How does BotableX decide when to ask a human?
The bot always answers first. When its confidence is below your threshold it retries, then quietly asks your team for guidance while it keeps talking to the customer.
Updated 2 September 2026
- 01
A message arrives
On any channel. The bot reads it together with the conversation so far and the customer's record.
- 02
The bot answers
It retrieves from its Knowledge Banks and composes a reply in your voice.
- 03
Confidence is checked
Below your threshold, it refines the question and retrieves again. Up to twice.
- 04
It asks a teammate, silently
A consultation opens for your team. The customer sees the bot, not a queue.
- 05
Guidance comes back in the bot's voice
Your teammate types a note; the bot turns it into a proper reply.
- 06
A human can take over at any time
Only a person can order that. The bot never hands off on its own.
The rule that everything else follows
In BotableX the bot is always the one talking to the customer, unless a human explicitly takes over. That single rule produces two things customers notice: an answer arrives immediately, on every channel, at any hour; and nobody is ever told “let me transfer you” and left waiting.
It also produces one thing your team notices: the bot cannot decide on its own to push a conversation to a person. It can ask for help. It cannot hand off.
What happens on every message
When a message arrives, the bot reads it together with the whole conversation, the customer’s record if your CRM is connected, your Global Mind (who you are, how you speak, what is off limits), and its own Agent Brain (its instructions and limits). It retrieves the most relevant entries from its Knowledge Banks and composes a reply.
Before sending, the reply passes two checks. A factuality check confirms the reply is supported by what was retrieved; an unsupported claim is dropped, not softened. A confidence score, from 0 to 100, estimates how well the reply answers the question.
The confidence threshold
Each bot has a confidence threshold you set. The default is 80. If the score is at or above it, the reply goes out. If it is below, the bot does not send a weak answer. It rephrases the question, retrieves again, and re-scores, up to two more times. Many hard questions are resolved here, because the second retrieval often finds the entry the first one missed.
Raising the threshold makes the bot more cautious and consult more often. Lowering it makes the bot more self-reliant. Most teams start at the default and adjust after a week of watching the Learning activity.
Silent consultation
If confidence is still low after the retries, the bot opens a consultation. In Live Conversations your team sees the conversation flagged Needs help, with the bot’s own question about what it is missing. The customer sees none of this. The bot may tell them it is checking, in its own words, and keeps the conversation open.
A teammate reads the context and types guidance, as brief as “Refunds over 500 need a manager; offer to schedule a call.” The bot rewrites that into a reply that matches your voice and the customer’s language, and sends it. If the guidance is too vague to act on, the bot can ask the teammate a clarifying question first.
If nobody answers within a couple of minutes, the bot gracefully tells the customer it will follow up, and the consultation stays open for your team.
Takeover
At any moment a person can press Takeover. From then on the bot is silent on that conversation and the person types directly. Press Release to AI to hand it back; the bot resumes with full context. A takeover is the only thing that stops the bot, and only a human can order it. Every takeover and every piece of guidance is logged.
Email is a little different
On email you can set a channel to Consultation Only. The bot composes every reply but never sends one; a person reviews the draft and sends it, edits it, asks for a revision, or writes their own. This is the right mode for sensitive mailboxes or the first weeks with a new bot.