The agent escalates
The agent itself decides the case calls for a person and uses its transfer-to-human tool — or the customer asks to speak with someone. The conversation becomes a card in the queue.
Not every conversation ends with the AI. When the subject calls for a human, the card moves to a queue, an operator takes it, and the platform pauses automatic replies in that chat. Same number, same history, no parallel conversation — and once it is closed, the agent goes back to answering.
Queues, operators, scoped supervisors, break types and an audit trail of the workday — all in the same console as your agents.
Camila took over · automatic replies paused
In a takeover, the same conversation carries on at the same number.
Four moves, all inside the same chat. There is no second channel for the human, and no second history for the person to read before replying.
The agent itself decides the case calls for a person and uses its transfer-to-human tool — or the customer asks to speak with someone. The conversation becomes a card in the queue.
When the card is assigned to a person, the platform pauses automatic replies in that chat. It isn't a setting someone has to remember to switch on: it is what taking the card does.
The operator opens the entire conversation — what the customer said, how the agent answered, what it looked up. Nobody asks the customer to explain the problem again.
Close the conversation and the chat returns to the agent, which keeps answering that customer's next messages. Transferring changes the card's owner and preserves the record of who said what.
Human and AI in the same chat: taking the card pauses the bot, closing it hands service back — no parallel conversation, no lost history.
The screen below shows the service inbox in the moment right after the handoff. At the top of the module are the active conversations and the pills that separate WhatsApp, webchat and the agent's own test chat; on the left, the conversation list with the open card highlighted; on the right, the customer's conversation with the phone number and channel in the header.
Inside the chat, the amber band states what happened: the AI was paused automatically because a person is answering, with the name of whoever took over and the actions to release the AI, transfer the card or close the conversation. Yesterday's history is still there — the customer's message and the agent's reply before the escalation — so nobody has to ask the customer to explain the problem again. Below it, the expired-window band warns that the customer's last message is more than 24 hours old and that, by WhatsApp's rule, only an approved template can go out from there.
Yesterday
Today
Across the top, the whole inbox: the conversations open right now, the shortcut into operator monitoring, and the pills counting how many conversations came in through each channel — WhatsApp, web chat and the agent's test chat — with the queue selector beside them. On the left, the conversations assigned to this person, the open card highlighted and an unread counter on the rest. Inside the chat, the amber band names whoever took over and offers the three ways out — hand the AI back, transfer the card, close the ticket — and the band below warns that this customer's 24-hour window has expired.
Whoever takes over starts with all of yesterday in front of them — the complaint, the agent's reply, the work order already tracked down — so the customer doesn't repeat the problem and nobody opens a side channel to piece the case back together. Because assigning the card is what pauses the AI, the informal habit of switching the bot off before replying disappears, and with it the risk of agent and person typing into the same chat. The expired-window notice shows up before the message is written, not after it's rejected.
When the ticket is closed, service returns to the AI — and the AI comes back knowing. The history is vectorized through embeddings and a worker extracts stable facts and customer traits, with a sense of time: the agent picks up aware that the visit was rescheduled and that a person stepped in along the way. A flow automation can return the contact to a specific node in the flowchart; that episodic, persistent memory is the subject of the patent filed with the USPTO.
The platform's call center structure exists to answer three questions: where the conversation goes, who is allowed to pick it up, and who is working right now.
Every slice of the operation gets its own queue. An escalated conversation arrives as a card and becomes someone's work — not a message lost in a shared inbox.
Every operator has their own active-chat panel, showing the conversations they own at that moment and the actions to reply, transfer and close.
A supervisor sees the queues they are responsible for — and only those. Scope is part of the account setup, not a verbal agreement.
The company defines the break reasons an operator can choose from. Every break is logged on the timeline for that day.
One screen shows which operators are connected and which chats are open at that moment. That is what lets a supervisor act during the shift instead of the next day.
Customer record, history by channel, AI-generated sentiment and tags, quick replies and conversation export: the operator works inside the intelligent CRM.
Five views of your human team's work, built on telemetry consolidated by a daily rollup. Each view can be emailed and scheduled.
| View | What the supervisor sees |
|---|---|
| Daily timeline | Online status, focus, activity, breaks, dropped connections and cards received per operator, hour by hour. |
| Operator comparison | Every operator from the same day side by side, on the same terms, so you can compare without building a spreadsheet. |
| Active hours and availability | How long each person was active and available to receive a card across the shift. |
| Operational workday | Browser opened and closed, breaks and chat windows by customer — the sequence of the day, not just the totals. |
| Workstation connectivity | How often connections dropped, for how long, and the longest interruption of the day — so a network problem doesn't get read as a performance problem. |
Telemetry records signals and counters — session, focus, break, dropped connection, card received. It does not record mouse coordinates or screen content. The monitoring notice shown to the operator is configurable by the company.
Chats per operator, human chats by period and customers by period, filtered by queue and by channel, with PDFs generated on the server, email delivery and recurring scheduling.
When service happens by voice, the phone conversations screen shows active calls, history by phone number and a turn-by-turn transcript of every call.
There is no global escalation rule on the platform. Each agent decides, based on the personality and the tools it was given — including the transfer-to-human tool.
That changes how you design the operation: instead of mapping every possible exit in a flowchart, you describe in plain language when that agent should bring in a person, and it judges case by case. A technical support agent and a collections agent can have completely different thresholds while working the same number.
The agent chooses the moment to bring in a person, but it does not choose what it is allowed to do: tools are declared one by one in its configuration, and whatever wasn't handed to it doesn't exist in the agent's world.
AI doesn't replace the contact center: it absorbs the repetitive volume and hands the operator the case that genuinely needs judgment — already read, already in context.
The conversation continues on the same number and in the same history. The moment the card is assigned to an operator, automatic replies are paused for that chat — the agent and the person are never talking at once. Whoever takes over decides whether to announce that they are human; the platform neither forces it nor hides it.
As long as the card belongs to a person, the bot stays paused in that chat. The real-time monitoring screen shows open chats by operator, and the supervisor can transfer the card to someone else or close the conversation — which hands it back to the AI agent.
No. The views are built on telemetry consolidated by a daily rollup: capture records signals and counters — session opened, window focus, break, dropped connection, card received — and never mouse coordinates or screen content.
The monitoring notice shown to the operator is configurable by the company, which decides how to communicate the practice to its team.
Queues, operators, real-time supervision and an audit trail of the workday, in the same console where your AI agents work. Tell us about your operation and we'll show you what the design looks like.