If you survey the AI mental health tool landscape today, you'll find a wide range of approaches to crisis handling — from "we display the 988 number in the footer" to "our algorithm detects distress and routes to a human." Most products claim to take safety seriously. Very few publish the specific criteria they use, the response times they commit to, or what happens downstream once a crisis signal is detected.
This opacity is a problem. When a digital mental health tool is deployed to a population of employees or health plan members, the organizations deploying it are implicitly accepting liability for what happens when a user in crisis interacts with it. Medical directors and VP Clinical roles signing off on these deployments should be asking specific questions about crisis escalation architecture — and getting specific answers, not marketing reassurances.
We've thought carefully about this at Neurodex, and we've published our approach internally and shared it with our clinical advisors for review. This article describes what we built and why. We're not claiming it's the only valid approach, but we're publishing the specifics because the field needs concrete examples to debate, not more vague commitments to safety.
The Problem with "Display 988 and Hope"
The 988 Suicide and Crisis Lifeline is a genuinely valuable public resource. Call or text 988 connects individuals in crisis with trained counselors, typically within a few minutes. It's the right resource to reference for someone in acute distress. The problem isn't with 988 itself — it's with treating a static 988 reference as a crisis safety architecture.
A user who is engaging with a mental health app in the early stages of a crisis ideation cycle — not yet in acute crisis, but moving in that direction — is unlikely to spontaneously navigate to a footer link and initiate a 988 call. They're having a conversation with the app. If that conversation doesn't have the capability to detect deteriorating signals and actively route toward crisis resources at the right moment, the 988 mention is legally protective but clinically meaningless.
Effective crisis escalation in a digital mental health context requires: (1) detection — multiple input channels for crisis signals, not just explicit statements; (2) graduated response — different responses for different levels of risk, not a binary "fine" or "call 911"; (3) active handoff — warm transfer to human resources, not a link; and (4) back-channel notification — alerting the clinical operator, not just presenting resources to the user.
Neurodex's Tier System
We use a four-tier crisis signal classification that determines escalation response. This isn't a proprietary algorithm — it's a clinical logic framework that should be standard practice for any tool in this space.
Tier 1 (Passive Flag): PHQ-9 item 9 scored at 1 ("several days") during a routine check-in with no acute language in conversational content. Response: acknowledgment within the session, increased check-in frequency for the following week, and a clinical team notification logged within four hours. No immediate interruption of the session.
Tier 2 (Active Monitoring): PHQ-9 item 9 scored at 2 or higher, or passive detection of language patterns associated with hopelessness or worthlessness without explicit ideation statements. Response: structured safety check within the session using validated inquiry language, 988 resource presentation with active prompting (not passive display), clinical team notification within one hour, and a flag for clinician review before the next scheduled session.
Tier 3 (Crisis Intervention): Explicit ideation statements detected ("I've been thinking about ending it," "I don't want to be here anymore") or PHQ-9 item 9 scored at 3. Response: session immediately pivots to safety protocol, 988 is presented as an active prompt with click-to-call functionality, the user is asked directly about safety, and a real-time notification goes to the clinical team with a 30-minute response commitment. The session does not continue with CBT content until safety check is resolved.
Tier 4 (Immediate Escalation): Explicit statements of intent, means, or plan. Response: session halts completely, user receives 911 and 988 as immediate prompts, automatic notification to clinical team with 5-minute response commitment, and if the tool is deployed within an employer or health plan context, the clinical operator's on-call protocol is triggered.
What "Detection" Actually Means — And Where It's Imperfect
Tier 1 and 2 signals are relatively straightforward because they rely on validated instrument scoring (PHQ-9) rather than natural language interpretation. PHQ-9 item 9 has reasonably well-characterized sensitivity and specificity for suicidal ideation at different score levels. The structured measurement channel is more reliable than language detection for identifying passive ideation.
Tier 3 and 4 require language-based detection, which is where the architecture has inherent limitations that any honest account of this system should acknowledge. Natural language processing for suicidal ideation expression is imperfect. It catches explicit statements well. It struggles with implicit expressions — dark humor, indirect statements, cultural idioms of distress that don't map cleanly to the detection patterns trained on English-language clinical corpora.
We err toward false positives. A Tier 3 response triggered by ambiguous language that turns out not to reflect ideation is, from a clinical safety standpoint, acceptable. It creates friction for the user and workload for the clinical team. A false negative — a Tier 3 situation that was classified as Tier 1 because the language didn't trigger detection — is clinically unacceptable. Our threshold calibration reflects this asymmetry explicitly.
We're not claiming our detection is comprehensive or infallible. A user determined to conceal their crisis state from the system can do so. This limitation is shared by every digital tool in this category and by many in-person clinical encounters as well. The goal of crisis detection architecture isn't perfect coverage — it's systematic capture of the signals that a well-designed system can catch, combined with clear escalation pathways when those signals appear.
The Clinical Operator's Role
An important design decision in Neurodex's crisis protocol is that it's not designed to fully automate crisis response. The system detects, escalates, and presents resources — but for Tier 2 through Tier 4, a human clinician on the operator's side is in the loop within defined response windows.
This matters for a few reasons. First, the clinical judgment about what to do when a crisis signal fires depends on context that the system doesn't have: the patient's history, their existing therapeutic relationship, whether they're known to present with passive ideation at times of stress but not progress to acute risk. A clinical team that knows a particular member can make a more calibrated response than any algorithm.
Second, the regulatory and liability landscape for autonomous crisis response by AI tools is genuinely unresolved. The system doing a safety check through a validated protocol is different from the system making a clinical determination about disposition. We've kept Neurodex squarely in the former category while requiring the latter to be in human hands.
This means that deploying Neurodex to a population requires a clinical operator with defined response protocols — typically a behavioral health medical director, clinical care team, or EAP provider — who has agreed to the response time commitments and has the infrastructure to receive and act on escalation notifications. We don't sell to operators who want to set up the tool without clinical oversight. That's a hard constraint in our deployment requirements, not a preference.
What the Field Should Standardize
We'd like to see a few specific things become standard practice for AI mental health tools, not because we invented them but because they represent the minimum viable safety architecture that any clinical deployment should require.
First: published, specific crisis tier criteria. Not "we take safety seriously" — the specific signal types and score thresholds that trigger each response level, so operators can evaluate whether the product's sensitivity matches their population's risk profile.
Second: defined response time commitments for each tier, with accountability mechanisms. Not "we notify your clinical team" — within how many minutes, by what channel, with what expectation of operator response.
Third: session interruption architecture for high-acuity signals. A tool that detects a Tier 3 signal and continues delivering CBT content while also displaying 988 is not safe. The session architecture has to respond to the signal, not just log it.
Fourth: disclosure to users at onboarding about the crisis protocol. Users should know, before they begin using any digital mental health tool, what happens if the system detects a crisis signal — who gets notified, what the response will look like, and what resources are available. This isn't just ethical — it's clinically important because users who understand the protocol are more likely to engage honestly with safety check questions when they arise.
The absence of a published industry standard for AI mental health crisis escalation isn't a benign gap. It's an environment where tools with inadequate safety architecture can make marketing claims that suggest parity with tools that have built real protocols. Health plans and employers selecting digital behavioral health tools for their populations deserve the ability to compare safety architectures on specific criteria. Until a standard exists, publishing the specifics is the least any serious tool in this space can do.
