When an AI support bot hands you off to a human, what should it actually tell them about you first?

This explores what an AI support agent should pass along when it hands a customer to a human, and which kinds of context tend to get lost on the way. The collection has no paper that studies handoff packets directly. Several findings from neighboring areas still point to a clear answer: the obvious package (transcript, issue category, account details) leaves out what most often decides whether the human can actually help.

The strongest evidence comes from a randomized field experiment on Alibaba's Taobao. When human workers took over after the AI failed, service quality recovered for technical problems the AI couldn't handle, but not for customers who were already frustrated. Timing mattered too: early handoffs worked, and late ones that came after frustration had built up did not When human workers take over from AI, does it actually fix the problem?. So the most important thing to transfer may not be what the problem is. It may be how the customer feels, how long they've been stuck, and how many failed attempts they've already sat through. A transcript technically contains all of this, but a human skimming it under time pressure won't see it unless it's flagged.

A second candidate is what the AI *doesn't* know. One study found that language models have no explicit record of what remains unknown about a user, which feeds sycophancy and made-up answers. Adding a list of labeled unknowns to the prompt cut harmful advice and hallucination roughly in half Do language models know what they don't know about users?. Applied to escalation, a good handoff would include a list of open questions: what the AI assumed, what it never confirmed, and where it guessed. A related finding about workplace AI shows why this has to be deliberate. When work passes through AI, signals of effort and uncertainty tend to vanish from the finished output, while polished, identity-bearing signals survive Which workplace cues survive AI mediation and which disappear?. A neat AI summary can look more certain than the conversation behind it actually was.

The handoff should also be designed rather than dumped, because AI context is built fresh for each request from shifting sources, unlike the stable state of conventional software How does AI context differ from conventional software context?. Which pieces reach the next party is a product decision, not something the model does on its own How do AI products decide what to remember between conversations?. Some of what shaped the AI's behavior is also invisible to the customer, such as system-prompt rules and policies Who actually controls what AI systems say?. A human agent who doesn't know the bot was told to refuse refunds can't explain to the customer why it refused. Research on delegation between agents adds one more point: passing along a short, evidence-grounded summary works better than passing everything Can delegation teach models to manage context more actively?.

Finally, escalation might be better treated as an ongoing partnership than a single hand-over. Work on human-agent systems argues that nobody knows the ideal moment to hand off, so good designs spread the decision across several touchpoints, such as co-planning, action guards, and verification When should human-agent systems ask for human help?. Others describe AI systems that assign tasks *to* humans, which raises the question of who is responsible for the case after the switch Can AI systems delegate tasks back to humans?. Taken together, the corpus suggests a handoff should carry four things: the customer's emotional state and history of failed attempts, an explicit list of unknowns, the hidden rules that shaped the AI's answers, and a clear statement of who now owns the case. The Taobao result adds that no handoff package can make up for escalating too late.