Put yourselves in the shoes of a health plan’s behavioral health hotline responder. This is a person, typically a clinician, whose job it is to provide immediate support as soon as members are flagged as a risk – a routine call that is made every day.
Suicide remains one of the leading causes of death in the United States, with the CDC reporting over 48,000 suicide deaths in 2024. For a health plan’s response services team, the interaction likely begins before the member ever picks up the phone. A member might complete a behavioral health screening and report suicidal thoughts, or enter something concerning into a digital therapy exercise. The responder knows the member needs immediate attention, but without the right context, they’re starting from a disadvantaged position.
All the responder knows is that the member is in need of help. Their primary objective is to build trust quickly enough to understand what’s going on and how best to intervene, which is hard to do when you’re starting from square one. A flurry of questions needs to be answered as soon as possible: Does this person have a history of mental health disorders? Was there a recent traumatic event, a change in medication or a pattern of disengaging from care? When was the last time they received treatment?
Those answers may already exist in the health plan’s data, but the problem has historically been synthesizing or pulling it in a way that is relevant to the situation.
Suicide risk is notoriously difficult to assess from any single interaction. A massive 2020 study of more than 3.7 million patients across five health systems found that models using structured electronic health record data achieved an average sensitivity of 38% at 90% specificity for predicting suicide attempts, with high-risk patients identified a mean of 2.1 years in advance. More recent research has found that combining patient self-reports with EHR data improved machine learning prediction further.
These findings demonstrate the potential of using data to identify members who may be at higher risk, but they don’t tell a responder what’s happening with a member in the moment a crisis signal appears. AI can do much more to improve proactive interventions by ensuring responders have a clearer picture of the member they’re reaching out to before they even make contact.
But that’s as far as AI should go. That interaction cannot, and should never, be automated. We as an industry have to keep voice AI as far away from crisis lines as possible; this is a human exchange that must rely entirely on communication between two people. A voice agent can’t establish trust, interpret the nuances of an inherently risky conversation, or reliably make the judgment calls that a trained human responder is there to make. The most appropriate and helpful role for AI in this interaction is giving responders a way to mitigate as much risk as possible.
Think of all that AI has done to improve the experience of seeing a doctor: AI-enabled systems are giving clinicians snapshots of the patients they’re about to see before they walk into an exam room, and ambient scribes have released them from their computer screens to have an actual, human-to-human experience. Crisis response teams deserve the same kind of context, particularly because the stakes are so high.
An AI-enabled system for crisis responders could bring together recent behavioral health activity, changes in symptoms, utilization, prior interactions and other relevant signals into a concise summary before the responder engages. If a member answered positively to question nine on the PHQ-9, for example, that signal could trigger outreach. Before making the call, the responder could see the member’s recent symptom scores, behavioral health activity, prior interactions, and gaps in care. If the member recently expressed alarming thoughts in a therapy exercise, the responder can see that context and ask what was going on when the member wrote it. If utilization has changed sharply over the past several weeks, that context can be available, too.
Supplied with resources like this, the responder would be able to spend more of the conversation listening, asking questions and helping the person in crisis through the immediate situation.
There’s good evidence that what happens after the initial crisis matters, too. A large study of patients treated in emergency departments for suicide-related concerns found that a safety-planning intervention combined with structured follow-up calls was associated with 45% fewer suicidal behaviors over six months and more than twice the odds of attending an outpatient mental health visit. A separate study of more than 76,000 patients found that outpatient follow-up within seven days of psychiatric hospitalization was associated with a lower subsequent risk of suicide.
The opportunity for health plans is to move from an inbound, manual response approach toward one that acts on signals as they appear. When an alert reaches response teams, they should have a clear picture of the person they’re about to reach out to before they act. The sequence moves from signal to context to deep human connection much faster, and much more efficiently, with the responder still making the call and leading the intervention.
This month, the healthcare industry will once again focus attention on our country’s suicide crisis. While health plans can do much more to recognize risk and connect members to the help they need, they should also reassess a fundamental aspect of intervention: when a signal appears, how quickly can responders understand who they’re about to call?
In a crisis, every one of those seconds matters.
About Jeremy Kreyling
Jeremy Kreyling is the SVP of Healthcare Informatics at NeuroFlow, the market leader in behavioral healthcare technology infrastructure, where he leads analytics strategy to integrate behavioral health into value-based care models. He brings more than 20 years of healthcare data and analytics leadership, with prior roles at NOCD, ProgenyHealth, and Change Healthcare.
