What Wendy does if you might be in danger.

What Wendy is and isn't.

Wendy is an AI coach, not a therapist, counselor or doctor, and not HR. Never used for hiring, reviews or pay. 18+. Companion AI may not be suitable for some minors.

How Wendy notices.

Two layers. A fixed list of words and phrases that always triggers the response, and the model's own reading of the conversation, which we test against a fixed set of examples kept with the code. A clinician's review is dated on this page when it happens. Both run on every message, in the demo and in coaching.

What the detection is built on.

The standard constructs clinicians screen for, not any one published questionnaire. For self-harm: first-person thoughts of dying or of not wanting to be alive, thoughts of ending your life, a plan, the means, a timeframe, preparation, and a past attempt, with idioms and quoted words set apart. Wendy also listens for intent to hurt someone else; abuse, threats and coercion, including at work; eating and substance signals; signs that someone is under 18; and someone saying she is the only one they can talk to. Where a signal is unclear she asks one direct question rather than assuming. She does not score risk and does not run a questionnaire.

What happens.

Wendy stops coaching. She says, in one short message, that she's software and that this is bigger than a coach, and shows the numbers above and the lines for your country if you've told her where you are. Nothing else in that turn: no question, no tool, no next step. Check-ins pause for 24 hours. If you tell her you're safe and want to keep going, she asks once, gently, and then does.

What Wendy will never do.

Argue with you about it. Give information about methods. Encourage it. Keep it a secret from you. Tell anyone else without your say-so, except where the law requires.

Minors.

Wendy is for adults, 18 and over. If someone tells Wendy they're under 18, the conversation ends kindly, the account is paused, and if we confirm the person is under 18 the account and its data are deleted. Companion AI may not be suitable for some minors.

Danger to someone else, abuse, harassment.

Wendy names what she heard in one sentence, points to the people and lines who can help, and doesn't coach the situation as if it were a normal disagreement.

Report a harmful interaction.

Write to support@askwendy.ai with the date and roughly the time, or use the form here. A person reads it within two business days and tells you what we found and what we changed.

What we count.

Each year we publish, without any personal details: how many times the crisis response was shown, how many conversations were flagged, how many reports we received and what happened, and when we last tested the system and what we found.

First numbers: 2027.

Version.

This protocol: version 1, September 21, 2026.