The Four-Tier Risk System: How AuraLift Decides When to Stop Coaching

Every AI wellness product faces a recurring question: at what point does the conversation in front of us stop being a coaching conversation? The cost of getting this wrong is high, coach when you should refer, and you delay needed care; refer when you should coach, and you teach users that the product is brittle. AuraLift’s answer is a four-tier risk system. This piece walks through what each tier means, what triggers movement between tiers, and how the system fails safely when it gets things wrong.

Why a tiered system exists

The naive way to design an AI wellness product is binary: it’s either coaching or it’s 988. Every conversation starts in coaching mode and stays there unless an unambiguous crisis word, “suicide,” “kill myself”, triggers an emergency handoff.

This design fails in two predictable directions. It under-routes the user with rising suicidal ideation who hasn’t used those exact words, because the binary detector waits for a keyword that may never arrive. And it over-routes the user processing a difficult memory who casually says “I felt like dying”, meaning they were tired, not suicidal, and then receives a crisis-line referral that breaks the conversational frame.

A four-tier system is built to handle the long middle band where actual users live. Most coaching conversations stay at Tier 1 their entire life. A subset rises to Tier 2 around difficult content. A smaller subset crosses into Tier 3, where AuraLift is no longer the right tool. A small but consequential subset reaches Tier 4, where the only correct move is to route to crisis support and stop being a coaching product for that user, in that moment. The point of the tiers isn’t to make the system clever. It’s to make the system right at every point along that gradient.

Tier 1, Coaching baseline

Tier 1 is the default. This is where most conversations begin and most stay. The user is bringing material that fits coaching: stress about work, frustration with a relationship, an uncomfortable feeling they want to understand, a habit they want to build, a 3am thought spiral. AuraLift’s job in Tier 1 is the standard one, present, structured, drawing from CBT/DBT/ACT/mindfulness traditions, holding the conversation without forcing it.¹

Tier 1 includes content that is emotionally heavy but coaching-appropriate: grief that isn’t complicated, work burnout that doesn’t involve crisis, relationship conflict that isn’t abusive, anxiety that’s elevated but not impairing. The line we draw is functional: if the user is broadly intact and the content is not safety-relevant, Tier 1 is the right place to be.

Tier 2, Heightened distress

Tier 2 is what happens when the conversation moves into territory where coaching can still continue, but with more care. The triggers are softer than Tier 3 or 4: trauma content (a specific incident the user is processing), high-intensity affect (sustained crying, panic symptoms in real time), descriptions of overwhelming stress, moments where the user names feeling unsafe but the unsafety is emotional rather than physical.

At Tier 2, AuraLift slows down. The pace shifts toward grounding and pacing. The system explicitly checks in: this sounds heavy, are you okay continuing here, or would something different be more helpful? Recommended techniques narrow toward distress-tolerance (TIPP, grounding, paced breathing, sensory practices) and away from anything that might intensify activation (deep cognitive work, exposure-style questions, anything that asks the user to recall more material).² The coaching does not stop. It changes register.

Tier 2 also opens an explicit reminder of resources: the user is told that what we’re doing is coaching, that other resources exist, and that if anything in the conversation feels like more than they want to hold here, here is the path to professional support. This reminder is not a hand-off. It’s a soft offer of off-ramps.

Tier 3, Clinical referral

Tier 3 is the explicit acknowledgment, in the conversation, that AuraLift is not the right tool for what’s being shared. The triggers are firmer:

• Symptoms that meet a clinical threshold by description: persistent severe depression, symptoms suggestive of OCD, eating-disorder behaviors meeting medical concern, dissociative symptoms, psychotic-spectrum content (hallucinations, paranoid delusions).
• Substance use disorder content that is active and impairing.
• Disclosure of an abuse situation that needs human professional support: domestic violence, child abuse, or elder abuse, especially when ongoing.
• Medication-related questions or distress that need a prescriber.
• Any pattern of escalation across the conversation where coaching is plainly not stabilizing things.

At Tier 3, AuraLift does something specific: it says, in plain language, that what the user is describing benefits from professional support, and it provides resources. It doesn’t pretend it can do the work. It stays warm, it stays in the conversation, but it stops offering coaching techniques as the answer to a question that needs a clinician.

Tier 3 is not a permanent banishment from the product. The same user can return and have a Tier 1 conversation about a different topic the next day. What Tier 3 says is: for this thread, this is not the right room. Resources offered include the SAMHSA national helpline, Psychology Today’s therapist finder, the user’s primary care physician (if disclosed), and any culturally appropriate options the user has named.

