The handoff is the product

Our own number is public: agents resolve up to 70% of conversations end to end. Which means the honest question about any deployment is what happens to the other 30%. That is where trust is won or lost — with customers and with your own team.

No cold starts

A handoff in Nimbus is not a transfer to a queue. It is a package:

• The full transcript, including what the agent already tried.
• The evidence it gathered — the invoice it checked, the doc it cited, the account state it read.
• A suggested next step, marked as a suggestion, with the rule that stopped the agent from taking it autonomously.

Your agent (the human kind) opens the conversation mid-stride instead of asking the customer to start over. Median human first-response after a handoff drops from minutes to seconds, because the reading is already done.

Escalation rules you write

When to hand off is policy, not model temperament. Teams write rules like:

1. After two failed resolution attempts — never let a bot argue with a customer.
2. On any sentiment signal past a threshold — frustration goes to humans, fast.
3. On any action outside its guardrails — the evidence arrives with the request.

The metric to watch

Track post-handoff CSAT separately. If customers who reached a human score lower than your pre-AI baseline, your handoffs are arriving cold and the deployment is borrowing goodwill. Ours runs at 4.8 — slightly above the human-only baseline — which is the whole argument for treating the handoff as the product, not the exception path.