Passing context
The cardinal sin is making the customer repeat themselves: the handoff package, the agent-assist afterlife, and closing the loop when the human resolves it.
Every study of support experience finds the same top complaint, decades running: having to repeat yourself. A handoff that drops the customer into a queue with no context converts your agent's good work into the customer's wasted breath. The fix is a designed artifact:
If your team works tickets rather than live chat, 'handoff' = a well-formed ticket with the package attached; the 8-second test and the off-hours honesty from the last lesson apply identically — a human still opens that ticket cold and needs to say a useful first line from the summary alone.
The handoff package
- The one-line situation: what the customer needs, in the agent's best summary — 'Order HL-1042 shows delivered, customer says nothing arrived; wants replacement or refund.' Written for a human who has 8 seconds before they say hello.
- The facts gathered: order number, relevant account state, what the agent already looked up (with timestamps) — so the human doesn't re-run the same lookups in front of an already-impatient customer.
- What was already said: policies quoted, options offered, anything the agent committed to ('label emailed at 14:02'). The human must never contradict the agent unknowingly — contradiction reads as chaos, and chaos reads as 'this company doesn't know me.'
- The trigger and the mood: why the handoff fired (customer-asked / distress / low-confidence / policy / loop) and a one-word temperature (calm / frustrated / distressed). Routing hint included: the mailer-damage refund goes to the returns queue, not general.
- The full transcript, attached — the package is the summary on top of the transcript, never instead of it. Trust-but-verifiable, for the human's judgment calls.
After the handoff: the agent's second job
Handoff isn't the agent's exit — it's a role change. Agent-assist mode: the same retrieval and drafting that served customers now serves the human rep — suggested policy citations, a drafted reply they can edit, the account context assembled. (Support teams often get more measurable value from assist mode than from deflection, and it's the same system you already built — this is the resolution-agent territory the Agentic AI course develops.) And the loop closes: when the human resolves the case, the resolution gets tagged — because Module 5's improvement engine runs on knowing what the right answer turned out to be for every conversation the agent couldn't finish. Today's handoff transcript is next month's eval case, knowledge-base patch, or scope amendment.
Before launch, run 10 live handoff drills with your actual support team: agent conversation → package → human picks up mid-flow. Ask the reps two questions afterward: 'What did you have to ask that the package should have told you?' and 'Did anything in the package mislead you?' Their answers are the spec for package v2 — and the drill doubles as the support team's introduction to their new coworker, which Module 6's launch depends on going well.