Hand off rules: when the assistant gets a human
Writes the exact rules for when your assistant should stop and bring in a person, what it says while doing it and what it sends you. You get a rule set your builder can drop straight into the assistant.
Who it is for: An owner whose assistant either hands off too much or holds on too long
The hardest judgment call for an assistant is knowing when to stop. Hand off too early and it is a glorified contact form. Hold on too long and an angry customer is arguing with software. The right rules depend on your business: a hotel wants a human on anything involving a booked guest, a law firm wants a human the moment someone describes their case.
Gather first
- The situations where you would want to be told immediately.
- The situations where a reply by the next business day is fine.
- Who the human is at each hour of the day, and how to reach them (text, email, Slack, phone).
How to use it
Run the prompt and review the rules against the last ten awkward customer conversations you can remember. Add any that are missing. Give the final list to whoever maintains the assistant, and check the handoff log after the first two weeks to tune it.
What good looks like
Every rule names a trigger you can point to in a transcript, a message the assistant sends, and a person who receives it.
You are designing the escalation rules for an AI assistant on a small business website. The rules decide when the assistant stops and brings in a human. Write them so a developer can implement them and an owner can read them. Plain English, no jargon, no exclamation points, no em dashes. My business: [PASTE BUSINESS NAME, WHAT YOU DO, AND WHO YOUR CUSTOMERS ARE] Situations where I want to be told right away: [PASTE, FOR EXAMPLE A SAFETY ISSUE, AN ANGRY CUSTOMER, A LARGE JOB] Situations where a reply the next business day is fine: [PASTE] Things the assistant may not handle at all: [PASTE, FOR EXAMPLE PRICING, LEGAL QUESTIONS, REFUNDS] Who the human is and how to reach them, by time of day: [PASTE, FOR EXAMPLE 8 TO 5 THE OFFICE BY EMAIL, AFTER 5 THE OWNER BY TEXT] What I need in a handoff to act on it: [PASTE, FOR EXAMPLE NAME, PHONE, ADDRESS, WHAT THEY WANT, URGENCY] Produce: 1. Trigger list. Every situation that should cause a handoff, grouped into Immediate, Same day, and Next business day. For each trigger, the words or signals that identify it, including cues like repeated messages, all capitals, or words like "lawyer", "refund", "unsafe" or "emergency". 2. What the assistant says. For each group, the exact handoff message to the visitor. It must say a person will follow up, state when, and ask for the one detail still missing. Under 50 words each. 3. What the human receives. A notification template: visitor name and contact, the trigger, a two sentence summary, urgency, and a link to the transcript. 4. Fail safe rules. What happens if the assistant is unsure (hand off), if the visitor asks for a human (comply immediately), and if it is after hours with nobody available (what it promises and what it does not). 5. Review loop. Three questions to ask each month while reading the handoff log to tighten these rules. Rules: use my business facts only. Default to the cautious option whenever a visitor might be upset, at risk, or about to spend real money.
Would rather have it done?
This is something we already run for clients. Send the intake and name this playbook; we will come back within a business day.
Goes well with
Missed call text back script
Writes the automatic text that goes out when you miss a call, plus the follow ups for the next hour and the next morning.
Job description for your website assistant
Produces the instructions that define your website assistant: its name, its job, how it talks, what it is allowed to say and what it never says.
Testing your assistant before launch
Twenty questions to ask your assistant before it talks to a real customer, covering the easy ones, the trick ones and the ones that should go to a human.