance based on the model you choose and the length of the answer: a short reply uses little, a long one or one from a more advanced model uses more. When your package runs out the agent stops and hands the conversation to a human, so there are no surprise bills. The provider is OpenAI and you can use models from the GPT-5 family.
You can upgrade your package whenever you like or set up auto-renew. You see exactly how much of your package each message used, along with the tokens spent and estimated cost, broken down in the panel.
The agent is fed only from your knowledge base via RAG and grounds its answers in the source you wrote. When no suitable source exists it can be directed to hand off to a real agent instead of inventing an answer. You can also review and correct every reply from the reports, making it even more accurate over time.
No. You feed the knowledge base by pasting text, crawling your website, entering question-answer FAQs, or uploading files (PDF, DOCX, CSV). Then you write the agent's personality and rules, try it in the playground, and go live. It's all done from the panel, with no coding.
The moment the operator types into that chat from the panel or the phone, the AI goes quiet (human takeover). The agent also hands off when unsure, on a keyword, or when the customer asks. When you give replies a 👍/👎 and write the correct one, the correction is added to the knowledge base so the same mistake isn't repeated.
It works directly on WhatsApp and Telegram. Through Bitrix24 Open Channels it can also answer across all channels connected to Bitrix, such as WhatsApp, Telegram, Instagram, Facebook and web chat.
Yes. You can define several agents per organization and assign each line (number or Open Channel) to its own agent and knowledge base. That way your different brands, branches or teams run with independent personalities and knowledge.
AI is metered by usage through an AI Assistant Package that is separate from your lines. Each reply draws from your package based on the model you choose and the answer's length; a short reply uses little, a long one on an advanced model uses more. When the package runs out the agent stops and hands off to a human, so there are no surprise bills.
An AI agent that speaks from your own knowledge, doesn't make things up, and hands off to a human when needed. Start free, no credit card required.
