Quick answerThe agent needs the full context of what the human already tried and why the conversation is being handed back, not just the raw transcript. Open by acknowledging the escalation explicitly rather than picking up as if nothing happened, do not repeat any question or solution the human already covered, and set a low threshold for re-escalating to a human again if the hostility does not de-escalate quickly, since sending a frustrated customer back to the agent a second time without real improvement compounds the original frustration.
Why this is the reverse of the more commonly designed direction
Most teams design carefully for the warm handoff from the agent to a human, making sure the human gets full context when a conversation escalates upward. Far fewer design deliberately for the reverse: a conversation coming back down to the agent after a human already tried and, for whatever reason, handed it back. That direction gets less attention because it happens less often, but it is arguably higher risk, since the customer in this scenario has already been escalated once and did not get full resolution, which is a worse starting emotional state than a typical first contact. The context problem is closely related to what already has to be solved for keeping conversation context intact when a customer switches from chat to phone, just applied to a channel switch between a human and the agent instead of between two channels.
Why this is not the same as a routine sentiment escalation
This is a related but distinct scenario from detecting and de-escalating an upset customer mid conversation, which is about the agent recognizing rising frustration in a conversation it has owned from the start. Here the agent is inheriting a conversation that was already hot when a human handled it, and is receiving it back for reasons that may have nothing to do with the agent's own performance, for example the human's shift ending or the human determining the remaining steps are procedural rather than requiring further human judgment. The agent needs to know which of those it is walking into.
What context has to travel with the handback
The minimum viable package includes what the human agent tried, what the customer's stated or apparent reason for dissatisfaction is, and an explicit reason the conversation is coming back to the AI agent rather than staying with a human. Without that last piece, the AI agent cannot distinguish between a handback that means the hard part is done and a handback that means the human could not resolve it either, and treating those two situations the same way will go wrong for one of them.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
How the agent should open the conversation
Open by naming the situation plainly: acknowledging that the customer already spoke with a human, briefly confirming what was already covered so the customer does not have to repeat themselves, and stating clearly what happens next. Picking back up with a generic greeting, as if this were a fresh conversation, is one of the fastest ways to make an already frustrated customer feel like the company is not tracking its own case history.
Setting a fast re-escalation threshold
Because this customer has already been escalated once without full resolution, the bar for escalating again should be lower than it would be for a first time contact, not the same or higher. If the hostility does not measurably improve within the first exchange or two, treat that as a strong signal to route back to a human immediately rather than continuing to attempt automated resolution. A second failed automated attempt on top of an already escalated conversation does measurably more damage to trust than the same failure would in a first contact scenario.
FAQ
Should the agent apologize on the human agent's behalf if the human could not resolve it?
The agent should acknowledge the situation honestly without assigning blame to the human, focusing on what happens next rather than relitigating why the prior attempt did not fully resolve things.
Should this scenario be logged differently than a normal escalation?
Yes, tag it distinctly from both a standard escalation and a standard resolution, since a pattern of frequent handbacks in the same category is a strong signal that the underlying issue type may need a policy or capability fix rather than repeated one off handling. Feeding these tagged conversations into the same ongoing human QA transcript sampling process you already run is a low effort way to make sure handback conversations get reviewed rather than falling through a gap between two ownership handoffs.