Tier 4, Crisis routing

Tier 4 is the smallest tier in volume and the most consequential. It is triggered by content that signals immediate safety concern:

• Active suicidal ideation, especially with plan, means, or stated intent.
• Self-harm in progress or with stated intent in the next hours.
• Imminent danger from another person, domestic violence escalating in the present, threats of harm.
• Acute psychosis with safety implications.
• Medical emergency content: overdose described in real time, severe dissociation with risk, severe medical decompensation.

At Tier 4, AuraLift exits coaching mode. The interface transitions to crisis-routing mode. The user is given the 988 Suicide & Crisis Lifeline (call or text 988), the Crisis Text Line (text HOME to 741741), and 911 if appropriate.³ International users see localized options from findahelpline.com. The product does not return to coaching content during a Tier 4 event. If the user de-escalates, AuraLift can re-enter coaching after explicit confirmation, but never automatically and never without the user being clear about their state.

The decision about Tier 4 is asymmetric on purpose: it is better to reach Tier 4 in a case that turns out to be Tier 2 than to stay in Tier 2 for a case that is actually Tier 4. False-positive Tier 4 routing is a usability cost. False-negative Tier 4 routing is a catastrophic cost. We tune toward the safer error.

How signals are detected

Detection is layered, intentionally, because no single layer is reliable enough. Three layers work together.

Lexical signals. Curated lists of high-risk language, regularly reviewed, cover phrases that should not be missed. This is the simplest and most brittle layer, false positives are common (“this work is killing me”), but the cost of missing a phrase like “I have a plan” in context is unacceptable, so the layer stays.

Model-based signals. A classifier evaluates the rolling context of the conversation and produces a risk score. Unlike the lexical layer, the classifier reads meaning rather than words: a user who says “I just don’t see the point anymore” in the context of a description of escalating despair scores higher than the same phrase in the context of a frustrating workweek.

Conversational metadata. Patterns in how the user is interacting (sudden escalation in message length, repetitive content, abrupt withdrawal, time-of-day patterns) feed into the same scoring. None of these alone are diagnostic; aggregated, they help.

Decisions about tier transitions use all three layers. The system is biased toward escalating early at Tier 2 (low cost, easy to walk back), more conservative at Tier 3 (genuine cost to misroute someone whose conversation should have stayed coaching), and biased toward triggering at Tier 4 (the asymmetric stakes above).

Failure modes we plan for

The system fails. Like all systems, it fails sometimes. The four failure modes we plan for:

• Missed escalation. The system stays at Tier 1 or 2 when it should have moved to 3 or 4. The mitigations: the lexical layer catches keyword-level signals the classifier might miss; periodic re-evaluation reclassifies the conversation as it accumulates context; explicit safety check-ins offer the user an off-ramp regardless of tier.
• Over-escalation. The system jumps to Tier 3 or 4 prematurely on a conversation that should have stayed coaching, breaking the user’s flow. The mitigations: at Tier 3 the system explains why and offers to continue at a different register; at Tier 4 the user can confirm de-escalation and return to coaching; the experience of being misrouted is acknowledged as a real cost.
• Stuck at a tier. The system enters a higher tier and doesn’t exit when it should, for example, treating every subsequent message as Tier 3 because an earlier message triggered it. The mitigation: tier evaluations are stateful but not sticky; downward transitions happen automatically when content patterns change.
• Adversarial input. Users who try to bypass the safety layer through phrasing tricks, role-play, or hypotheticals. The mitigation: detection is layered and intent-aware; explicit role-play of suicidal content is treated as a safety signal regardless of framing; the system errs toward the safer interpretation when ambiguous.

What this system isn’t

It isn’t a clinical risk assessment.

A licensed clinician, in a clinical encounter, conducts a structured risk assessment that considers history, ideation, intent, plan, means, protective factors, and context. That assessment is a piece of professional clinical judgment with legal weight. The four-tier system is not a substitute for it and is not meant to be. What it does is route, fast, toward the right resource, including the resource of a clinical risk assessment. The system is designed to recognize when one is needed and step out of the way.

It also isn’t a treatment system. AuraLift does not treat. The four tiers are gates that govern how AuraLift coaches and where it stops; they are not a treatment plan and do not replace one.

How the system is governed and updated

The four-tier architecture is reviewed on a quarterly cadence by the AuraLift product and safety teams, with input from AuraLift’s Clinical Advisory Board on the safety and clinical-boundary questions. New cases, both successful routings and missed escalations identified through retrospective review, feed back into the detection layer and the boundary definitions.

Concretely, this means: the lexical lists are not static (they expand as language evolves and as we encounter new edge cases), the classifier is retrained, the tier boundaries are re-examined, and the resources offered at each tier are re-validated against the latest-available crisis-line standards. This is governance work, not product work, it doesn’t ship as a feature, but it’s the most consequential thing the team does.