Train it, test it, switch it on per channel.
8 answers
You give it your own material; it answers from that, not from the open internet.
Upload PDFs (catalogues, price lists, policies, FAQs), and add plain-language context about how you work. The agent answers from that material and says it does not know rather than inventing an answer.
Update the source document and the agent updates with it. There is nothing to retrain by hand.
A sandbox that runs the real agent with no messages sent and no records written.
Ask it the ten questions your customers actually ask. Nothing is delivered, nothing is logged against a contact, and the toggle for live replies stays where it is.
This is the fastest way to find the gap in your uploaded material, a wrong answer in the sandbox is a missing paragraph in your PDF.
WhatsApp, Instagram, Facebook and the website widget are switched on independently.
Many teams start with the website widget, where the stakes are lowest, then add WhatsApp once the answers read well.
Leaving the channel list empty means the agent replies everywhere it is able to.
Separate context, separate knowledge base, separate triggers.
Website visitors are usually strangers comparing options; WhatsApp contacts are usually existing customers. They deserve different answers, so the widget keeps its own material.
Set the widget's context and upload its documents from the website section rather than the main agent settings.
The agent can answer with structured blocks, not just a wall of text.
Where it helps (plans, options, next steps), the agent renders tiles, chips and links instead of a paragraph. Links are restricted to URLs found in your own uploaded material.
You do not configure any of this per company; it comes from the documents you upload.
Your plan includes an AI credit balance; it accrues and does not expire.
Every AI action (replies, generated content, checks), draws on one shared balance. Subscribing and each renewal add credit, and unused credit carries forward.
At zero, AI features stop quietly rather than sending a broken reply, and the account owner is emailed. Nothing else in the platform is affected.
On explicit request, on repeated failure, or on a rule you set.
If a customer asks for a person, or the agent cannot answer from your material, the chat stays in the inbox for a human, the AI does not keep guessing.
You can also send specific topics straight to a human with a trigger, which is how most teams handle pricing negotiations and complaints.
Almost always a gap in the source material, not a model problem.
Reproduce the question in the sandbox, then look at what your uploaded documents actually say about it. Vague or contradictory source text produces vague or contradictory answers.
Add the missing fact, or state the rule explicitly in the context box, and re-test. Changes apply to new conversations immediately.
Tell us the number, the time and what you expected. That is usually enough for us to spot it.