Design the handoff before the happy path

Most agent projects design the escalation last, as an error case. It is the opposite: escalation is the path where your agent either earns trust or destroys it, because it is the moment the customer was already frustrated.

Handing over a transcript is not handing over context

Dropping a human into a chat log and expecting them to read it is a worse experience than no agent at all. They are now doing the customer’s job of explaining, under time pressure, while the customer waits.

What the human needs is a briefing:

handoff
  conversation_id
  reason              -- low_confidence | user_asked
                      -- | policy_block | budget_exceeded | sentiment
  summary             -- what the customer wants, one paragraph
  verified_facts      -- account id, order ref, what was confirmed
  attempted           -- tools called and what they returned
  blocked_on          -- the specific thing that stopped it
  suggested_next

Generated at the point of escalation, shown above the transcript rather than instead of it. The human reads five lines and takes over knowing what has been established.

Escalate on the right signals

The customer asked. Immediately, without a retention attempt. Trying to keep someone who has asked for a human is the single most damaging pattern in this category.

Low confidence. The model is guessing rather than retrieving. Better to hand over than to answer plausibly and wrong.

Policy boundary. Anything the agent is not permitted to do — a refund over the ceiling, an account change, a complaint.

Repetition. The customer has asked the same thing twice. It did not work the first time.

Preserve what was verified

If the agent confirmed identity, the human should not ask again. If an order reference was validated against the system, carry it through as structured data rather than as a sentence in a summary.

That single detail — not re-verifying — is what makes a handoff feel like continuity instead of a restart.

Measure the handoff, not the deflection rate

Deflection rate rewards the wrong thing: an agent that refuses to escalate scores well and is hated. Track the resolution rate after handoff, and how often customers who were escalated come back. Those tell you whether the agent is helping or just absorbing the first minute of everyone’s frustration.