Building safety into mental health AI

Introducing Sona, a new voice in mental health, trained by clinicians, and measured against the industry’s most demanding safety benchmarks. 

AI is moving into mental health faster than the industry has developed standards for evaluating it. How you build in that environment is a choice many are being faced with. 

We spent a decade in substance use care learning that outcomes depend on architecture: matching people to the right level of care, at the right moment, with the right clinical oversight in place. When we built Sona, we applied that same thinking to how AI can deliver care safely. 

Sona is a voice-first AI mental health specialist, developed and supervised by Pelago’s clinical team. It meets people in the everyday moments that traditional psychotherapy models were never built to reach, and it draws on the full intensity of Pelago’s clinical care capabilities when someone needs more. It remembers every conversation, operates under continuous clinical oversight, and escalates to a human clinician the moment a situation calls for one. 

For employers and health plans, Sona acts as an AI-powered assessment, triage, and treatment layer that extends evidence-based support to an entire population. With Sona, our customers pay for care that is actually delivered, matched to each person’s individual needs, with our fees tied to the outcomes we produce.  

The clinical decision behind a voice-first product 

Sona is voice-first by design, because people are most honest when they are talking. They hesitate, their tone shifts, they say things out loud that they might not write down. Sona listens to the full texture of every conversation and draws on clinical signals from tone, pacing, and emotional register as it unfolds. For mental health, the information that lives between words is often the difference between catching something early and missing it entirely. 

Behind member conversations with Sona, Pelago’s clinical team is present, reviewing performance, refining protocols, and retaining judgment over every clinical decision. Sona doesn’t replace a therapist, diagnose, or prescribe treatment. It extends the reach of a clinical team into the everyday moments where support has never been easy to access. 

How Sona handles risk 

Sona’s risk handling operates on three levels:

Detection. Sona continuously screens for signals of elevated risk within every conversation in real time, including indirect and ambiguous language, the way risk actually presents.

Escalation. When a conversation warrants human judgment, Sona routes it to one. Defined protocols govern when and how members move to Pelago’s care team, and to crisis resources when needed. These thresholds were set by clinicians, and they are deliberately conservative.

Oversight. Pelago clinicians supervise Sona’s operation on an ongoing basis, reviewing performance, refining protocols, and retaining authority over every clinical decision.

We designed Sona this way because we believe AI in mental health should be held to a clinical standard, with the supervision and accountability that standard implies.

Measuring safety in the open 

Over the past year, digital health companies have built open safety benchmarks for mental health AI and published them for anyone to use. VERA-MH evaluates how AI systems handle high-risk and emotionally complex conversations. MindEval scores AI support across clinical dimensions. We evaluated Sona against both:

VERA-MH v1.1.0

VERA-MH is an early effort to define a standard for conversational mental health AI safety. We evaluated Sona on v1.1.0, the current version, which is more demanding than the original v1.0.

Sona’s risk detection sub-score, which measures whether the system identifies potential risk in a conversation, was 100. We believe identifying risk is the single most important job of any AI operating in this space, and it is the capability we have invested in most heavily.

MindEval

MindEval evaluates multi-turn supportive conversations across five clinical dimensions, each scored from 0 to 6. Sona scored highest of any tested system on every dimension.

What benchmarks can and can’t tell you

Benchmarks are point-in-time evaluations. They measure how a system responds to a defined set of scenarios, and a strong score is necessary but not sufficient evidence of safety in the real world. Real safety comes from the full architecture: detection, escalation, human oversight, and continuous monitoring of live performance, all operating together over time. 

When Pelago began delivering substance use care virtually, most digital health companies were using the PHQ-8, a version of the standard depression screening tool that omits the question about suicidal ideation. The reasoning was liability. Asking meant owning the answers and complex clinical workflows that would follow. But we disagreed with this approach. Substance use is itself a risk factor for suicidality, and omitting the question to avoid the complexity of the answer felt like the wrong tradeoff. 

In 2023, we published a peer-review study on implementing real-time suicidality screening across our substance use care program — the first virtual care platform to do so at scale. We screened 100% of eligible members, found that 8% screened positive, and demonstrated that universal screening was feasible without requiring new clinical infrastructure. Every positive screen was non-acute, and every one received safety planning and clinical follow-up. 

We applied that same thinking to Sona. In a Phase 1 study of Sona’s suicidality detection and escalation protocol, 13% of participants endorsed suicidal ideation. When a positive screen was detected, a Pelago clinician was notified immediately through a real-time escalation system. The average time from detection to outreach was 11 minutes. That study has been submitted for peer review and will be the first published account of how to keep humans in the loop for suicidality detection and escalation in a generative AI mental health tool.

Early 4-week pilot data from Sona shows what happens when the architecture works as intended. Among members with baseline depression or anxiety, 71% reduced at least one GAD/PHQ severity tier, with an average symptom score reduction of 53%. On the CDC’s Healthy Days measure, Sona members gained an average of 2.5 healthy days per person per month. The share of members reporting frequent distress — 14 or more unhealthy days in the past 30 — fell from 41% to 28%. 

The mental health system has spent two decades expanding access to one kind of care. We built Sona because we believe the next decade belongs to matching care to need, with technology doing what it does best and clinicians doing what only they can. Measured openly. Accountable for results.