What this page is
This page is the protocol itself.
Several states require a company like ours to publish how it detects a crisis and what it does next. This is that document. It is not a summary of one.
Atticus is not a substitute for clinical care. Our software is not a doctor, a nurse, a therapist, or a counselor. Members always talk to software unless we say otherwise. The software says so itself, and it says so again whenever a member asks.
Atticus provides care management. We explain a member's own health information, we make recommendations, and we coordinate care. We do not diagnose mental health disorders and we do not treat them.
Detection
What we look for.
| Signal | Status | What it means |
|---|---|---|
| Suicidal thoughts and intent | Active | Words that state a wish to die, a plan, or a method. This is the signal the check was built for. |
| Thoughts of self-harm | Active | Words about hurting oneself, and check-in answers that report the same. Program check-ins carry a self-harm item. |
| Hopelessness and severe distress | Active | Distress without stated intent. This does not open a crisis screen. It notifies a person on the care team. |
| Risk of harm to another person | Active | Words describing violence toward someone else, including a member's report of being in danger at home. |
| Eating disorder symptoms | Active | Purging, going days without eating, and restricting to lose weight. These notify a person rather than opening a crisis screen — a symptom report is not an emergency, and treating it as one teaches people to stop mentioning it. |
| Financial harm to another person | Active | Someone taking or draining a member's money or benefits, and being pressured to sign or pay. Almost always a member telling us it is happening to them or to someone they care for. |
Every row above is live as of 10 September 2026. The last two were published here as unfinished before they were built, and that is the standard we hold: a protocol that only lists what already works is not a protocol.
Method
How the check works.
Before anything else
Every journal entry, chat message, reflection, and written exercise passes the check before it is saved and before any model reads it.
A curated phrase list
A fixed list of phrases runs first. It is hand-written from Stanley-Brown safety plan training materials and Columbia protocol interview examples. No model wrote it.
A second read for ambiguous words
Some phrases mean one thing in one sentence and another in the next. Those go to a second, narrower read. Clear cases and clear negatives skip it.
It fails toward safety
If the second read errors or times out, we treat the message as positive and escalate. The list favors recall over precision for the same reason.
The list recognizes more than standard American English. Members write the way they speak, and a safety check that only reads one dialect misses people.
Prevention
What our software will never produce.
Content that encourages, instructs, or treats as normal suicide, self-harm, an eating disorder, or harm to another person. This is a hard rule in the software, not a guideline for its authors.
Identity
It never claims to be a person.
Our software will not say it is human, a clinician, or a licensed professional. It will not imply that what a member tells it carries the confidentiality of a therapy session. It answers honestly when asked what it is.
Referral
Where we send people.
988 Suicide & Crisis Lifeline
Call or text 988, any hour, any reason. This is the referral our software offers first, by name, as a button a member can press.
988 chat
For a member who would rather type than talk. Opens the Lifeline's own chat at chat.988lifeline.org.
911
For immediate physical danger to the member or to someone else. Offered alongside 988, never instead of it.
Crisis Text Line
Text HOME to 741741 to reach a trained volunteer crisis counselor by text.
The Trevor Project
1-866-488-7386, for young LGBTQ people. This is the youth alternative we offer to a member we know to be under 25.
Spanish
988 answers in Spanish. Our own crisis screens carry Spanish text alongside English, not behind a language setting.
Every one of these is a working link in the app, not a number printed on a screen. A member in crisis should not have to copy anything.
One firm limit
We do not refer a crisis to the police.
A crisis referral goes to a crisis service. It never goes to a law enforcement agency because someone wrote something.
Colorado is the only state that writes this down. We apply it to every member in every state, because the reason holds everywhere: a member who fears the police will not tell us the truth, and a safety net nobody trusts catches nobody.
911 stays on the list above for immediate physical danger, offered to the member as a choice. That is a different thing from us calling the police about a member.
One thing this does not change. Some members of a care team are mandatory reporters by law. When someone tells us about abuse of a child or an older adult, we alert the reporter on the care team, and that clinician decides and makes the report their state requires. Atticus never files the report itself and never contacts an agency directly. We tell you this here rather than let you read the rule above as a promise it does not make.
Escalation
What happens after the first referral.
| Step | Status | What happens |
|---|---|---|
| In the moment | Active | The member sees a crisis screen with 988 and 911 as buttons, and their own safety plan if they have written one. The message is never left unanswered. |
| A person is notified | Active | The same event pages a care manager and a behavioral health role on the care team. Software does not close a crisis by itself. |
| Recorded as evidence | Active | Every detection and every referral writes an audit record we can show an accreditor or a regulator. |
| Second-tier follow-up | In progress | A member who keeps expressing risk after a referral needs more than the same referral again. A designed follow-up step, informed by clinical practice, is being built. Today the human escalation above carries this. |
Handling
What we do with what a member wrote.
Identifiers are stripped first
Text sent for the second read is stripped of record numbers, dates of birth, phone numbers, email addresses, and street addresses. The member's identifier is never sent.
Nothing is retained by a vendor
The second read runs inside our own cloud account under a signed business associate agreement. No conversation history is kept on the vendor's side.
The member never reads a verdict
A member sees a crisis line and their safety plan. They never see a label, a score, or a reason. We do not tell a person what a classifier decided about them.
It is a safety route, not a diagnosis
The check matches words and opens a door. It does not produce a mental health finding, and nothing it produces goes on a member's record as one.
Annual report
How many referrals we made.
We publish a count once a year, with no personal information in it, and we keep the prior year's number on this page.
Crisis referrals in calendar year 2025
Atticus served no enrolled members in 2025. The number is zero because there was nobody to refer, not because nothing was detected.
First reporting year is 2026
The 2026 count is published here by December 31, 2026, and filed with the state agencies that ask for it.
Reporting
Tell us when the software gets it wrong.
Anyone can report a safety problem with our software, whether or not they are a member.
Email safety@atticushealth.com. Tell us what the software said or failed to say, and when. A report about a live danger is handled the same day; everything else gets an acknowledgment and an outcome.
A member never has to accept this check to get care, and refusing anything on this page is never a reason we withhold care.
Related
The rest of our compliance record.
Accreditation status, security controls, and how we handle health information sit on their own pages.