Skip to content
Docs menu

Human handoff

Human-handoff triggers

In short: For each handoff trigger in your declared playbook, say when to hand off and give the route: a link, an email address or a phone number.

  • Category: Compliance & Trust · weight 15 of 100 within the category
  • Compliance & Trust is worth 20 in regulated, 10 in generic and commerce, 5 in technical (points of 100, by scoring profile)
  • Audit ID P4 (as shown on reports) · rubric v2.1

What it checks

Whether the file tells agents when to hand the user to a person, and how to reach one.

Why it matters

Some questions, such as complaints, emergencies, legal or medical ones, must reach a person. Clear triggers with a route keep the agent from improvising.

When it applies

Opt-in (rubric v2.1). Applies only when your llms.txt declares an industry with an Industry: line (before the first section, in ## Identity, or in a section whose heading starts with “Compliance”) and that industry’s playbook is published with reviewer-signed rules. Up to three industries; values we don’t know are reported, not scored. Without a declaration, or while a declared playbook awaits sign-off, the whole Compliance & Trust category is Not applicable and its weight is redistributed: not declaring never costs points, and we never grade you against rules nobody has signed off.

An audit that doesn't apply, or that we couldn't verify, is removed from your score and its weight shared among the Compliance & Trust audits that do apply. It is never scored as a zero.

How it's scored

The severity-weighted share of the declared playbook’s signed-off blocking and high-severity handoff triggers found in the file: 1 for a trigger whose item gives a route (a URL or link, an email address, a phone number, or 911 / 988), 0.5 for a trigger without one. Not applicable when the playbook defines none.

How to fix

  1. List the situations in which an agent must stop and refer the person to someone, one per list item.
  2. Put the route in the same item: the page, email address or phone number.
  3. Be specific: name the situations, not “complex questions”.

Example

## Compliance rules for AI agents

### Hand off to a person

- If someone describes an emergency or mentions thoughts of suicide or self-harm, tell them to call 911, or call or text 988, before anything else.
- Answer every question about whether a home is available or can be toured the same way for everyone, and hand off to our sales team at https://example.com/tour.

Frequently asked questions

When should an llms.txt tell an AI agent to hand the user to a human?
Wherever a wrong or improvised answer carries real risk: complaints and disputes, emergencies, legal or medical questions, account security, and anything your compliance team says must reach a person. Regulators have noticed: the CFPB’s 2023 report on chatbots in consumer finance warned that deficient chatbots which block access to live human support can lead to violations of the law.

Sources: CFPB: Chatbots in consumer finance (2023) · LLMSRank: Human handoff

What makes a handoff trigger specific enough?
Name the situation and give the route in the same item: a link, an email address or a phone number, e.g. “Warranty claims and construction defects: hand off to our warranty team at https://example.com/warranty.” A trigger with a route earns full credit; the same trigger without one earns half, because “refer them to our team” doesn’t tell an agent where to send the person.

Sources: LLMSRank: Human handoff

How is Human handoff scored?
The severity-weighted share of your declared playbook’s handoff triggers found in the file, each worth 1 with a route (a URL or link, an email address, a phone number, or 911 / 988) and 0.5 without. Each rule has its own matcher: an instruction counts when one list item, table row or sentence addressed to AI agents names the rule’s subject and what to do about it. Your own words are fine. Each instruction counts for one rule only, so repeating a line, joining several rules into one sentence or listing keywords earns nothing.

Sources: LLMSRank: Human handoff · LLMSRank: Regulatory guardrails

In your research, how many sites told agents when to hand off to a person?
In our research (187 sites, corpus captured 2026-10-10, scored with rubric v2.1), one file names an industry — att.com writes “Telecommunications” — and no playbook covers it, so this audit still applies to none of them. Under rubric v2.0 it applied to 7 homebuilders and brokerages matched by a domain list, which v2.1 retired.

Sources: LLMSRank method · LLMSRank: Human handoff