We built support that knows when the person is in charge.
An agent handles the conversation until a person takes over in the same thread. When the work returns to the agent, its memory stays with it.
- One agent, WhatsApp + web chat
- Support knowledge base
- Memory across handoffs
- Agent lane + human lane
- Same-thread takeover
- Late AI replies blocked
- Audited run log
Customers shouldn't have to start again.
YoboLabs customers write in on WhatsApp and web chat. The job: one agent answers first on every channel, and a person can take over the same conversation and hand it back, without the customer repeating themselves. It runs in its own messaging service, so more than one product can use it.
One agent on every channel, with a person one step away.
- One agent, every channelWhatsApp and web chat feed one inbound pipeline. Each line is bound to the agent that answers it.
- Answers from the knowledge baseThe agent answers from the YoboLabs knowledge base, and asks a clarifying question when it needs one.
- Memory across the threadPersistent memory holds the conversation. When a person hands it back, the agent picks up where they left off.
- An agent lane and a human laneOn the board, an escalated conversation moves from the agent lane to the human lane. The operator replies from the CRM inbox.
- Handoff and hand-backThe agent can ask for a person itself, or the customer can. The customer is told a person is coming, and the same thread carries on.
- Progress the customer can seeOn WhatsApp, the customer sees step-by-step progress and a typing indicator while the agent works.
One conversation, start to finish.
Every channel feeds one inbound pipeline. A conversation is either AI-driven or human-driven, and only the side in control may write to the customer.
- AI answers
- Escalated: holding line sent
- A person answers, same thread
- Handed back
- AI answers, same memory
Lines marked real are recorded text or recorded behaviour. Lines marked illustrative show the shape only.
Who may speak to the customer, and when.
- Late repliesDropped, and audited. An AI reply is written only if the conversation is still AI-driven and this is the current run.
- Holding lineAlways sent. When a chat is escalated, the customer is told a person is coming.
- Same threadNo restart. The person takes over the same conversation. On hand-back the AI resumes with the same memory.
- Who can askThe agent can ask for a person itself, or the customer can. Each escalation opens one episode.
- One winnerOne operator wins a take-over. Anyone trying at the same moment sees a notice, and nothing is written.
- PermissionsTaking over and replying are separate permissions. The operator's identity is checked on the server, never taken from the browser.
- Time limit5 minutes. An AI turn that runs longer is released, so the conversation keeps moving.
- Open chatsAny recent chat. An operator can take over any chat where the customer wrote in the last 24 hours, even if nobody escalated it.
In production 13 days after the first commit.
The first real multi-turn WhatsApp conversation ran on day 3, with context held across turns. Human takeover and hand-back were proven in browser tests on 7 July, and the service went to production on 14 July. A web chat now goes to the agent, escalates and lands in the CRM inbox.
What it runs on.
- Channels
- WhatsApp and web chat
- Platform
- Cadra
- Operator desk
- CRM inbox
- Hand-back
- Same thread, same memory
Milestones
- Spec reviewed and the first code committed.
- First real multi-turn WhatsApp conversation through the new service.
- Human takeover and hand-back proven in the browser.
- Messaging service live in production.
- WhatsApp shows step-by-step progress and typing while the agent works.
- “Open” chip in the operator inbox, and take-over of any recent chat, live in production.